One Goal, One Machine
The ABC proposition: apps, a back office and context, run by agents
How a company of one picks one goal and builds one machine to move it: Apps, a Backoffice and Context, run by agents. 100 chapters across 10 parts, one diagram per chapter. Private draft.
Preface
How to read this one.
How a company of one picks one goal and builds one machine to move it: Apps, a Backoffice and Context, run by agents. 100 chapters across 10 parts, one diagram per chapter. Private draft.
- ↓ Scroll vertically for the 100-chapter journey.
- → Swipe horizontally on any chapter for its plate + details.
- ✦ Random Chapter and Random Book pinned to the footer.
Contents
tap a chapter to jump
100 chapters.
One Goal, Not Twelve
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most small businesses do not lack goals. They have twelve, written on a whiteboard in January in a confident hand, and by March eleven of them have quietly become furniture. The twelfth survives only because nobody has looked at it either. Nobody decided to abandon them. They simply lost a series of small, polite arguments with Tuesday.
ABC starts somewhere less ambitious and considerably more useful: one goal for the year . Not a vision, not a mission, not a values statement printed on a mug. One goal, stated so plainly that a stranger could check, on any given morning, whether you are winning. Everything we build, every agent we switch on and every morning status we send exists to move that goal. If a piece of work does not move it, the work does not get built, however clever it looks in a demo.
This sounds restrictive. It is. That is rather the point. A one-person business has one head, one pair of hands and roughly one good idea's worth of attention per day. Twelve goals divide that attention twelve ways and call the result strategy. What actually happens is that each goal gets a slice too thin to move anything, and at the end of the year all twelve have advanced by roughly the width of a whiteboard marker.
The ancient Stoics had a word for the discipline of choosing what deserves your effort: prohairesis , the faculty of choice. They thought it was the only thing truly yours. We think it is also the only thing a small business cannot outsource. A machine can draft, sort, match, publish and report. It cannot tell you which of your twelve goals matters. That choice is the first and most valuable thing you will do with ABC, and it costs nothing but the nerve to let eleven good ideas go.
People worry that a single goal makes them narrow. In practice it does the opposite. With one goal settled, every other question gets easier, because it has a referee. Should we build the booking app? Does it move the goal? Should we answer that partnership email today? Does it move the goal? The answers are not always no. They are simply always quick.
The rest of this book assumes you have picked yours. If you have not, the onboarding will make you, gently, in about five minutes. It will ask what you would be proud of on the last day of the year, and then it will keep asking until the answer is a single sentence. You will be surprised how much lighter the whiteboard looks afterwards.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
A Number With a Date
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A goal you cannot measure is a mood. "Grow the business" is a mood. So is "build the brand", "get serious about content" and the perennial favourite, "scale". Moods are pleasant to have and impossible to fail, which is exactly why they never get done. Nothing is ever late for a mood.
ABC turns a mood into a goal with two small edits. The first edit adds a number . Not "more revenue" but "£1,000 a month of recurring revenue". Not "an audience" but "five hundred people on the email list". The number does not need to be large. It needs to be countable by someone who is not you, on a day when you are not in a good mood. If two honest people could look at the evidence and disagree about whether you hit it, it is not a number yet.
The second edit adds a date , and the date matters more than the number. A number without a date is a wish that has learnt arithmetic. "£1,000 a month" could be achieved next spring or in 2041, and both would technically count. "£1,000 a month by 30 June" forces a different kind of thinking. It tells you how many weeks you have, which tells you how much you can try, which tells you what you must stop trying today. Deadlines are not cruelty. They are the only way a plan finds out how big it is.
There is a temptation, once the date is set, to make the number heroic. Resist it. A goal that would require everything to go right is not a goal; it is a raffle ticket with your name on it. Pick the number you would be genuinely pleased to hit and slightly nervous about missing. That nervousness is useful. It is what gets you out of bed on the Thursday when nobody is watching and the work is dull.
The goal also needs to be written down where the machine can read it. In ABC it lives at the top of your context, as one line, with the number, the date and the source of truth that will prove it. Every agent reads that line before it does anything. When the morning status arrives, the first thing it reports is the distance between today's figure and that line. Not your activity, not your effort, not how busy you felt. The distance.
This is the least glamorous chapter in the book and possibly the most important. Everything that follows, the apps, the back office, the context, the agents, is machinery. Machinery needs a target, or it simply runs very efficiently in a circle. Give it a number and a date, and you will find it has a remarkable sense of direction.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The 1G
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every proposition is easier to believe when the person selling it is using it. So here is ours. For 2026, Mat Siems has one goal, known around the workshop as the 1G . It is written down, it has a number and a date, and everything in this book was built to move it. The rest of the business is arranged around it like planets around a slightly demanding sun.
The 1G does not move on its own. It is held up by three pillars, and their initials give the whole system its name. A is for Apps : a fleet of small, single-purpose applications, each built quickly from a single dataset and each expected to earn its keep. B is for Backoffice : one place where every asset, number, file and payment lives, so the business has a single version of the truth instead of twelve spreadsheets with opinions. C is for Context : the private, written memory of the business, the goals, decisions, people and rules, which every agent reads before it acts.
Around those three sits a fourth idea, borrowed from games and called the Prodgame . The year is treated as a game with levels, runs and a score. It sounds whimsical. It is in fact a very serious way of making sure the work is visible, measured and occasionally fun, which are the three conditions under which humans tend to keep going.
None of this is unique to Mat. The shape is the proposition. Your goal will be different, and your apps, back office and context will contain different things, but the arrangement holds: one goal in the middle, three pillars around it, and a machine that reports every weekday morning on whether the pillars are doing their job. When a client finishes onboarding, the first thing they receive is a short book that draws this same picture with their own goal at the centre.
There is a reason for using initials rather than something grander. ABC is easy to remember and impossible to inflate. Nobody has ever been intimidated by the alphabet. When the system is described in three letters, nobody mistakes it for a platform that will take eighteen months to implement, and nobody needs a consultant to explain the acronym. A, B and C. Apps, Backoffice, Context. You could explain it to a twelve-year-old, and if you are lucky, the twelve-year-old will ask why every business does not work this way.
A pillar that does not hold anything up is a sculpture. In ABC, each pillar must earn its place by moving the 1G, or it is trimmed until it does. That rule applies to Mat's machine, and it will apply to yours.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What Gets Dropped
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Choosing one goal is mostly choosing what to stop. This is the part nobody puts on the whiteboard. It is easy to write down the thing you are going to do. It is much harder to write down, in the same pen, the eleven things you are not going to do, and then to look at them every day without flinching.
So let us hold a small funeral. Not a sad one. The kind with sandwiches. On the list are the podcast you were going to start, the second newsletter, the redesign of a website that works perfectly well, the partnership with a company you met once at a conference, the course, the community, the side project that was going to be "just a weekend", and the app idea you had in the shower, which was excellent, and which will still be excellent next year. Each of them was a good idea. That is what makes this hard. Bad ideas drop themselves.
ABC makes the funeral routine by asking the same question of every project, every feature and every request: does it move the goal? If it does, it goes into the backlog and the machine gets to work on it. If it does not, it is dropped, kindly, with a note explaining why, so that future you does not dig it up in a moment of boredom and wonder who buried it. The note matters. A decision that is written down stays decided. A decision that lives in your head is reopened every time you are tired.
Dropping is not the same as deleting. Ideas that do not move this year's goal go into a parking file in your context. Nothing is lost. When the goal changes, the parking file is the first place the machine looks. Some of those ideas will turn out to be next year's goal. Most of them will quietly reveal, after six months in the car park, that they were never going anywhere at all. Either way, you did not spend a single Tuesday on them.
The Stoics practised something they called voluntary discomfort: going without, on purpose, to discover how little they needed. Dropping eleven goals is the business version. You will feel the loss for about a week. Then you will notice that your days are shorter to plan, your list fits on one screen and the work you do finish is the work that counts.
There is a final benefit nobody expects. When you say no to eleven things, the twelfth gets your whole attention, and work that receives your whole attention tends to get done. That is not discipline. It is arithmetic, with sandwiches.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Scoreboard
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
If the goal is a number, it needs somewhere to live where you can see it every day. ABC calls this the scoreboard , and it has one job: to show today's real figure against the goal, without commentary, without adjectives and without anybody having typed it in by hand.
That last condition is the whole trick. A scoreboard that relies on someone updating a cell on Friday afternoon will be accurate on Friday afternoon and fiction for the other six days. Worse, it will drift towards optimism, because the person updating it is also the person who wants it to look good. So the ABC scoreboard is fed from real sources: the store export, the payment processor, the affiliate statement, the analytics account. Each number on the board is connected to a feed, and each feed is connected to evidence you can click.
In Mat's own machine the board is a backlog item with a plain name, BL-076, and a plain rule: null beats invented . If a feed is not connected yet, the board shows nothing, an honest empty box, rather than a guess. This feels uncomfortable at first. An empty box looks like failure. It is not. It is a precise statement of what you do not yet know, and the first job of the back office is to fill those boxes with real numbers, one feed at a time. A board full of guesses would look better and mean less.
The scoreboard is layered, and the layers matter. At the top sits the goal number, unchanging, with its date. Beneath it sits today's real figure. Beneath that sit the feeds that produce the figure, and beneath those sit the receipts: the exports, the statements and the logs that prove each feed is telling the truth. Anyone who doubts the top line can walk down the stack and check. Most people never do. They simply relax, because they know they could.
There is a psychological point here too. A number you look at every morning changes your behaviour more than a number you look at once a quarter. Not because it nags you, but because it removes the fog. When you know exactly where you stand, decisions shrink. You stop worrying in general and start working in particular.
The morning status begins with the scoreboard, every weekday at 08:45. One line: the goal, today's figure, the distance between them. Sometimes the distance is discouraging. That is fine. A discouraging number you can trust is worth a great deal more than an encouraging one you cannot, and it has the considerable advantage of being true.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Evidence or It Did Not Happen
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask anyone in a small business what they did last week and you will get a story. It will be a good story, mostly about being busy. What you will rarely get is a link. The week happened, clearly. It is just not entirely clear what it left behind.
ABC has a blunt rule about this, borrowed from every sceptical auditor who ever lived: evidence or it did not happen . Work counts when it produces one of three things. A merged pull request, which proves code changed and was accepted. A live URL, which proves something is out in the world and anyone can open it. Or a file: an export, a statement, a document, a log. If a piece of work cannot point at one of those, it is not finished. It might be very nearly finished. The machine does not have a column for very nearly.
This is not an attack on effort. Effort is lovely. It is simply not the thing that moves the goal. Three hours spent researching a feature or rearranging a backlog may be necessary, but on the day it is done it is a story told about work. The output is what changes the scoreboard, and the output is what the morning status reports.
The rule applies to people and agents alike, and it is especially useful for agents. An agent will cheerfully tell you it has updated the page, fixed the bug and emailed the supplier. It is usually right. Occasionally it is confidently, fluently wrong. So every claim an agent makes in ABC carries its link. No link, no claim. The status does not include a line it cannot prove.
There is a diagram for this, and it is two circles. One circle is work that is done. The other is work that is visible. Plenty of work is done but invisible, sitting on a laptop where nobody can check it. Plenty is visible but not done, a draft shared in a chat, admired and abandoned. Evidence is the small overlap in the middle, where the thing exists and someone other than you can see that it exists.
Living by this rule changes the shape of a day. You stop starting six things and start finishing one, because one finished thing is worth a line in the status and six started things are worth nothing at all. You also stop needing to remember what you did, because the evidence remembers for you. The week becomes a list of links rather than a feeling.
It is a slightly unforgiving standard, and that is precisely its charm. A business that counts only what it can prove is a business that cannot fool itself, and fooling yourself is the most expensive hobby a company of one can have.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Weekday Rhythm
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most businesses have a calendar. Very few have a pulse. The calendar tells you where you are supposed to be at three o'clock. The pulse tells you whether anything is actually alive. A company of one can survive a messy calendar for years. It cannot survive long without a pulse.
ABC gives the one goal a pulse in the form of the weekday rhythm , a loop with three beats. Plan the day. Ship one thing. Report. Then do it again tomorrow. That is the whole of it, and its plainness is deliberate. A rhythm with nine beats is a process document. A rhythm with three is something you can keep on a wet Wednesday in February.
The loop starts with the report, oddly enough. At 08:45 the morning status arrives with the scoreboard, what moved yesterday, what is stuck and what needs a yes. That status becomes the plan. You read it, pick the one thing that most moves the goal today, and say so. The agents take the rest of the list. By late afternoon there is a merged change, a live page or a file, and that piece of evidence goes into tomorrow's report. The loop closes itself.
Notice that the loop asks for one shipped thing, not one perfect thing. A day that ends with something small and real beats a day that ends with something large and imaginary. Over a year of working days, one shipped thing a day adds up to a quantity of finished work that would have alarmed you in January, delivered without a single heroic weekend.
Which brings us to the weekends. The rhythm runs on weekdays only, and this is not laziness, though we would defend laziness if pressed. It is design. A machine that reports seven days a week trains its owner to check it seven days a week, and an owner who checks the business on Sunday morning is not resting, merely working in pyjamas. The agents may well run jobs over the weekend. The status waits until Monday.
There is a second reason. Rhythm needs rests to be rhythm. Without the gap, the beats blur into one long, anxious hum. The weekend is the rest in the bar. It is where the ideas that do not fit in a working day get room to arrive, usually in the shower, and occasionally on a walk that had no purpose at all.
The loop is also forgiving. Miss a day and nothing breaks: tomorrow's status simply reports one fewer link, without comment. There is no streak to protect, no guilt to manage, no app sending you a sad owl. You just pick it up on the next beat. The Stoics would approve. They were great believers in doing the next right thing, and rather less keen on catching up.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Attention Is the Budget
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask a small business owner what they are short of and they will usually say money. Ask them what they ran out of yesterday and the honest answer is something else. They did not run out of pounds at four in the afternoon. They ran out of the ability to think about anything harder than which biscuit to have next.
In a company of one, money is rarely the scarcest resource. Attention is. You have perhaps three or four good hours a day in which you can make a real decision, write something worth reading or solve a problem that needs your actual brain. Everything else is emails, logins and small fires. ABC treats those good hours like cash, because that is what they are, and it budgets them accordingly.
The tool for this is an old one, a simple two-by-two. On one axis sits effort, how much attention a task will cost. On the other sits goal impact, how far it moves the number. Everything you could do lands somewhere on the grid. The top corner, low effort and high impact, is where you spend first. High effort, high impact gets a planned slot. Low impact work goes to the agents, or to nobody. The bottom corner, high effort and low impact, is where attention goes to be wasted beautifully, and it is surprisingly crowded.
The machine does much of this sorting for you. When the morning status lists what needs a decision, it puts the cheap, valuable choices first and keeps the list short. Anything an agent can safely do without you, it does without you. Anything that merely wants your opinion is batched, so that ten small approvals cost one cup of tea instead of ten interruptions.
That is the real job of the agents in this book, and it is worth saying plainly. They are not there to make you faster at everything. They are there to keep your best hours for the work only you can do, and to stop the rest of the day from quietly borrowing against them.
It helps to notice what the leaks look like. Switching between six tabs is a leak. Checking a dashboard that changes once a day, eleven times a day, is a leak. Answering a message in four minutes that could have waited four hours is a very polite leak. None of them feels expensive. Together they cost you the afternoon.
The Stoics said we are frugal with our money and careless with our time, which is the one thing it is right to be miserly with. They had never seen a notification badge, but they would have recognised it immediately. Guard the good hours. Everything else can wait in a queue.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Green Every Weekday
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every machine has a light somewhere that tells you whether it is well. The kettle has one. The car has several, most of which you have learnt to ignore. ABC has one that matters more than all the others, and it is either green or red.
The light belongs to the control plane , the main code repository where the machine itself lives: the skills, the agents, the dashboards and the scripts that send the morning status. Every change to it, whether written by Mat or proposed by an agent, triggers a set of automated checks. Does it build? Do the tests pass? Has anything that used to work stopped working? The answer comes back as a colour. Green means yes. Red means somebody has some explaining to do.
The rule is short enough to fit on a mug. Main is green every weekday, and no red run survives a working day. If a change goes in at ten and turns the light red, it is fixed, or reverted, before the day ends. Not tomorrow. Today. The diagram shows the whole sequence: you push a change, the machine runs the checks, and by five o'clock the light is green again.
This sounds fussy, and it is, in the way that brushing your teeth is fussy. The reason is compound interest, the unpleasant kind. A red light left overnight becomes normal. A second failure arrives and hides behind the first. By Friday nobody is sure which change broke what, and the agents, who rely on the checks to know whether their own work is safe, are building on sand. One red evening costs ten minutes. One red week costs a week.
There is a gentler reason too. Green every weekday is a small promise, and small promises kept daily are how trust is built, not least with yourself. Big promises are easy to make and rarely tested. A promise you keep every single working day, without ceremony, quietly tells you that the rest of the machine is probably in order as well.
It also keeps the agents honest. In ABC an agent can propose a change, run the checks and open a pull request, but it cannot wave its own work through a red light. The checks do not care how confident the agent sounds, which makes them the most impartial colleague in the building. If the work is good, the light goes green and a human merges it. If not, the agent tries again.
Like the weekday rhythm, the rule takes weekends off. A red light on a Saturday is a problem for Monday morning, and the 08:45 status will mention it first. The machine does not need you to stand guard over it. It just needs you never to go home with the light on red.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Goal Is the Product
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Nobody has ever bought a drill because they wanted a drill. The old marketing line says they wanted a quarter-inch hole. Look a little closer and they did not want the hole either. They wanted the shelf up, the books off the floor and a partner who would stop mentioning it. The drill was simply the least interesting part of the transaction.
Software is sold as if this were not true. It arrives with feature lists and a pricing page with three columns and a recommended badge. The client is invited to admire the drill. Six months later the drill is in a drawer, the shelf is still not up, and the subscription renews quietly every month like a gym membership for a body that never visits.
ABC is built on the opposite assumption. The goal is the product. What a client buys is not an app, a back office or a team of agents, though they get all three. What they buy is a number that moves: the one goal from the first chapter, with its date, on a scoreboard fed by real evidence, reported every weekday morning. If the number moves, the product worked. If it does not, no quantity of features will rescue it.
This changes how everything in this part fits together. One goal gives the machine a direction. A number with a date makes it measurable. Dropping the other eleven frees the attention to pursue it. The scoreboard and the evidence rule keep it honest. The weekday rhythm gives it a pulse, the attention budget protects the hours that matter, and the green light keeps the machinery underneath in working order. Together they make a machine for moving one number.
That is also why the proposition is the same shape for every client. Their goals differ wildly. One wants a thousand pounds a month of recurring revenue, another wants to stop working Sundays, another wants forty paying members by the spring. The machine does not much mind. It needs a number, a date, real feeds and a human who makes one decision a day. Everything else is configuration.
It is also a more honest thing to sell. A client who buys software can be disappointed in a hundred subtle ways. A client who buys a moving number can be disappointed in exactly one, and they will know about it at 08:45 on a weekday, at the same moment we do. That is uncomfortable for a seller and rather reassuring for a buyer.
So much for the goal. The next part turns to the people who need one: who they are, what a normal Tuesday costs them, and why the tools they already pay for have not helped. The pain, it turns out, is remarkably consistent.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Who It Is For
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
You can usually spot the person ABC is built for by their browser. It has thirty-one tabs open, three of which are the same spreadsheet, and one of which has been playing a muted video since Thursday. They are not disorganised. They are simply doing the work of a department with the staff of a sole trader, and the tabs are where the department lives.
ABC is for the company of one : a business with one decision-maker, whatever the headcount on paper. Sometimes it is literally one person. Sometimes it is one person, a part-time assistant and two freelancers who appear on alternate Wednesdays. Either way, one head carries the whole picture, and that head is full. Four portraits make the point, and they are composites rather than case studies, though you may recognise a friend.
The first is the founder . She has a product, a few hundred users and a plan for a few thousand. She also does the support inbox, the bookkeeping, the social posts and the investor update she keeps postponing. Her problem is not ideas. It is that every idea arrives on the same afternoon as a refund request, and the refund wins.
The second is the consultant. He sells his expertise by the day and is very good at it. Between client days sit proposals, invoices, follow-ups and the slow work of being findable. He earns well in the weeks he is booked and panics politely in the weeks he is not, because the pipeline only gets attention when it is empty.
The third is the landlord with a handful of properties, run from a laptop and a phone that never quite stops. Rents, repairs, renewals, certificates, tenants who message at ten at night about boilers. None of it is difficult. All of it is constant, and most of it lives in a WhatsApp thread that would make an auditor weep.
The fourth is the creator: videos, a newsletter, a shop, some affiliate links and a quietly growing sense that the business side is eating the creative side. They publish beautifully and reconcile badly. Their numbers are spread across five dashboards, none of which agree, all of which want a password.
Put the four side by side and the differences fade. Each has one goal, whether they have written it down or not. Each runs on too many tools and not enough system. Each loses hours a week to the dull middle of the work. And each would pay a reasonable sum to receive, every weekday at 08:45, a short and honest note saying where everything is. That shared shape is why one machine can sit at the centre of all four.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Tuesday Problem
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Mondays have a reputation and Fridays have a mood. Tuesday has neither, which is exactly what makes it dangerous. Nothing is on fire on a Tuesday. No client is shouting, no deadline is screaming, and the coffee is fine. You sit down at nine with a clear head and every intention, and at six you stand up again, slightly puzzled, having done a full day's work and shipped nothing at all.
This is the Tuesday problem , and most small businesses have it four days a week. It is drag: the steady, invisible friction that turns eight hours at a desk into one real hour of work that moves anything. The day was busy. It just was not productive, and busy is very good at disguising itself.
So where do the other seven hours go? Follow a typical Tuesday and they leak out in small, reasonable-looking pieces. Twenty minutes on email before you start, which becomes forty. A quick check of the bank, the store and the analytics, each in a different tab with a different login. A call that could have been a message. A message that needed a call. Finding the file from last week, then finding the newer version of the file from last week. Looking up what was agreed with a supplier, because it was agreed in a chat you cannot now find.
Then there is switching, the most expensive habit nobody budgets for. Every time you move from writing to invoicing to answering a question and back, your head has to reload the whole picture. The reload is not free. Do it thirty times a day and you have spent an hour or two simply remembering where you were. It feels like work. It is actually the overhead of work.
The sad part is that the one real hour is often good. It is where the actual business happens: the decision, the design, the conversation that matters. It is just very small, and it usually comes at the end of the day, when you are tired, because everything else got there first.
ABC does not try to make Tuesday exciting. It tries to make it shorter on the overhead and longer on the hour. The morning status replaces the round of tabs. The agents take the finding, chasing and tidying. The one thing to ship is chosen before nine, not discovered at five. None of this is dramatic, and that is the point. The Tuesday problem is not solved by heroics. It is solved by quietly removing the reasons the day keeps evaporating.
If you can get two real hours out of a Tuesday instead of one, you have doubled the business's output without hiring anyone. That is a modest-sounding change with an immodest effect.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Twelve Tools, No System
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every tool a small business uses was bought for a good reason. The invoicing app is genuinely better than a spreadsheet. The scheduling tool saved hours. The email platform, the store, the project board, the password manager, the form builder and the note-taking app with the lovely fonts all earned their place on the day they arrived. Nobody ever bought a bad tool on purpose. They simply bought twelve good ones, one at a time.
That is how the SaaS pile forms. Each tool is excellent in its own little box. None of them knows the others exist. The customer in the store is a slightly different customer in the email platform, and a third, more out-of-date customer in the invoicing app. Every tool has its own login, its own export and its own idea of what happened last month. Together they are not a system. They are twelve neighbours who have never been introduced.
The diagram for this is a stack, and it gets heavier as it goes down. At the top, twelve logins. Beneath them, twelve exports in twelve formats. Beneath those, the human glue: you, copying and pasting between them, reconciling the differences by eye. And at the bottom, the thing the whole pile was supposed to provide and does not: a single truth . Ask how many active customers you have and you will get three answers, depending on which tab you open first.
The odd consequence is that more tools make a small business slower. Each new subscription promises to save time, and locally it does. But it also adds one more place to check, one more set of numbers to reconcile and one more password to reset at the worst possible moment. Past a certain point the time spent maintaining the tools exceeds the time they save, and the business is running its software rather than the other way round.
The usual response is to buy a thirteenth tool that connects the other twelve. Sometimes this helps for a while. More often it becomes the most fragile thing in the building, a web of automations that nobody fully understands, which breaks quietly on a Friday and is discovered on a Wednesday.
ABC approaches the pile differently. It does not ask you to throw your tools away, because many of them are good and some of them hold your customers. It asks which one holds the truth for each kind of thing, and then builds one back office that reads from all of them, keeps one copy that wins and reports from that copy alone. The tools become feeds. The back office becomes the system. The details of how belong to a later part of the book.
For now, the diagnosis is enough. You do not have a tool problem. You have twelve good tools and no one in charge of them.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Copy-Paste Tax
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Somewhere in every small business there is a person whose real job, unlisted on any contract, is to be the bridge between two pieces of software. They open one tab, highlight a number, press two keys, open another tab and press two more. They do it with the calm of someone who has stopped noticing. It is the most expensive calm in the building.
Call it the copy-paste tax . Every tool in the pile from the last chapter is good at its own job and hopeless at talking to its neighbours, so the talking is done by hand. An order arrives in the shop and gets typed into the spreadsheet. The spreadsheet total gets copied into the invoice. The invoice number gets pasted into an email, and the accountant types it into something else. Each move takes a minute. Nobody calls a minute a cost.
So price it honestly, with an illustrative week rather than a heroic one. Say fifteen small transfers a day, two minutes each, plus the occasional ten minutes spent working out why two numbers disagree. That is the best part of an hour a day, roughly four or five hours a week, which is more than half a working day spent as a human clipboard. Your figures will differ. Keep a tally on a sticky note for one week and you will probably be surprised in the wrong direction.
The hours are only the visible half. The other half is that data moved by hand arrives late and slightly wrong. A digit transposes. A row gets pasted into the row below. A price is updated in one place and not the other, and for a fortnight the business quietly holds two opinions about what it charges. None is a disaster. Together they produce a low hum of distrust, where every number has to be checked before it can be used, which is the copy-paste tax charging interest.
There is also a tax on attention, and it is the steepest of all. Each transfer is a small context switch: leave the thing you were thinking about, become a clerk for ninety seconds, try to return. A morning broken into fifteen pieces is not a morning. It is fifteen short visits to your own work.
ABC treats every recurring copy-paste as a bug report. If a number travels from one place to another on a schedule, a machine should carry it, and the back office should be the place it arrives. The first job is not clever automation. It is to list the transfers, with their frequency, and fix the top three.
The clipboard was designed to hold one thing for a moment. It was never meant to be your integration strategy.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Context Lost Overnight
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every morning, a small business wakes up with mild amnesia. It knows its name and roughly what it sells. It is much less sure what it decided on Thursday, why the price changed in March, or what exactly was promised to the customer who rang at ten past five. The owner remembers some of it. The owner is not a filing system, however hard they try.
Look at where yesterday actually went. A decision was made in a chat thread, in between two jokes and a photograph of a sandwich. The reasoning behind it lives in a voice note. The follow-up sits in a notebook that is currently in a coat pocket. A half-finished plan is in a document called new plan FINAL 2 , and an important caveat is in an email reply that nobody will ever search for. Five places, none of them wrong, none of them readable as a whole.
This is context lost overnight , and its symptom is a particular kind of déjà vu. You find yourself having the same discussion you had three weeks ago, reaching the same conclusion, and feeling faintly proud of it. Nobody remembers the first time, so the business pays for every decision twice: once to make it and once to make it again. Sometimes three times, if a holiday falls in between.
It gets worse the moment anybody else is involved. A freelancer starts on Monday and spends the first morning asking questions whose answers exist only in your head. A colleague covers for a week and makes a perfectly sensible choice that contradicts one you made in January. Neither is at fault. They are reading a business that has not been written down, which is like reading a novel where every third page is kept in a different house.
Now add agents. An AI agent arrives each morning with no memory of yesterday unless someone has given it one in writing. It is the most willing newcomer you will ever hire and also the most forgetful. Ask it to draft the pricing email and it will cheerfully invent a pricing policy, because the real one is in a voice note it cannot hear. The agent is not careless. It simply has nothing to read.
The fix is not a better memory. It is a smaller number of places, ideally one, where decisions, rules and reasons are written in plain text that a person or a machine can read before acting. Part V describes how ABC builds that place. For now it is enough to notice the cost of not having it, which is measured in repeated conversations and confident mistakes.
Yesterday is the cheapest thing a business owns. It is a pity to keep throwing it away.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The AI Toy Box
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most small businesses now own a surprising amount of artificial intelligence. There is a chatbot subscription, an image generator someone tried at Christmas, a writing assistant built into the email client and a meeting recorder that produces summaries nobody opens. Ask what any of it has shipped this month and the room goes quiet in a thoughtful sort of way.
This is the AI toy box . Each tool is impressive. Each was bought, or at least signed up for, with genuine excitement. And each is used the way people use a slot machine: open the window, type a hopeful sentence, pull the lever and see what comes out. Sometimes it is brilliant. Sometimes it is nonsense delivered with total confidence. Either way, the result gets copied somewhere by hand, and the window is closed until the next time somebody feels lucky.
The problem is not the models. The models are remarkable, and they get better every few months. The problem is that a tool sitting in a separate tab, with no access to your data, no knowledge of your goal and no place to put its output, is not part of the business. It is a very clever visitor. It knows nothing about your customers, cannot see your back office and has never read the decision you made on Thursday. So it guesses, politely and at length.
The test that separates a toy from a teammate is short. Is it wired into the work ? A teammate reads the same context you do before it starts. It works on real records, not on text you pasted in. Its output lands somewhere useful: a draft in the approvals queue, a pull request, a row in the ledger, a line in the morning status. And it can be checked afterwards, because there is evidence of what it did. A toy does none of these things. It entertains, occasionally helps, and leaves no trace.
There is also the matter of repetition. A toy has to be asked every time, in a slightly different way, by someone who remembers how. A teammate has a written procedure, a skill it loads, so the fortieth invoice reminder is drafted the same way as the first. Consistency is boring, which is why nobody puts it in the advertisements, and it is precisely what a business needs.
None of this means throwing the toys away. Some will graduate, once they are given a job, a procedure and a place to report. Most will turn out to have been a hobby, and a hobby is fine, provided nobody mistakes it for a plan.
A toy box is fun on a Saturday. On a Tuesday, you want colleagues.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Why Agencies Disappoint
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
At some point every growing small business decides to call in professionals. An agency is found, a pleasant meeting held, and a discovery phase begins. Six weeks later there is a slide deck with personas on it, one of whom looks suspiciously like the owner, and an estimate that would buy a reliable used car.
None of this is villainy. Agencies are built for clients with budgets, project managers and an IT department to receive the finished work. Their model works when the client can absorb a long discovery, a large invoice and a formal handover. A company of one can absorb none of those things. It has no committee to sit through the discovery, no spare quarter of revenue for the invoice and nobody to catch the handover.
The handover is where it usually goes wrong. The agency delivers something polished, explains it in one long call and moves on to the next client. The owner is left holding a system they did not build, in a codebase they cannot read, hosted on an account they are not quite sure they control. Three months later something small breaks. The agency is happy to fix it, at its day rate, starting the week after next. The software was delivered. The business did not really receive it.
Plot the market on two axes and the problem becomes obvious. On one axis, cost . On the other, ownership : how much of the result you can run, change and keep without asking anyone. Agencies tend to sit at high cost and low ownership. Off-the-shelf software sits at low cost and low ownership, because you rent it and it bends to nobody. A brave few build it all themselves and lose their evenings for a year. The corner that small businesses actually need, low cost and high ownership with the work done for them, is mostly empty.
That corner is where ABC is aimed. Discovery is an onboarding interview that takes minutes, not weeks, and it produces a short book about your business that you can correct with a pen. The first build is deliberately small, a V1 that runs one workflow well. Everything is built in your own repositories, with written specs, readable history and skills you can open and edit. And instead of a handover, there is a morning status every weekday, so the system never becomes something that happened to you once.
This is not an argument that agencies are bad. Many are excellent at what they are designed for. It is about fit. A small business does not need a bigger project. It needs a smaller one that it keeps.
The best handover is the one that never has to happen, because you owned it all along.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Why Now
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every few years someone announces that small businesses are about to be transformed by technology, and every few years small businesses carry on much as before, only with more passwords. So it is fair to be suspicious of a chapter called Why Now . The honest answer is that three separate things changed, and they changed at roughly the same time.
The first was the models . Through 2025 and into 2026, language models went from impressive conversationalists to capable workers. They can now read a long brief, follow a written procedure, write and fix real code, and say when they are unsure, at least most of the time. They still make mistakes, confidently. But they make far fewer, and the mistakes are of a kind that checks and tests can catch.
The second was the tools around the models. A clever model in a chat window is a toy, as we saw two chapters ago. What arrived in this period was the wiring: agents that can open a repository, run the tests, propose a pull request, read files, use a browser and hand work to one another. Skills became a normal idea, a written procedure an agent loads like a recipe. Suddenly the model was not just answering questions. It was doing jobs, with evidence.
The third was cost . Building a small custom app used to mean weeks of a developer's time, priced accordingly. Now an agent can do much of the build in an afternoon, under supervision, and hosting a small app costs pounds rather than hundreds. When the cost of an experiment falls that far, the arithmetic of a small business changes. You no longer have to bet the year on one big system. You can afford a fleet of small ones, and you can afford to throw away the ones that do not work.
Put those three together and something genuinely new becomes possible. One person, with a clear goal and a well-written context, can run the kind of operation that used to need a team: someone to build the apps, someone to keep the books tidy, someone to write the content, someone to send the morning report. The agents do the middle of that work. The person does the choosing, the approving and the parts that need taste.
None of this removes judgement. If anything it raises its value, because the bottleneck is no longer labour. It is deciding what is worth doing, which is why this book began with one goal rather than with software.
Waiting is still an option. The pain will simply be a year older, and slightly better organised in your head.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What Good Looks Like
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
It is worth describing the destination before the rest of the book describes the engine. Not a vision, not a mood board, just an ordinary Wednesday, after the machine is running. It is less dramatic than you might hope. That is the good news.
The day begins at 08:45, when the morning status arrives. It is short and it is the same shape as yesterday's. One line for the goal: today's figure, the target and the distance between them. Then what moved overnight, each line with a link to its evidence: a merged pull request, a live page, a file in the back office. Then what is stuck. Then the one thing that needs you. You read it with the first coffee, which is still hot when you finish.
The thing that needs you is one decision . Perhaps an agent has drafted the clips for a new app and wants a yes before they are scheduled. Perhaps the back office has matched an affiliate statement and found a payment it cannot explain. The decision arrives framed: what is being asked, the recommended option, the evidence, and what happens if you say no. You read it, change a word, and approve. It takes a minute.
Then the morning is yours, and this is the part people find strangest. There is no inbox triage disguised as work, no hunting through tabs for the latest version, no copying numbers between tools. Those jobs were done overnight or are being done now, quietly, by agents that will report back rather than interrupt. You spend the morning on the work that needs a human: a conversation with a customer, a piece of writing, a hard think about next month.
By noon, one thing has shipped . Not ten. One. A new app is live at its URL, or a feed has been connected to the scoreboard, or a chapter of something has been published. It is small, it is real and it can be linked. Over a year, one shipped thing a working day adds up to a quietly alarming amount.
The afternoon has room in it. You might go for a walk, which is often more productive than doing more. The machine does not need supervising in the way staff sometimes do; it needs direction, and it got that at 08:45. If something breaks, it will say so in plain words, and red will not survive the working day.
In the evening you switch off, and this is the real measure. Not hours worked, but whether you can close the laptop without a low background worry that something important is slipping. When the status is honest and the gates are clear, there is nothing left to worry about until tomorrow at 08:45.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Cost of Doing Nothing
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Doing nothing has an excellent reputation. It is cheap, it is safe and it never needs a meeting. Nobody has ever been blamed for keeping things as they were. That is precisely what makes it dangerous, because staying as you are is also a decision, and like every decision it has a price. It simply sends the invoice in small pieces.
This part of the book has been an itemised version of that invoice. The Tuesday that drags. The twelve tools with no system. The copy-paste tax, charged by the minute. The context that evaporates overnight and gets rebuilt at the cost of a second conversation. The AI toy box, paid for monthly and used occasionally. The agency project that arrived polished and could not be run. Each line on its own looks affordable. That is how it was allowed onto the bill.
So add it up, honestly and for yourself. Take the hours lost each week to handling, hunting and re-deciding, and multiply them by what an hour of your attention is genuinely worth. Add the subscriptions nobody uses. Add the sales that slipped because the second follow-up was never sent. Purely as an illustration, five lost hours a week at a modest hourly worth soon becomes a monthly figure larger than most software budgets. Your sum will differ. It is very unlikely to be zero.
The money and the time overlap, and the overlap has a name: drift . Drift is what happens when a business is busy every day and no further forward at the end of the month. Nothing has gone wrong. There is no crisis to point at. The goal simply stays the same distance away, week after week, while everyone works hard and feels vaguely tired. Drift is the most expensive state a small business can be in, because it does not feel like a cost at all. It feels like normal.
The Stoics had a habit of imagining the future plainly, without drama, to see what their present choices were buying. Try it here. Picture next October with nothing changed: the same tabs, the same clipboard, the same Tuesday. If that picture is fine, keep it. Many businesses are genuinely fine. But if it makes you wince slightly, the wince is information.
The good news is that the cure for drift is not a grand transformation. It is a machine, built in small pieces, each one earning its keep before the next is added. The first of those pieces is the most visible, and the easiest to start with on a quiet morning.
Part III begins with the apps: small, single-purpose and expected to pay their way. Nothing on the bill shrinks until something ships.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
A Is for Apps
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
When people hear that a business runs on apps, they tend to imagine one enormous application: a platform, with a logo, a roadmap and a launch party. ABC means almost the opposite. The A in ABC stands for a fleet : many small applications, each built quickly, each doing one job, and each expected to justify its existence in plain numbers.
Mat's own fleet lives in a single folder, ~/APPS, and reads like a sensible shopping list. There is a lists app, a travel app, a scoring app, a couple of experiments in personal and contextual AI, and the cockpit that watches over the rest. Around those sit the scrollers: small, phone-first apps that turn a ranked dataset into something you flick through on a train. None of them is trying to be everything. Each is trying to be useful to one kind of person for one kind of moment.
So what counts as an app in ABC? The definition is deliberately modest. An app has one job , describable in a sentence without the word "and". It has a live URL that a stranger can open. It has a single source of data behind it, so that the site, the clips and any book drawn from it all agree. It has an entry in the registry, with a status and an owner. And it has a number it is held to, which for a fleet app is the gloriously unambitious target of a pound a day. Without all of these, it is not yet an app. It is an idea that has had a nice afternoon.
Why small? Because small apps are cheap to build, cheap to run and cheap to be wrong about. A large platform is one bet, placed once, with your whole year on it. A fleet is many bets, each small enough to lose without noticing. The ones that work get more attention. The ones that do not are retired without ceremony, and the lessons go into the next build.
There is a second reason, which is less obvious. Small apps are easy for agents to build and maintain. A single-purpose app with one dataset and a written spec fits comfortably inside an agent's working memory. A sprawling platform does not, and that is when agents start guessing. Keeping each app small keeps the machine honest.
None of this rules out ambition. The ambition simply lives in the fleet rather than in any single ship. A hundred modest apps, each earning a little and each feeding the same back office, add up to something no single platform could promise with a straight face.
The rest of this part follows one app from idea to income. It starts, as everything in ABC does, with something small enough to finish.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Small Apps, Small Bets
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every founder has met the large bet. It arrives as a platform, a marketplace or "the Uber of something", it needs eighteen months and most of your savings, and it is usually described in the future tense. It may well work. You will only find out at the end, when finding out is expensive.
ABC prefers the opposite shape. Instead of one big bet, it places many small bets : single-purpose apps, each built in days rather than months, each costing roughly the price of a decent dinner to run. A ranked guide to travel adapters. A list of the hundred best board games for two. A tool that does one sum well. None will change the world. Several might quietly pay for themselves, which is much easier to arrange.
The maths is pleasingly simple. If one large bet has a modest chance of success, you have a modest chance of success. If a hundred small bets each have a modest chance, you have a very good chance that some of them land, and you learn something useful from the ones that do not. The failures are cheap. The winners are real. And you find out which is which in weeks, not years.
It helps to picture a simple square. Along one edge runs the cost to try; along the other runs the upside. Large platforms sit high on both: expensive to try, enormous if they work. Hobbies sit low on both. Small apps sit in the corner everyone wants and few people aim for: cheap to try, real upside . Not a fortune each. A real, countable amount, which is the only kind the scoreboard accepts.
There is a quieter benefit too. A large bet ties your mood to a single outcome. Every slow week feels like a verdict. A fleet of small bets spreads that weight. When one app stalls, another is climbing, and the morning status reports both without drama. You stop asking whether you are a success and start asking which of your experiments is working, a far better question for breakfast.
None of this means building carelessly. Each small app still gets a governed dataset, a proper deploy, green CI and a line in the registry. Small is a statement about scope, not about standards. A tiny app that is broken is not a small bet; it is a small embarrassment, and the machine is built to prevent those.
The Stoics advised holding outcomes lightly while doing the work well. Small bets make that advice practical. You cannot control which app the world will like. You can control how many honest attempts you make, how cheaply you make them and how quickly you read the results. Do that for a year, and luck starts to look suspiciously like method.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Seed Dataset
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most apps begin with a logo. Someone opens a design tool, spends an afternoon choosing between two blues, and then discovers in week three that they do not actually know what the app contains. The logo is lovely. The content is a placeholder that says "Lorem ipsum" with great confidence.
ABC begins at the other end, with the data. Every fleet app starts life as a seed dataset : a single file, usually called seed.json, holding somewhere between fifty and a hundred carefully chosen items. If the app is about camping stoves, the seed is the camping stoves, each with a name, a rank, a short description, the facts that matter and the link that might one day earn a commission. Next to it sits a one-page app card saying who the app is for and what question it answers.
The seed is governed , which is a slightly grand word for a simple idea. Every item has the same fields. Every fact has a source. Nothing is invented to fill a gap; if a value is unknown, the field is left empty, for the same reason the scoreboard prefers null to a guess. The seed is reviewed like code, changed through pull requests and kept in the repository where the history shows who altered what and why.
This sounds like a lot of care for camping stoves. It is the right amount, because the seed is not one output. It is the source for all of them. From that one file the machine builds the scroller, a vertical feed of cards you can flick through on a phone. It builds the website, with pages, search. It writes the script for the short clips that bring visitors in. It can even become a small book. Four surfaces, one dataset, and not a single fact typed twice.
That last point is where the savings live. When a price changes or a product is discontinued, you fix one line in the seed and every surface updates on the next build. Compare that with the usual arrangement: the website says one thing, the video another and the original spreadsheet a third, and nobody remembers which came first.
There is a creative benefit too. Choosing the hundred items is where the taste goes. It forces the real editorial decisions early: what counts, what ranks higher, what gets left out. Those decisions make an app worth visiting, and they are easier to make in a plain file than in a half-built interface with a lovely blue logo.
So when ABC proposes a new app, the first deliverable is never a design. It is a file you can open, read and argue with. Get the seed right, and the rest of the app is mostly a matter of watering it.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The App Factory
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A factory is a slightly unfashionable word for a small business. It conjures smoke and someone in a hard hat shouting about quotas. The ABC version is quieter. It has no chimney, it runs on weekdays and most of its workers are skills that an agent loads when it needs them. But it is a factory in the sense that matters: the same steps, in the same order, every time.
The App Factory takes a niche idea and turns it into a deployed app that earns. It has five stations, and each one is a named skill. The first is /abc-seed, which builds the governed dataset from the previous chapter. The second is /abc-scroller, which pours that seed into a ready-made shell: a vertical, phone-first feed of cards with a progress rail, categories and a place for the calls to action. The third is the build, where the prototype becomes a proper Next.js app with its own repository, tests and deploy. The fourth is /abc-monetise, which adds the revenue layer. The fifth is /abc-clip, which writes the short videos that bring visitors to the door.
The order is not decorative. Each station consumes what the previous one produced and adds exactly one thing. The seed gives you content. The shell gives you something to look at by teatime. The build makes it real, reachable at a live URL that anyone can open. Monetising comes after that, once there is something worth paying attention to, and clips come last, because traffic to an unready app is wasted.
What makes it a factory rather than a series of good intentions is that every station has a definition of done . The seed is done when the items validate. The shell is done when it renders on a phone. The build is done when CI is green and the URL loads. Monetising is done when the links are tagged and the disclosures are in place. Clips are done when the scripts exist and the calendar is set. Nobody has to decide whether a step feels finished. The skill already knows.
This removes the most expensive part of building small apps, which is not building. It is deciding. Without a factory, each new app reopens the same questions: which framework, which layout, where the links go, how to announce it. With one, those questions were answered once, written down and turned into skills. The founder's job shrinks to the only question a machine cannot answer well: which niche next?
The first app through any pipeline is slow, and that is fine. The tenth is noticeably quicker. By the thirtieth, the hard part is choosing which ideas deserve a seed, which is exactly where a human's attention should be spent. Everything downstream of that choice can run on rails.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
A Pound a Day
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask a founder what their app should earn and you will usually hear a number with a lot of zeros and the word "eventually". Ask ABC and you will hear something that sounds almost rude: a pound a day . One pound. Roughly the price of a disappointing biscuit, every day, from every app in the fleet.
It is a deliberately modest target, and the modesty is doing real work. A pound a day is achievable by a small, useful app with a few hundred interested visitors a month and a handful of affiliate clicks. It does not require virality or a funding round. It requires the app to be genuinely helpful to a narrow group of people and to stay switched on. That is a bar most honest apps can clear, which is exactly why it makes a good bar.
Now multiply. One app at a pound a day is a biscuit habit. A hundred apps at a pound a day is a hundred pounds a day, which over a year comes to £36,500 . That is not a fortune, but it is a salary, built from parts so small that no single one of them can sink you. If an app falls to zero, the fleet barely notices. If an app climbs to five pounds a day, you have learnt something and you build three more like it.
These figures are illustrative, of course. The scoreboard does not care what this chapter says; it reports what the affiliate statements and ad accounts actually paid, and until a feed is connected it reports nothing at all. But the shape of the arithmetic holds, and it changes how you think about each new app. You stop asking "could this be big?" and start asking "could this reliably earn a pound?", which you can answer before lunch.
The pound a day is also a test of whether an app deserves its slot. Every app costs something to keep alive: hosting, monitoring, a line in the registry, a little attention in the monthly review. An app that cannot clear a pound a day after a fair trial is costing more than it earns. The target gives that feeling a number.
There is a Stoic comfort in this. Seneca warned against letting the size of an ambition run ahead of the work in front of you. A pound a day keeps the work in front of you. It is small enough to aim at every morning and, multiplied, large enough to change a year.
Some apps will beat it handsomely. Some will never reach it. The fleet does not need heroes. It needs a great many dependable little earners, each doing one useful thing, each quietly sending its pound home. Boring, it turns out, compounds rather nicely.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Registry
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every small business has an app it has forgotten about. It was built in a burst of enthusiasm, deployed somewhere, and is still running on a server you pay for, at an address nobody can remember, earning nothing and serving three bots. It is not a failure, exactly. It is a ghost, and ghosts are quietly expensive.
ABC deals with ghosts by keeping the registry . Every app in the fleet has an entry, and the entry is the app's official existence. In Mat's own machine it lives in a plain data file inside the ABC OS cockpit, alongside the fleet in ~/APPS: mscore, mslists, mstravel, agentai, personai, aicontext2026, ais and abc itself. The rule is short enough to remember on a bad day: if it is not in the registry, it does not exist . It is not monitored, not reviewed and certainly not credited on the scoreboard.
An entry is built in layers, and each layer answers a practical question. At the top sits the app itself: its name, what it is for and who owns it. Below that sit its port and its URL, so anyone can find it and nothing collides with anything else. The cockpit itself sits on port 25000, and every app has its own number in the same orderly fashion. Below that sit status and health: is it live, is CI green, did the last deploy succeed. At the bottom sits the revenue feed, the connection to wherever the money is actually measured.
That bottom layer is the point of the whole stack. A registry entry that stops at "live" tells you the app is switched on. One that ends in a revenue feed tells you whether being switched on is worth anything. Until the feed is connected, the revenue field is empty rather than estimated. You will recognise the principle by now.
The registry also makes the fleet readable by agents. When ABC Master writes the 08:45 status, it does not need to remember which apps exist. It reads the registry, checks each entry's health and reports the ones that need attention. When a new app comes out of the factory, adding its entry is part of the definition of done. An app that ships without one has not really shipped.
There is a tidiness to this that is easy to underrate. A founder with twenty apps and no registry has twenty small worries. A founder with twenty apps and a registry has one table, which can be read in a minute and trusted completely. The worries do not vanish, but they line up, and worries in a line are easier to deal with.
Ghosts, meanwhile, get an exorcism. Any app found running without an entry is either registered properly or switched off. Most are switched off. Nobody misses them.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Scale, Hold or Kill
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The hardest thing to do with an app you built is switch it off. You remember the evening you made it. You remember the clever bit in the layout. You suspect, against the evidence, that it is about to take off, possibly after one more small change. This is how fleets fill up with apps that are loved, maintained and earning nothing.
ABC handles this with a monthly review and three verdicts. Once a month, every app in the registry is put through the same short hearing. The first question is the only one that really matters: is it earning? Not "is it promising", not "do people seem to like it", but is the revenue feed showing money coming in against the pound-a-day target. The answer is a number, read off the scoreboard, not recalled from memory.
From that one question, the verdicts follow. Scale is for apps that are earning, or clearly climbing towards it. They get more: more items in the seed, more clips, a sister app in a neighbouring niche. Hold is for apps that are young or close, given another month with one specific change to try, written down so the next review can see whether it worked. Kill is for apps that have had a fair trial and did not earn. They are switched off, their entry is marked as retired, and a short note records what was learnt.
The important word in all of this is fair . A new app is not judged in its first fortnight; it needs time for search engines to notice it and for clips to bring visitors. But a fair trial has an end date, set when the app ships, and it is not extended because somebody is fond of the colour scheme. Sentiment is welcome elsewhere in life. In the review it is not a field.
Killing an app sounds brutal. In practice it is a relief. A retired app stops costing money, stops appearing in the morning status as a small grey worry and stops occupying your attention. The seed is kept, so nothing is truly lost, and the lesson goes into the context so the factory gets a little wiser about which niches to try next.
The review takes surprisingly little time. With a registry and real revenue feeds, ABC Master can prepare the table before you sit down, with a suggested verdict for each app and the evidence beside it. You agree, disagree or ask a question. Most months, for most apps, you agree.
Marcus Aurelius reminded himself that everything he loved was on loan. App owners could do with the same reminder. You built it, you tried it, it did not earn. Thank it, switch it off and spend the attention somewhere it will be repaid.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Clips as Distribution
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The traditional way to get visitors to a new app is to pay for them. You set up an advertising account, choose a budget, watch it vanish at an alarming pace and receive in return a dashboard full of numbers that look encouraging until you divide one by another. For a fleet of apps meant to earn a pound a day each, that arithmetic never works.
ABC uses clips instead. Every app that leaves the factory comes with fourteen short vertical videos, made for the places people actually scroll: TikTok, YouTube Shorts and their cousins. Fourteen is not a magic number. It is two weeks of one clip a day, which is long enough to learn something and short enough to finish. The skill that writes them is /abc-clip, the last station in the factory, and it works from the same seed as everything else.
That is what makes the clips cheap. Nobody sits down to invent fourteen ideas. The seed already holds a hundred ranked items, each with facts, a description and a reason to care, so the scripts almost write themselves. Clip one might be the top three. Clip two might be the surprising entry at number forty. Clip three might settle an argument the niche is always having. Each clip has a hook in the first second, a few quick beats and one clear call to action, which is nearly always "the full list is at this address".
The footage comes from the app itself. Pages are captured as the scroller moves, cards slide past, rankings change, and the result looks like what it is: a real, working thing you can visit. No stock video of people laughing at salads. The clips come with captions and hashtags, a posting calendar and a note of what each one is trying to test.
That last part matters, because clips are not only distribution. They are a feedback loop. A clip brings visits. Visits show which items people tapped, which pages they stayed on and which links they clicked. That tells you what the next clip should be about, and occasionally what the next app should be about. The loop runs every day of the calendar, and the monthly review reads its results.
There is a temptation to judge clips by views, because views are the number the platforms put in large type. Resist it. A clip with modest views that sends fifty people to the app is worth more than one with a hundred thousand views that sends nobody. The scoreboard counts what arrives, not what was watched.
None of this requires you to become a personality, dance or point at floating text. The clips are about the subject, not about you. That is a considerable relief to most founders, and to most audiences.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Monetise Last
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There is a particular kind of website that greets you with a cookie banner, a newsletter pop-up, a sticky video advert and a chat bubble asking whether it can help, all before you have read a word. It has monetised everything except the reason you came. You close it, and so does everyone else, which is why it needs so many adverts.
ABC takes the opposite approach and calls it monetise last . An app earns the right to make money by being useful first. Until it has real content and real visitors, it carries no adverts, no affiliate links and no sign-up forms. Revenue is a reward for usefulness, not a feature bolted on at the start to make up for the lack of anything else.
When the time comes, the order is fixed, and part of the method. First come affiliate links, starting with Amazon Associates, because they fit naturally into a ranked list of things people might buy and they pay only when someone was genuinely helped to a decision. The link sits beside the item, with the facts, and nobody minds it there. Second comes advertising, through AdSense, once the app has enough pages and traffic to pass a readiness audit. Third comes email capture, the slow one, which turns occasional visitors into people you can tell when the list changes.
Each layer is added by the /abc-monetise skill, and each comes with its own housekeeping. Affiliate links are tagged correctly for each country. Disclosures are written in plain words and placed where they will be read, not hidden in a footer. Applications to each programme are tracked, because programmes have rules and the rules change. None of this is glamorous. All of it is the difference between earning and being quietly suspended.
The fixed order also stops a common mistake, which is adding everything at once and learning nothing. When affiliate links go in alone, you can see what they do. When ads follow a month later, you can see whether they help or merely annoy. Change one thing at a time and the scoreboard can tell you which change mattered.
There is an old shopkeeper's rule that you look after the customer and the till looks after itself. It is not quite true; tills need watching. But it points in the right direction. People return to things that help them, recommend things that help them and, now and again, buy through things that help them. An app that annoys them gets none of the three.
So the fleet earns its pounds the patient way. Useful first, then visited, then linked, then measured. It is slower in the first week and considerably faster in the first year, which is the only timescale a pound a day was ever meant to be judged on.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Fleet as an Asset
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A single app is a gamble. Ten apps are a habit. A hundred apps, each built from the same factory and held to the same modest target, start to look like something an accountant would recognise: an asset. This chapter closes Part III by making that small change of word, because it changes how you treat everything that came before.
Most people judge an app the way they judge a lottery ticket, by whether it won. That is the wrong frame for a fleet. You do not judge a portfolio by its best holding or its worst. You judge it by the total and by whether it is quietly getting bigger. A fleet is a portfolio , and portfolios reward patience, spread and a refusal to fall in love with any single position.
The compounding comes from three directions at once. Traffic compounds, because every page that ranks today keeps ranking tomorrow without a fresh invoice. Clips compound, because fourteen short videos per app keep being found long after the week they were posted. And the factory compounds, because every app it builds teaches it something, so the next seed is cleaner, the next shell is faster and the next monetise step is already written down. The hundredth app costs a fraction of the first.
This is why the £1 a day target matters more than it sounds. One pound is not a business. A hundred pounds a day, from a hundred small things that each need almost nothing from you, is a rather different conversation. None of them is heroic. Together they are the kind of income that does not vanish when one platform changes its mind on a Thursday.
There is also ownership. A company of one cannot own a factory floor or a fleet of lorries. It can own code, domains, datasets, rankings and an audience. Those are real things, with real value, and they sit in your registry with a status and a revenue line. You can count them, and one day you could sell some. Most people who work for themselves end the year owning nothing but a tired diary.
None of this works if the numbers are fiction, and that is the hinge into the next part. A portfolio is only an asset if you can prove what it holds and what it earns: which apps exist, which domains they sit on, which affiliate statement paid for which click. That proof does not live in the apps. It lives somewhere much less glamorous, a single back office with one copy of the truth.
So Part III ends with a fleet and Part IV begins with a filing system. That is the right order. First you build things worth owning, then you build the place that remembers you own them.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
B Is for Backoffice
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask a small business owner where the logo lives and you will learn a great deal about their week. It is in an email from 2023. Or the designer has it. Or it is on the old laptop, the one in the drawer, which still works if you hold the charger at an angle. The logo is not lost, exactly. It is simply everywhere except where you need it.
The second pillar of ABC exists to end that particular game of hide and seek. B is for Backoffice : one place where every asset, number and file the business owns is kept, labelled and findable by a person or an agent in under a minute. Not a shared drive with good intentions, but a system with a structure written down.
In Mat's own machine the back office is called bo28. It holds four kinds of thing, and the agent that tends it is called boai. Assets: the apps, domains, datasets, images and documents the business owns. Data: the tables that describe them, from the app registry to the list of affiliate programmes. Finance: every invoice, payout and statement, matched into one ledger. Files: the originals, stored once, with a record pointing at them. Each gets its own chapter; for now, they all sit under one roof.
Why one roof? Because a company of one cannot afford a second. Every extra place where things are kept is another place to check, another login to remember and another version of the truth to disagree with the first. Twelve tools feel like choice. In practice they feel like twelve small cupboards, each slightly too full, and none of them labelled.
There is a quieter reason too. The agents need somewhere to look. An agent asked to report this week's revenue, or find the hero image for a new app, or check which domain renews next month, is only as useful as the place it searches. Point it at a tidy back office and it answers in seconds. Point it at your inbox and it will answer anyway, confidently, and possibly wrongly.
None of this is glamorous, and that is deliberate. The apps are where the fun happens. The context is where the thinking happens. The back office is where things are put away properly, so that the fun and the thinking have something solid to stand on. It is the cupboard under the stairs, rebuilt by someone who likes labels.
By the end of this part you should be able to picture your own version: one place, four kinds of thing, a scoreboard fed from real numbers and an agent that keeps it all in order. Then the logo will be exactly where you left it, which for most businesses would be a small miracle.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
One Source of Truth
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every small business has a spreadsheet called something like "Apps FINAL v3 (Mat's copy)". It sits next to "Apps FINAL v2", which is newer, and "apps-list-USE-THIS-ONE", which nobody has used since the spring. Each is mostly right. Together they are a committee, agreeing only that someone else should sort it out.
The back office sorts it out by insisting on one source of truth : for every kind of thing the business tracks, there is exactly one table that is allowed to be right. Apps live in the registry. Money lives in the ledger. Domains live in the domains table. Every other copy, the export, the dashboard, the slide in the pitch deck, is a view of that table and is allowed to be wrong only for as long as it takes to refresh.
This sounds obvious until you try it, at which point you discover the real question is not whether to have one truth but which copy gets the job. The rule ABC uses is simple: the copy closest to the event wins . Revenue is true where the payment happened, so the store or the processor is the source and the back office records what it reports. An app's status is true where the app is deployed, so the registry reads it from the deployment, not from someone's memory. When two copies disagree, you do not average them or pick the more flattering one. You walk back towards the event until you find the one that was there.
Getting from twelve spreadsheets to one table happens in three moves. First, gather every copy and admit how many there are. Second, write down a single schema : the columns, what each one means and which ones are required. This is the step that does the real work, because most disagreements turn out to be two people using the same word for different things. Third, import once, retire the old copies loudly and make the new table the only place edits happen.
The retiring matters more than the importing. A spreadsheet left lying around is a spreadsheet someone will update, kindly and in good faith, and within a month you have two truths again. So the old copies are archived with a note at the top saying where the real one lives. A small act of tidiness, a great deal of argument saved.
For the agents, one truth is not a nicety. It is the difference between an answer and a guess. An agent asked how many apps are live will look somewhere. Give it one table and it will be right. Give it twelve spreadsheets and it will be confidently, articulately wrong in a different way each morning, which is somehow worse than silence.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Assets, Not Files
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Somewhere on your computer is a file called IMG_4471.jpg. It might be the best photograph your business owns. It might be a blurred picture of a parking meter. You will not know until you open it, and you will not open it, because there are four thousand others just like it. That folder is not a library. It is a very well-organised landfill.
The back office draws a firm line between a file and an asset . A file is a thing: some bytes with a name and a date. An asset is a thing with an owner, a status and a use. The photograph becomes an asset the moment someone records that it belongs to the mstravel app, that it is approved for use, that it is licensed for the web, and that it is the hero image on a particular page. Same bytes. Entirely different object.
What turns one into the other is metadata , the small set of facts written down beside the file. In ABC the list is short on purpose: what it is, who owns it, which app or project it serves, its status, where it is used and where the original lives. Nobody fills in thirty fields, however good the intentions in January. Six fields, filled in every time, beat thirty fields filled in once.
The record lives in the back office; the bytes live in storage. The two point at each other. It sounds fussy, but it lets the business answer useful questions in seconds. Which images are unused? Which documents belong to an app we killed last month? Which datasets feed more than one surface? No folder can answer those questions. A table answers all of them.
It also changes what happens when things move. A file that is renamed or shifted breaks every link that pointed at it, silently, usually on a Friday. An asset that moves simply has its location updated in one place, and everything that refers to it carries on. The asset has an identity; the file merely had an address.
The agents are the main beneficiaries. Asked to find a hero image for a new fleet app, an agent can search assets by status and licence and return three good candidates with their owners and their history. Asked the same thing of a downloads folder, it will return IMG_4471.jpg and a hopeful tone of voice.
None of this requires heroic tidying of the past. Start with whatever you touch this week. Each time a file earns its keep, give it its six fields and let it graduate. The landfill will not vanish overnight, but the library beside it will grow, one useful thing at a time, until you notice you have stopped visiting the landfill altogether.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Database Is Boring
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There is a particular kind of conversation that begins with "we should probably move the data to something more scalable". It is usually held by two people with three hundred rows between them. It ends, some weeks later, with a managed cluster, a monthly bill and a dashboard that nobody can log into because the person who set it up has gone on holiday.
ABC avoids that conversation by making the database the least exciting part of the stack. In Mat's back office the foundation is SQLite : a single file, on a disk, that has been quietly running a remarkable share of the world's software for over twenty years. It needs no server, no administrator and no meeting. A company of one, with tables measured in thousands of rows rather than billions, will not outgrow it this decade.
Three things sit on top of the plain database, and each one is as boring as the last. First, a schema written down : every table, every column and what each one means, in plain text in the repository, where a human or an agent can read it before touching anything. Second, an API: a small set of agreed doors through which the apps, the admin screens and the agents read and write, so nobody reaches into the tables directly. Third, admin screens: simple lists and forms for the occasions when a person needs to look or correct something by hand.
Then there is the backup, which is the only part allowed to be slightly obsessive. A copy every night, kept somewhere other than the machine it came from, with a handful of older copies kept beside it. And, crucially, a restore that has actually been tried. A backup you have never restored is a hope with a timestamp. Restoring one on a quiet weekday, deliberately, when nothing is wrong, is the cheapest insurance policy a small business will ever buy.
Why insist on boredom? Because every exciting choice in the foundation becomes a recurring cost in attention. A clever database needs clever people to look after it, and clever people are either expensive or busy, often both. A plain file with a written schema can be understood by any competent developer in an afternoon and by an agent in seconds. That is what makes the back office something you own rather than something you rent from your own complexity.
It also makes the machine honest. When the schema is written down, the agents can check their work against it. When the database is one file, the backup is one copy. When the API is the only door, every change leaves a trace. Boredom, it turns out, is just reliability that has stopped asking for attention, which is exactly what you want underneath everything else.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Money In, Money Out
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Money in a small business arrives in the most inconvenient formats it can find. A payment processor sends a CSV with its own idea of what a date looks like. An affiliate programme sends a statement in a currency you did not earn in, for a month that ended six weeks ago. An invoice you sent is paid in two halves, from an account whose name bears no resemblance to the customer's. Every one of them is technically correct.
The back office takes all of this and puts it in one place: the ledger . Every pound that comes in and every pound that goes out becomes a single line, with a date, an amount, a source and a link to the document that proves it. Invoices, payouts, refunds, subscriptions, affiliate statements and the occasional hosting bill all end up in the same table, in the same shape, however strangely they arrived.
The work in between is matching , and it is exactly the kind of work machines do better than tired people. A statement arrives. The agent reads it, works out what each line is, and looks for its partner: the invoice it settles, the app that earned the commission, the subscription that renewed. When the match is obvious, it is made and logged. When it is not, the line sits in a short queue with the agent's best suggestion beside it, waiting for a human to say yes or no. Nothing is guessed silently. Null beats invented here too.
Once the ledger exists, the reports are almost free. Revenue by app, by month, by source. Which affiliate programmes are paying and which are merely sending emails. Which costs have crept up since spring. These are not Sunday-evening spreadsheets, just views of a table that is already right. The scoreboard reads from the same place, so the number you see at 08:45 and the number in the ledger cannot drift apart.
The human part is small and deliberate. You do not type amounts in. You do not reconcile line by line. You approve the handful of matches the agent was unsure about, and you approve anything that sends money out. That is the line ABC draws everywhere: agents prepare, humans decide on payments.
For a fleet of small apps this matters more than it might seem. A hundred apps earning a pound a day produce a great many tiny payments from a great many sources. Matched by hand, they would cost more in attention than they earn. Matched by the machine, they become what they should be: a steady, boring column of evidence that the portfolio is paying its way. The next chapter follows one of those columns all the way to the scoreboard.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Revenue Feed
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
For a long time the most important number on Mat's scoreboard was an empty box. It sat there every morning at 08:45, honest and slightly reproachful, because the rule was null beats invented and nobody had yet connected the feed that would fill it. This chapter is about the morning that box got its first real number, filled without anybody typing anything.
A feed is simply a route by which a fact travels from where it happens to where you can see it. The revenue feed starts at the store, because that is where the sale actually occurs. Every store worth using can export its orders: date, amount, product, currency, refunds. That export is the raw truth, and the whole job of the feed is to carry it to the scoreboard without changing its meaning on the way.
The journey has five stops. The store holds the orders. The back office is told where to collect them. A nightly export runs while you sleep, fetching yesterday's orders and saving the file itself as evidence, untouched, with a date in its name. The orders are then matched: each one recorded in the ledger, tied to the app or product that earned it, with duplicates and refunds handled by rule rather than by mood. Finally the total is written to the scoreboard, where it appears in the next morning status as a figure with a link behind it.
That link is the point. Click the number and you reach the ledger lines. Click a line and you reach the order in the export. Click the export and you are looking at the file the store produced. Nobody has to trust the figure, because anyone can walk it back to its source in three clicks. Most people never bother. They simply stop worrying.
The first feed is always the hardest. Formats will surprise you. A currency will appear that you did not expect. One night the export will fail, and the right behaviour is for the box to show that it failed, not yesterday's figure pretending to be today's. Each of these is a small fix, made once and written into the feed's procedure, so it never needs fixing again.
After the first, the others come more easily. Affiliate statements, ad revenue and email subscribers each follow the same five stops, with a different source at the start and the same scoreboard at the end. The pattern is the asset. Build it once with care and the rest is mostly copying.
And then one weekday morning the status arrives and the box is full. The number is small. It is also true, which is considerably rarer, and you will find it far more motivating than any figure you ever estimated.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Media and Pipelines
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every business with a camera eventually acquires a downloads folder that frightens it. Images generated at midnight, screen recordings of a demo that went rather well, a voice memo containing the best idea of the quarter, all filed under names like final_final_2.mp4. The media exists. It is even good. It is just not findable, which for practical purposes means it does not exist at all.
The back office treats images, video and audio as assets like any other, with one difference: they arrive constantly and in bulk. A fleet app needs a hero image, a set of cards, an icon and fourteen clips. Multiply that by the fleet and you have thousands of pieces of media a year, most made by machines in seconds. Without a system, that volume becomes weather.
So media travels through a pipeline , and the pipeline has four stops. Capture is wherever the thing is made: a model generating an image, a screen recorder, a phone. Tag is the moment it gets its six fields, what it is, owner, app, status, use and location, ideally added by an agent at the point of creation rather than by a person months later. Store puts the original somewhere permanent, once, with the record in the back office pointing at it. Reuse is the payoff: the same image appearing on the site, in a clip and in a book, found by search rather than by memory.
The vault mirrors this structure for the human side. Under Media there is a place for assets, split into images, videos and audio, and a place for pipelines, where the planning documents live: what a clip batch needs, which prompts produced which style, what was approved and what was quietly retired. Plan and output sit side by side, so next month's batch starts from last month's lessons.
The real gain is in reuse, and it compounds. A clip made for one app can be cut differently for another. An approved hero image becomes the starting point for the next variation. A voice recording can be transcribed into context and quoted in a chapter. Media that sits in a downloads folder is spent the moment it is used. Media that sits in a pipeline is closer to capital: it keeps being drawn on.
There is a modest cost discipline here too. Generating media is cheap per item and expensive in aggregate, and the easiest saving is not making the same thing twice. When an agent can search what already exists before it asks a model for something new, the bill shrinks without anybody noticing.
None of this needs a studio. It needs four stops, six fields and the patience to tag things as they arrive. Do that, and the frightening folder gradually empties into a library you can actually use.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Permissions and Privacy
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every small business has a shared folder that began life with good intentions. It was called "Team" or "Admin", and within a year it contained the contracts, the bank statements, a passport scan, three versions of the logo and a spreadsheet of client phone numbers that a freelancer was given access to in 2023 and never lost. Nobody decided who could see what. It simply accumulated, like dust or opinions.
ABC treats access as a decision rather than an accident. The rule is short enough to fit on a sticky note: private by default, shared on purpose . Nothing in the back office is visible to anyone, human or agent, until there is a reason for it to be. When there is a reason, access is granted for that reason, written down, and taken away again when the reason ends.
The test is one question, asked every time: does this person or this agent need it to act? Not "might it be useful", but need it, for a job they have actually been given. The bookkeeper needs the invoices and the statements. She does not need the passport scan. The agent that drafts product listings needs the product images and the affiliate tags. It does not need the bank feed. Each grant is narrow, because narrow grants are easy to reason about and wide ones are not.
All of this lives in one plain file, a permissions register , kept alongside the rest of the back office. It lists who and what can reach each area, why, and since when. In Mat's own machine there is a page in the vault that does exactly this, and the projects themselves carry a line that says, without ceremony, that this work must remain private and must not be pushed anywhere public. It reads like a fussy note on a fridge.
Agents make this more important, not less. A person with too much access usually lacks the time to misuse it. An agent has nothing but time. It will read everything it can reach, very quickly, and occasionally quote it back in an unexpected place. So the machine is built on the assumption that whatever an agent can see may one day end up in a draft, and the access list is drawn accordingly.
Privacy, in the end, is mostly housekeeping. Once a quarter the register is read through, line by line, and anything without a current reason is removed. It takes ten minutes and feels faintly like tidying a sock drawer.
The result is a back office where you can answer the question nobody asks until something goes wrong: who could have seen this? The honest answer should be short, written down, and slightly boring.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Back Office Agent
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every business has someone who tidies up. In a large company it is a department with a manager and a quarterly review. In a company of one it is you, at ten past eleven on a Sunday night, renaming files called "final_FINAL_v3" and wondering where the receipt for the microphone went. ABC gives that job to an agent, and the agent does not mind Sundays.
In Mat's machine the back office agent is called boai , which is bo28 with an AI attached, pronounced however you like. It has a deliberately small job description, and it runs the same loop every day: tidy, match, report. It does not have ambitions. It does not suggest a rebrand. It does the three things, in order, and then it stops.
Tidying comes first. New files arrive in the back office from everywhere: exports, statements, images, invoices dropped in by hand. boai gives each one a name that follows the house rules, files it in the right place, and attaches the basics that turn a file into an asset: what it is, which app or client it belongs to, and its status. Anything it cannot place goes into a short pile marked "unsure" rather than a confident guess in the wrong drawer.
Matching is the second step, and the one that saves the most time. A payout arrives; boai looks for the invoice it settles. An affiliate statement lands; boai lines its figures up against the app that earned them. A product image appears; boai links it to the listing that uses it. Most of these matches are obvious, and the obvious ones are made without fuss. The doubtful ones are not made at all. They are flagged, because a wrong match is worse than an open one.
The loop ends in a report , and the report is the point. A short note each morning, feeding into the 08:45 status: what was tidied, what was matched, what is still unmatched and why, and which feeds on the scoreboard changed. Nothing in it is invented. If the store export did not arrive, the report says so, plainly, and the scoreboard keeps its honest empty box.
There are things boai does not do, and they are written down. It does not pay anyone. It does not delete anything, only archives. It does not change a number on the scoreboard by hand. Every action that moves money or cannot be undone waits for a human to approve it, which is exactly where a human is most useful.
What you get, in the end, is not a clever assistant. It is a reliable clerk who never tires of the dull part. That turns out to be rarer, and considerably more valuable.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Boring Is a Feature
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Nobody has ever thrown a launch party for their filing cabinet. Nobody posts on a Friday afternoon about how smoothly the invoices matched this week. This is the clearest sign that a back office is working: it has stopped being interesting. The day you forget it exists is the day it has finally become good.
Part IV has been, by design, the least thrilling stretch of this book. One source of truth instead of twelve spreadsheets. Assets rather than files. A database chosen for its lack of surprises. Money flowing into one ledger. A revenue feed that fills the scoreboard without anyone typing. Media that can be found again. Permissions written down. An agent that tidies, matches and reports. Not one of these would survive a pitch meeting on charisma alone. Every one of them quietly removes a reason to worry.
It helps to picture two dials. One measures drama : how often something surprises you, breaks, goes missing or needs a late evening to fix. The other measures reliability : how often things are where they should be, when they should be, without being chased. Most small businesses live with high drama and middling reliability, and mistake the drama for momentum. The back office belongs firmly in the opposite corner. Low drama, high reliability, and a mild sense of anticlimax.
This is not an argument for doing nothing. Boring takes work. It takes a schema written down, a backup that is actually tested, a rule about which copy wins, and the patience to fix the same small thing properly once instead of improvising around it forever. In Mat's machine, the same instinct shows up everywhere: main CI goes green every weekday, and no red run is allowed to survive a working day. Nobody celebrates a green tick. That is precisely why it stays green.
There is a Stoic flavour to all this. Seneca suggested that a well-ordered mind is shown by its ability to stay in one place and linger with itself. A well-ordered back office is much the same. It stays put, it knows where things are, and it does not need to be reassured.
The reward for all this tedium is attention. Every hour the back office does not demand is an hour returned to the one goal, which is the only place an hour was ever meant to go. Boring, it turns out, is a very generous feature.
Which brings us to the third pillar. Apps earn the money, the back office keeps the records, but neither knows what the business is for, who it serves or what was decided last March. That knowledge lives somewhere else, in writing, and the whole machine reads it before it acts. Part V is about context.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
C Is for Context
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Hire a brilliant new assistant, sit them at a desk, and give them nothing to read. By lunchtime they will have reorganised something that did not need reorganising, emailed a client you stopped working with in 2024, and politely asked what the business actually does. None of this is their fault. Brilliance without context is just confidence with nowhere to go.
Agents are exactly the same, only faster. They arrive knowing a great deal about the world and nothing whatsoever about you. They do not know your goal, your prices, your tone, your clients or the fact that you decided, after one bruising quarter, never to offer discounts again. ABC's third pillar exists to fix that. C is for Context : the private, written memory of the business, kept in one repository and read by every agent before it does anything.
The context holds four kinds of thing. First, the goal : the one line with a number and a date that sits at the very top, so every agent knows what it is working towards. Second, the people: who you are, who you work with, who your customers are and how they like to be spoken to. Third, the decisions: what was chosen, when and why, so nobody, human or machine, has to argue it out again. Fourth, the rules: the things that are always true here, from house style to privacy, from which tools to use to which repositories must never be pushed anywhere public.
In Mat's machine this is a single repository called context-2026. It is mostly plain text, unglamorous and private. It contains no clever code. It contains what the business knows, written down in sentences a stranger could follow, organised so that an agent can find the relevant part quickly and ignore the rest.
The arrangement changes how agents behave. An agent that has read the context drafts in the right voice, quotes the right price and declines, of its own accord, to build the feature that does not move the goal. It feels less like instructing a tool and more like briefing a colleague who has been paying attention. The difference is not the model. It is the reading.
Context also protects you from yourself. Tired founders reopen settled questions. A written decision, sitting where it can be found, closes them again with very little fuss. The context remembers what you agreed with yourself on a clearer day.
The next nine chapters cover how this pillar is built: how to write things down once, how to arrange them in rings, where the hand-written notes live, and why the memory must never be allowed to edit itself. All of it rests on one modest idea. Before the machine acts, it reads.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Write It Down Once
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most businesses make the same decision several times a year without noticing. What do we charge for a rush job? Do we work with agencies? Which font goes on invoices? Each time the question arrives, it is answered from scratch, slightly differently, by whoever happens to be tired at the time. The business does not lack decisions. It lacks a memory of having made them.
ABC's rule is simple to the point of embarrassment: write it down once . When a decision is made, it goes into the context, in one place, in plain text, with the date and a sentence on why. Not in an email thread. Not in a chat message that will scroll out of sight by Thursday. Not in your head, which is excellent at many things and terrible at retrieval under pressure. One place, one paragraph, done.
Plain text matters more than it sounds. A decision written in a proprietary tool is a decision that lives at the mercy of a subscription. A decision written as a short markdown file can be read by you, by any agent, by any editor, today and in ten years. It can be searched, linked, versioned and diffed. It does not need an app to open it. It is the most durable format the computer industry has ever produced, mostly because nobody has managed to improve it.
The leverage comes from reuse. A pricing decision written once is read by the agent that drafts proposals, the agent that writes product pages and the agent that answers enquiries. A tone-of-voice note written once shapes every email, every listing and every clip caption. You did the thinking on one afternoon. The machine reuses it a thousand times without asking you again. There are very few places in a business where an hour of work keeps paying out indefinitely. This is one of them.
The habit has a small cost. After each decision, you spend two minutes writing it down. That is the whole bill. Compared with the hours spent rediscovering, re-arguing and quietly contradicting yourself, it is the cheapest leverage you will ever buy.
There is a gentle discipline in this. Writing a decision down forces you to state it clearly, and a surprising number of decisions dissolve at that stage, revealed as half-formed preferences rather than choices. That is useful too. Better to find out at the desk than in front of a client.
Over a year, the written decisions add up to something more valuable than any single one of them: a business that knows its own mind. New agents read it and behave. Future you reads it and stops arguing. And the question about rush-job pricing, when it arrives again in November, takes about four seconds.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Five Rings of Context
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask a founder to write down everything an assistant should know about the business and you will receive either one sentence or a forty-page document with a table of contents and no ending. Neither helps. The first leaves the agent guessing. The second buries the useful part somewhere around page twenty-three, between a mission statement and a photograph of the office plant.
ABC organises context in five rings , starting at the centre and working outwards: me, my business, my clients, my market and the world. Each ring answers a different question, and each is written separately, so an agent can read the rings it needs and skip the rest. Drafting an invoice needs the inner rings. Writing a market note needs the outer ones. Nobody needs all five for everything.
The first ring is the person. Who you are, how you work, what you are good at, what you refuse to do, the hours you keep and the tone you write in. It is short, personal and the most private ring of all. It is also the one that most changes the output, because an agent that knows you prefer two lines to two paragraphs will stop sending you two paragraphs.
The second ring is the business, and it is the working centre . The one goal sits here, with its number and its date. So do the offer, the prices, the decisions written down once, the apps in the fleet and the rules of the house. Most agents spend most of their time reading this ring, which is why it gets the most care.
The third ring is the clients: who they are, what they bought, what they asked for, what annoyed them and what they were promised. The fourth is the market: competitors, channels, prices elsewhere and what is changing. The fifth is the world, kept deliberately thin: the handful of outside facts, such as tax dates, platform rules and seasonal patterns, that actually affect your work. The world is large. Your ring of it should not be.
The rings also tell you how often to look at things. The person changes slowly. The business changes monthly. Clients change weekly. The market changes when it feels like it. The world changes constantly, which is exactly why you only keep the parts of it that touch you.
There is a quiet benefit to writing in rings. It shows you, very plainly, where your attention has been. Most founders find a thick business ring, a decent client ring and an outer ring with nothing in it but a vague unease. That is not a failure. It is a map, and you cannot improve a map you have never drawn.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Vault
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There is a particular kind of horror reserved for the moment you discover that an automated tool has helpfully "cleaned up" a folder. Two years of notes, gone, replaced by a neat set of freshly generated files that are tidy, accurate and completely missing the bit you wrote at two in the morning that turned out to be the whole idea. Machines are very good at tidying. They are less good at knowing what matters.
ABC keeps its working notes in the vault : a standalone Obsidian vault, separate from the context repository, where the machine and the owner both write. That sharing is the useful part and the dangerous part, so the vault is divided into clear zones, and the zones are not a matter of etiquette. They are enforced.
The generated zones belong to the machine. In Mat's setup they follow a plain structure: Apps for notes on each app in the fleet, Bo for the back office, Context for the rings and the business, Media for assets and pipeline plans, and an infrastructure folder holding the vault standard, the permissions page and the dashboards. A single sync command rebuilds all of these from the real sources, so they are always current and never lovingly hand-edited into fiction.
The written zones belong to you. Daily, where the day's notes live, and Legacy, where older material rests in peace. The sync script never touches them. Not to tidy, not to "improve the formatting". They are yours in the way a diary is yours, and the machine treats them with the same wary respect.
The thing that makes this safe is a small, dull file called the manifest . Every time the machine generates a note, it records the path in the manifest. When the sync runs again, it deletes only the files listed there before rebuilding them, and nothing else. It never wipes the vault root. It never assumes a file is its own because it looks like the kind of thing it might have written. If a file is not on the list, the machine leaves it alone, whatever it is called.
This sounds like a lot of ceremony for some markdown files. It is the right amount. The vault becomes a place where you can open any note and know, from where it lives, who wrote it and whether it will still be there tomorrow. Generated notes are fresh and disposable. Written notes are slow and permanent. Both are useful, and neither gets in the other's way.
Obsidian was chosen for familiar reasons: plain files on your own disk, links between notes, and no hostage-taking if you change your mind. The vault is where the machine's knowledge and yours finally sit in the same room, on good terms, with a very clear seating plan.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Read-Only by Default
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Imagine a diary that rewrote itself overnight. Nothing dramatic: a word softened here, a date nudged there, an awkward decision quietly rephrased into a sensible one. Within a year you would have a perfectly readable account of a life you never lived. You would also have no way of knowing which parts were yours.
That is the risk ABC is designed to avoid, and the rule is blunt. The context repository is read-only to agents . They may open it, search it, quote it and act on it. They may never create, edit, move, delete, stage or commit a single file inside it. In Mat's machine this is written into the instructions every agent receives: context-2026 is to be treated as read-only, and no agent is to recreate folders inside it or slip in compatibility files and shortcuts. There is no exception for helpfulness.
The reason is trust. The context is the thing every agent reads before it acts, so it has to be the thing nobody can doubt. If agents could write to it, every line would carry an invisible question: did I decide this, or did a model infer it at four in the afternoon and save it with great confidence? Once that question exists, the context stops being a memory and becomes a rumour.
There is also a quieter problem, which might be called drift. A model that summarises a decision will summarise it slightly differently from the original. Another model reads the summary and summarises it again. After a few rounds, the business has a set of beliefs that no human ever held, each one a careful paraphrase of the last. Read-only context stops that loop before it starts, because the source never changes unless a person changes it.
So where do agents write? They write to the vault, in its generated zones, tracked by the manifest. They write to the back office, to apps and to reports. They write drafts and proposals. If an agent thinks something belongs in the context, it says so, in a note or a pull request for a human to read, and the human decides. The human gate sits exactly where the memory is kept.
This makes the context a little slower to change, and that is a virtue. The things written there should be things you meant: the goal, the rules, the decisions you would defend on a bad day. Slow to change is another way of saying hard to corrupt.
It is also, oddly, a relief. You never have to wonder what the machine did to your memory while you slept. It read it, carefully, and left it exactly where you put it.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Memory That Travels
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every agent session begins the same way: bright, willing and completely unaware of anything that happened before lunch. It is like hiring a brilliant temp every morning who has never met you, never seen the building and cannot find the kettle. They will work very hard, and they will politely ask the question you answered yesterday.
The fix is not a cleverer temp. It is a note on the desk. ABC calls this project memory : a small set of plain files, kept with the work, that tell any agent arriving fresh what it needs to know before it touches anything. Which port the app runs on, which folder is read-only, what was tried last week and why it failed. None of it is profound. All of it is the difference between an agent that starts informed and one that starts guessing.
In Mat's machine, every repository in the fleet carries one such file at its root, and every agent reads it before it does anything else. It says what the app is for, how it is built, where the boundaries are and what must never happen. When an agent learns something worth keeping, a build quirk or a naming decision, the instruction is simple: save the fact where the next session will find it. Not in the chat, which evaporates. In the file, which stays.
The discipline is in what gets saved. Memory is not a diary. It does not record that the tests were lovely. It records facts that change what the next agent should do: a rule, a path, a decision, a known trap. If a line would not alter tomorrow's behaviour, it does not belong there. This keeps the files short enough to be read in full, which is the only way they get read.
There is a pleasing side effect. Because the memory travels with the project rather than with any particular agent, the agents become interchangeable. A session that runs out of time, an upgraded model, a second agent brought in to help: each opens the same file and picks up the same thread. Nobody has to be briefed. Nobody has to remember. The work simply continues, which is what work is supposed to do.
It also turns out to be useful for humans. After the agents, the most frequent reader of project memory is Mat himself, returning to an app he has not opened in a month and discovering that his past self, or rather his past agents, left him a perfectly sensible note. It is the most reliable colleague in the building, mostly because it never goes home.
A fact saved today is a question nobody has to ask tomorrow. Over a year, that is a great many questions, and a great deal of very calm work.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Daily Note
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most people keep a daily log at some point in their lives. It usually lasts eleven days, peaks with a long entry about a sandwich, and is discovered years later in a drawer, where it explains nothing. The problem was never the habit. The problem was that nobody read it, so there was no reason to write it.
ABC gives the daily note a reader. In fact it gives it several, all of them tireless. The daily note is a single page in the vault, one per day, in the hand-written Daily/ folder that the machine reads but never edits. It answers three questions in plain sentences: what happened, what was decided, and what is next. It takes five minutes at the end of the day, or three if the day was dull, which is a perfectly respectable kind of day.
The first question is the record. What shipped, what broke, who called, what the scoreboard said. It is not a mood diary, although a line about mood is allowed if it explains a decision. The point is that tomorrow's reader, human or agent, can see the shape of the day without having been there.
The second question is the one that earns its keep. Decisions made in conversation have a half-life of about a week. Written in the daily note, with a sentence on why, they become part of the context, and the next time the same question surfaces, the answer is already waiting. A decision that was written down is a decision you only have to make once.
The third question closes the loop. "What is next" is a short list, usually no more than three items, and it is the first thing the agents read in the morning. Before the 08:45 status goes out, the machine has already looked at yesterday's note, compared it with the backlog and the scoreboard, and worked out whether today's plan still makes sense. If you wrote that the affiliate feed needs connecting, the status will say whether anybody did.
There is a quiet discipline in this. Knowing the note will be read makes it honest, and knowing it must fit in three short answers makes it brief. You will not write about the sandwich. You will write that the build is green, that you dropped the podcast idea again, and that tomorrow is for the clip calendar. Then you will close the laptop, which is the other half of the habit.
Over a month, the notes become a running memory of the business in its own words. Over a year, they are the best account anybody will ever have of how the goal was actually moved. One page a day, read at dawn, acted on by lunch. It is a very small machine for keeping a very large one honest.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Context Beats Cleverness
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every few months a new model arrives, and the internet greets it like a second coming. It is faster, it is smarter, it has passed an exam nobody you know has ever sat. Then you ask it to write a renewal email to your best customer, and it produces a lovely letter to somebody else entirely, in a tone you would never use, about a product you do not sell.
This is not the model's fault. It was asked a question with no context, and it did what any clever stranger would do: it guessed, fluently. ABC is built on a rule that sounds almost too plain to write down. Context beats cleverness. A modest model that knows your goal, your customers, your prices and your house style will beat a brilliant model that knows none of them, nearly every time, on nearly every job a small business actually needs done.
Picture a square: model power along one side, context along the other. Top right, both high, is lovely and expensive. Bottom left, both low, is where most business chatbots live. The interesting corner is high context with an ordinary model, because that is where good output becomes cheap and repeatable. The corner to avoid is a powerful model with no context, which produces confident nonsense at speed. It is the costliest way yet devised to be wrong.
So every ABC prompt is built context first. Before an agent writes a word, it is handed the goal line, the relevant notes from the vault, the project memory, yesterday's daily note and the evidence it will be judged against. The instruction itself is often short. Draft the reply. Match the house style. Cite the order number. The short instruction works because the long context has already done the thinking about who, what and why.
This changes where you spend your effort. Most people, faced with a disappointing answer, try to write a cleverer prompt, adding capital letters and the word "please" in increasingly desperate tones. In ABC, a disappointing answer is treated as a missing fact. What did the agent not know? Write it down once, in the right place, and every future answer improves, not just this one.
It also makes the machine resilient to the model market. Models will keep changing. Some will get cheaper, some retired, and one day you will switch providers because the price suits you. If your quality lived in the model, every switch would be a gamble. Because it lives in the context, the switch is mostly a shrug. The new model reads the same files and does the same work, usually a little better.
Cleverness is rented by the month. Context is written by you, kept by you and read by whichever model turns up. Bet on the part you own.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Sweeping the Vault
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every notes system starts as a library and ends as a loft. For the first month everything has a shelf. By the sixth, there are three notes called "Pricing", two of them contradicting the third, a dozen links pointing at pages that were renamed in spring, and a folder labelled "misc" that nobody has opened since. Nothing in the loft is wrong, exactly. It is just no longer true.
Context that is no longer true is worse than no context at all, because the agents believe it. An agent that reads last year's prices will quote last year's prices, with a citation. So ABC does what any sensible household does on a Sunday: it tidies up. The sweep is a scheduled job, run weekly, that walks through everything touched in the vault that week and asks one question of each note: is this still true?
The answer sorts the notes into three piles. Notes that are still relevant are kept, linked to their neighbours and filed in the right zone. Notes that were true once but are not any more are archived, moved out of the way with a line saying when and why, so the history survives without misleading anyone. Duplicates are merged, and the survivor keeps the best sentences from both. Broken links are repaired in passing, like picking up the post from the doormat.
In Mat's machine this is a skill called abc-vault-sweep. Raw material lands in an input folder during the week: chat exports, session logs, voice summaries. The sweep distils that pile into current state and into a small wiki of things worth knowing for good. It is a funnel: a great deal goes in at the top, rather less comes out at the bottom, and what comes out is the bit you would actually want an agent to read.
The sweep has firm manners. It works only in the zones the machine generated, the ones listed in the manifest. It never touches the hand-written Daily/ and Legacy/ folders, however untidy they become, because those are yours. And it never deletes anything it cannot account for. If the sweep is unsure whether a note is stale, it leaves the note alone and mentions it in the morning status, so a human can decide over coffee.
Weekly is the right rhythm. Daily is twitchy: notes need a few days to show whether they still matter. Monthly is too late: by then the loft has a loft. Once a week, the vault is brought back to a state where every note can be trusted, which means every agent reading it can be trusted a little more as well.
Tidying is not glamorous. Neither is a library where every book is on the right shelf, until you need one.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Context Is the Moat
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There is a particular kind of meeting, held in large companies, where someone asks what the moat is. It usually ends with a slide about proprietary technology and a long silence about the fact that a competitor with a credit card could rent the same model by teatime. For a company of one, the question is simpler and the answer is more comforting.
Anyone can rent the model. That is the point of renting. The most capable AI in the world is available to you, your competitor and a teenager in a bedroom, at the same price. If your advantage lived in the model, you would have no advantage at all. You would have a subscription.
What nobody can rent is your context. Stack the machine up and it is plain. At the bottom sits the rented model , the same for everybody. Above it sit shared tools: the editors, the hosting, the databases, available to anyone who looks. Above those sit your skills, the written procedures that say how your business does things, and right at the top sits your context : the goal, the decisions, the customers, the house style, the daily notes, the parking file. The lower layers are commodities. The top one took a year of ordinary Tuesdays to write, and it cannot be copied because it was never published.
This is what the whole of Part V has been building towards. Write it down once. Keep it in rings, from you outwards. Store it in a vault with clear zones and a manifest. Make the memory read-only to the machine. Let project memory travel between sessions and let the daily note feed the morning. Sweep it weekly so it stays true. None of these habits is impressive alone. Together they produce an asset that gets more valuable every week and that no competitor can buy.
It compounds in a very practical way. Every fact written down improves every future answer. Every decision recorded is a decision never reopened. Every sweep makes the agents slightly more trustworthy. A year in, the same rented model that gives a stranger generic advice gives you something that sounds like your business on a good day, because it has read your business on every day.
There is also a quiet freedom in it. Because the value lives in your context, you are never trapped by a vendor. Change the model or the tools: the context comes with you, in plain text, in a folder you own. That is the only kind of moat worth digging when you work alone.
With the memory in place, the machine has something worth reading. The next question is who does the reading. Part VI introduces the agents: one master, five pillars and a roster of skills that turn all this context into work.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
One Master Agent
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The first thing people do when they discover agents is make lots of them. One for email, one for invoices, one for social posts, one for research, and one, inevitably, called "Strategy", which nobody quite knows what to ask. Within a fortnight they have a small office full of eager assistants, none of whom talk to each other, all of whom want a word.
ABC starts from the other end. There is exactly one master agent , and in Mat's machine it is called ABC Master, invoked with a single command: /abc . It sees every pillar at once: the apps fleet, the back office, the context and the Prodgame scoreboard. Whether you want something done or want to know how things stand, you ask the master. You do not need to know which specialist exists or what it is called. That is the master's problem, and it enjoys it.
The master's real job is routing . A request comes in, say "why did the travel app earn nothing yesterday?", and the master works out who should answer it. It asks the back office for the affiliate statement, the apps pillar for the deploy log and the context for any note about a change. It gathers the answers, checks each one against its evidence, and replies in one voice, with links. You asked one question and got one answer. Four agents were busy behind the scenes, and you met none of them.
Why insist on one? For the same reason a business has one owner. When something goes wrong, there must be one place to ask and one place that answers. A committee of bots produces what committees always produce: overlapping work, conflicting advice and a summary that begins "on the one hand". One master means one account of the truth, one queue of work and one morning status at 08:45 that speaks for the whole machine.
The master is also where the rules live. It knows the goal, because it reads the goal line first. It knows what it may not do: it drafts payments but does not send them, opens pull requests but does not merge its own. Its instructions are written down as a skill in the repository, so you can read exactly how your chief of staff has been told to behave and change it with a pen.
For clients the arrangement is identical, only renamed. Each business gets its own master, named after the business and invoked like a command, sitting above its own pillars. The shape stays the same so that the habits transfer: ask one voice, receive one answer, check the evidence underneath.
A good orchestra has many players and one conductor. Nobody has ever complained that the conductor was a bottleneck. They complain when there are three.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Five Pillars of Agents
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Org charts have a poor reputation, mostly earned. They tend to be drawn after the fact, to explain a mess, and they grow boxes the way old houses grow extensions. But a small org chart drawn before the mess is a different thing. It is a promise about who does what, made while there is still time to keep it.
Below the master, every ABC system has the same five pillars of agents . Each pillar has a lead agent with a short job description, and beneath each lead sit the specialists that do the actual work. The pillars are the same for every client, which means the chart is familiar on the first day and still makes sense on the three hundredth.
Acquire handles the front door: the landing page, search, content and lead qualification. In the fleet, this is where the fourteen clips per app and the pages that bring strangers in belong. Deliver runs the core workflow, one specialist for each manual step that onboarding uncovered, so the work that used to eat your Tuesdays now happens in the background and arrives as a draft.
Operate keeps the back office tidy. It reconciles the numbers, cleans the data, feeds the scoreboard and drafts support replies for a human to approve. Build looks after the code: it opens pull requests, runs the checks and keeps main green, and it never merges its own work, which is the single most important sentence in its job description.
Govern watches everything else, and sits closest to the master for good reason. It keeps an eye on the monthly AI budget, with a sensible starting figure of around fifty pounds and an alert at eighty per cent. It keeps the approvals queue honest, keeps the audit log and checks that every claim in the morning status has evidence attached. An agent system without a governor is a cost line with ambitions.
Why five? Because five is enough to cover a small business and few enough to remember without a diagram. Fewer, and pillars start doing two jobs badly. More, and you are back to an office full of eager assistants. If a new specialist cannot be placed under one of the five, it probably should not exist. The leads help the master too: instead of guessing among thirty specialists, it routes to one of five leads, and each lead knows its own team. Two short hops instead of one long guess.
Five boxes, one master, and a rule that every box earns its place. It is not much of an org chart. That is precisely why it works.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What a Skill Is
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every business has procedures. Most of them live in one person's head, where they are stored in a format best described as "you sort of just know". The rest live in a document called something like "How we do invoices (FINAL v3)", last opened by someone who has since left. Neither is much use to a machine, or to you.
ABC writes procedures down in a form an agent can pick up and use. That form is called a skill . A skill is a small folder with a plain-language set of instructions at its heart, saying what the job is, how to do it and what finished looks like. Alongside the instructions sit whatever the job needs: a script for the fiddly part, a template, a list of the tools the agent may use. An agent with no skills is a clever generalist. An agent that has loaded a skill is a specialist for exactly as long as the job takes.
Draw it as two overlapping circles. One circle is the procedure: the knowledge of how your business does a thing. The other is the tools: the scripts, files and connections that let it actually be done. Where they overlap is the skill. A procedure without tools is a memo. Tools without a procedure are a shed. Together they are a job that gets done the same way every time.
The most important line in any skill is the definition of done . Not "make a diagram" but "one self-contained file, in the house blue, legible in light and dark, saved where the book can find it". Not "check usage" but "a report per project for the current week, with each figure marked measured or inferred". The definition of done is what lets an agent, or you over coffee, say with confidence that the job is finished. Without it, a job is finished whenever someone gets bored.
Mat's machine runs on dozens of these. abc-diagram draws the editorial diagram that sits beside every chapter of this book. abc-usage reports where the week's tokens went. abc-vault-sweep tidies the vault. The App Factory is really a chain of skills, from /abc-seed to /abc-clip, each handing a finished output to the next. None of them is magic. Each is a page of careful instructions and a few scripts, improved every time something goes slightly wrong.
That last part matters. When a skill produces a poor result, the fix goes into the skill, not into a one-off apology. Every later run benefits. Over a year, the skills become the most accurate description of how your business actually works that has ever existed.
A procedure in your head is a habit. A procedure in a skill is an asset, and it does not forget the steps on a Friday.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Skill Canon
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A skill written once is useful. A skill written four times is a problem. This is not a theory; it is what happens the moment you use more than one AI surface. You write a lovely procedure for drafting the weekly report in claude.ai. A month later you want the same thing in Claude Code, so you copy it. Then you tweak the copy. Then you need it in the browser, and the browser version is a copy of the old one, before the tweak. By spring you have four procedures with the same name and four slightly different opinions about what a report is.
ABC solves this with something deliberately unexciting: the skill canon . Every skill lives in one repository, called skills-2026, and that repository is the only place a skill is ever edited. From there it is synced outwards to every surface the business uses: claude.ai, Claude Code, Claude in Chrome and Codex. The surfaces receive skills. They do not get to keep their own.
The rule sounds bureaucratic until you try living without it. When a skill is wrong, you want to fix it in one place and know the fix has reached everywhere. When an agent behaves oddly on one surface and normally on another, you want to be able to rule out the obvious suspect, which is that it is reading a different set of instructions. With a canon, that suspect has an alibi.
Because the canon is a repository, it inherits everything good about code. Every change is a commit with a date and a reason. Every new skill arrives as a pull request that someone can read before it lands. Old versions are never lost, so if this week's edit to the invoicing skill turns out to have been a mistake, last week's version is one command away. Skills get the same care as software because, in a system run by agents, they are the software.
There is also a quieter benefit. A canon makes the skills legible to the owner. Open the repository and you can see, in plain text, every procedure the machine knows. Nothing is hidden in a settings panel on a website you visited once. If you want to know how the back office agent reconciles a statement, you read the file. It is usually shorter than you expected.
The sync itself is a skill, naturally, which has a pleasing circularity to it. It checks what each surface holds against the canon, reports anything out of date and puts it right. The report is boring on a good day. Boring is the goal.
The Stoics liked a man who was the same in the marketplace as at home. The canon asks the same of a skill: one character, whatever surface it turns up on. It is a modest virtue, consistency. It is also, in a business run by agents, most of the battle.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Anatomy of an Agent
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Strip the mystery off an agent and you find five parts, stacked like a sandwich that has been designed by an engineer. There is nothing magic in it, which is good news, because things with no magic in them can be understood, repaired and, when necessary, replaced.
At the top sits the persona . This is the agent's name, its job and its manners: who it works for, what it is responsible for and what it must never do. The back office agent, boai, has a persona that says it tidies, matches and reports, and that it does not send anything to a customer. The persona is short. Long personas tend to become autobiographies, and nobody wants an agent with an autobiography.
Beneath the persona sit the skills , the written procedures from the last two chapters, loaded from the canon. This is where most of the agent's usefulness lives. The persona says what the agent is for; the skills say how it does each part of the job, step by step, with a definition of done at the end of each one.
Below the skills come the tools and the memory. Tools are the hands: the ability to read a repository, query the back office, open a pull request, fetch a statement. Memory is what lets an agent pick up where the last session stopped, instead of introducing itself afresh every morning like a goldfish with a lanyard. Memory reads from the context, which agents may consult but never rewrite.
At the bottom sits the model. It is the engine, and like most engines it is the part people talk about most and need to think about least. Models improve every few months. When a better one arrives, ABC swaps it in underneath, and the persona, skills, tools and memory above it carry on unchanged. The agent gets cleverer without needing to be rebuilt.
That is why the diagram puts the skills in the middle and gives them the emphasis. They sit between who the agent is and what it runs on, translating the one into the other. Swap the persona and you have a different agent with the same abilities. Swap the model and you have the same agent with a faster engine. Swap the skills and you have changed what the business actually does. This layering also makes agents cheap to create. A new agent is mostly a new persona plus a selection of skills that already exist. The vault agent and the app agent share a surprising number of procedures, and neither had to learn them from scratch.
So when someone asks what an ABC agent really is, the honest answer is undramatic. It is a job description, a stack of procedures, a set of hands, a notebook and an engine. Rather like a good employee, in fact, except that this one reads the procedures.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Agents Hand Off
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most work in a business crosses a boundary. A sale becomes a job, a job becomes an invoice, an invoice becomes a payment someone has to match. In a human team, each crossing is where things fall down the back of the sofa. Somebody assumed somebody else had it. In an agent team, the crossings are the same, and so is the sofa, unless the handoff is designed.
Take a real one. The app agent finishes building a new app for the fleet and puts it live. Good. But an app that earns money needs its revenue counted, and counting revenue is the back office's job, not the app agent's. So the app agent hands off to boai with a request: add a revenue feed for the new app. boai builds the feed, connects it to the scoreboard and confirms. Five steps, two agents, one result: a new app whose earnings appear in the morning status from day one.
The interesting part is what travels with the request. A bad handoff says "please add a revenue feed" and leaves the receiving agent to work out which app, which account, which store and what counts as a sale. A clean handoff carries the context with the work: the app's name and live URL, where its payments land, the merged pull request that put it live and what done looks like for the feed. The receiving agent starts informed instead of starting with questions.
In ABC every handoff is written down, in the thread where the work lives, not whispered between agents in some private channel nobody can read. That matters for two reasons. The first is that the owner can see it, and can step in if the request is wrong. The second is that when something goes astray three weeks later, there is a trail. You can find the exact moment one agent passed the parcel and see what was in it.
There is a case for saying that the handoff is the product. Individual agents are getting better at individual tasks every month, and that is largely the model makers' doing. What they are not getting better at, on their own, is working together. The value of a system like ABC lies less in any single agent than in the joins between them: the fact that a new app reliably becomes a counted app, without the owner remembering to ask.
The handoff also closes properly. The receiving agent does not simply go quiet when it is finished; it confirms, with evidence, in the same thread. "Feed confirmed" means the feed exists, has run and has produced a number someone can check. Only then does the first agent mark its part done.
Relay teams win and lose races at the changeover. Agents are no different. Anyone can run fast with the baton. The skill is in letting go of it at exactly the moment the other hand has closed.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Humans at the Gates
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Agents in ABC are capable, quick and tireless. They are also, by design, not trusted with everything. This is not a criticism of agents. It is the same principle that stops a new employee, however brilliant, from wiring money on their first morning. Capability and permission are different things, and the gap between them is where trust gets built.
The line is drawn with one question: is it reversible? If an action can be undone cheaply, the agent does it. Drafting an email, tidying a spreadsheet, opening a pull request, writing a report, renaming a file, researching a supplier: all reversible, all done without asking. If the action cannot be undone, it waits for a human yes. Three kinds of step fall on that side in every ABC system: payments, anything going to production, and messages out to customers or the wider world.
The logic is simple once you say it aloud. A bad draft costs nothing; you delete it. A bad payment costs money and an awkward phone call. A bad deploy costs a broken site at the worst possible moment. A bad email to a customer costs something harder to get back. Agents are allowed to be wrong cheaply as often as they like. They are not allowed to be wrong expensively even once.
In practice the gate is a short queue. The agent prepares the thing, the invoice, the merge, the reply, and puts it in front of the owner with the evidence beside it. The owner reads it, perhaps changes a word, and approves. Most approvals take less time than reading this paragraph. The machine has done the work; the human has done the deciding, and the deciding is the part that needed a human.
This also applies to the machine's own code. No agent merges its own work into production. It proposes, the checks run, a human merges. However confident the agent sounds at seven in the morning, and agents always sound confident, the button belongs to a person.
The gates are not permanent fixtures. Each approval teaches the system what good looks like, and over months an owner may choose to move a line: letting the agent send routine reminders unsupervised, say, once it has drafted fifty perfect ones. That choice is always the owner's, and it is always written down, so nobody discovers the line moved by accident.
Epictetus divided the world into what is up to us and what is not. ABC divides it into what can be undone and what cannot, and puts a human at every door of the second kind. The agents handle the volume. The owner holds the keys. It is a very old arrangement, and it works.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Threads, Not Chats
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A chat is a lovely place to think and a terrible place to keep work. Everything scrolls away. Three ideas, two decisions and a request for an invoice end up in the same river, and by Thursday nobody can find which bit was a decision and which bit was thinking aloud. A chat remembers everything and organises nothing.
ABC splits work into threads . Every ask, however small, becomes its own thread with three things attached: a goal, a checklist and an output. The goal says what the thread is for in one line. The checklist shows the steps, each marked not started, in progress or done. The output is the thing that proves the thread finished: a merged pull request, a live URL or a file. Nothing else counts.
The chain runs left to right. An ask arrives. It becomes a thread. The thread grows a checklist. The checklist ends in an output. The emphasis falls on that last link, because that is where most work quietly dies. Plenty of activity happens in the middle; plenty of it never turns into anything you could point at. A thread that cannot name its output is not finished, however busy it looked.
Threads are coordinated from one place: a single project chat where the owner talks to the master agent. The owner says what they want in plain words. The master opens the threads, assigns the agents and reports back. The owner never needs to remember which thread holds what, because the project chat lists them all, with their state beside each one: working, blocked, ready for review, done.
This turns out to be very good for running several things at once. On a busy morning there might be a thread building an app, another fixing a back office feed and a third drafting a proposal, each moving at its own pace and none of them treading on the others. When one gets stuck, it says so in its own thread and waits for an answer, rather than holding up the rest.
Threads also make good history. Six months on, when someone asks why the pricing page changed, the answer is not lost in a chat log the length of a novel. It is in a thread with a title, a checklist and a link to the pull request that changed it. Work that was done in threads can be found again, which is more than can be said for work done in conversation.
There is something calming about it, too. A chat feels endless, because it is. A thread has edges. It starts, it moves, it ends, and when it ends you can close it and stop thinking about it. That small sense of completion, repeated many times a week, is a surprisingly good substitute for a holiday.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
When Agents Fail
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Agents do not fail like people. A person who is stuck tends to look stuck: they go quiet, or they ask, or they sigh loudly near the kettle. An agent that is stuck often does the opposite. It reports success, in a clear and cheerful voice, about a task it did not complete. It has fixed the bug, it says, and the bug is still there. It has updated the feed, and the feed is still empty. Agents fail quietly and confidently , which is the most expensive way to fail.
So ABC does not take an agent's word for anything. Every action is followed by a check, and the check is not performed by asking the agent whether it is sure. It looks at the world. Did the pull request merge? Does the URL load? Does the file exist, and does it contain what it should? Did the main branch's checks go green? These questions have answers that do not depend on anybody's confidence, which is exactly why they are asked.
The loop is three steps long: act, check, correct. The check carries the weight. Acting is easy; agents will act all day. Correcting is usually straightforward once you know what went wrong. The hard part, and the part most systems skip, is the plain, unglamorous moment of looking to see whether the thing actually happened.
This is why ABC is strict about evidence. An output counts only if it is a merged pull request, a live URL or a file. A claim with no link does not go into the morning status. A board with no feed shows an honest empty box rather than a guess. These rules feel fussy on a good day. On a bad day they are why anybody notices the bad day before a customer does.
Logs help too. Every agent action leaves a trail in the thread where the work lives: what it tried, what it found, what it claimed. When something goes wrong, the question is never "what happened?" asked into a void. It is a matter of reading back a few lines until the confident sentence with no evidence under it turns up.
The main branch is the last line of defence. In Mat's machine it must be green every weekday, and when it is not, fixing it outranks everything else on the list. A red main is a small fire, and small fires are the cheap kind. Left alone, they spread into every change that follows.
None of this is distrust in the personal sense. Agents are not liars; they are optimists with no sense of embarrassment. The remedy for an optimist is not a lecture. It is a checklist, a log and someone, or something, who goes and looks. Seneca advised reviewing each day before sleep to see what had gone wrong. ABC reviews each action before the next one, which is the same idea with less candlelight.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
A Team That Sleeps Never
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Put the whole of Part VI on one page and it reads like a hiring plan for a company that could not afford one. One master agent, ABC Master, who sees every pillar and routes the work. Five pillars of specialists beneath it, each with a lead and a short job description. A canon of skills in one repository, synced to every surface. Agents built from five plain parts. Handoffs that carry their context. Humans at the gates. Threads instead of chats. Checks after every action.
Now look at what that roster costs and when it works. The diagram has two axes. Along one runs cost; along the other runs hours on. A human team sits where you would expect: the more hours of cover you want, the more it costs, steeply, with night rates and the faint whiff of burnout. An agent roster sits somewhere a human team cannot reach: low cost, round-the-clock coverage , for roughly the price of a subscription and a modest monthly allowance for the models.
The practical effect is that the business keeps working while the owner does not. The app agent finishes a build at two in the morning. boai reconciles the overnight payments before anyone is awake. The vault agent tidies the notes. By breakfast the work is done, checked and waiting, and at 08:45 the morning status says so, line by line, with evidence.
This is not a story about replacing people. Most ABC clients are one person, or a handful, and the alternative to an agent roster was never a full team. It was the owner, doing everything, badly, at eleven at night. The roster does not take anyone's job. It takes the jobs the owner was doing grudgingly when they should have been asleep.
The discipline of the last nine chapters is what makes the night shift safe. Without the gates, an overnight agent could pay a supplier twice. Without the checks, it could report a feed fixed that is still broken. Without the threads, nobody could find what it did. An agent that works all night is only an asset if you can trust what it did in the dark, and trust is built from boring rules applied every time.
There is a limit, and it is worth naming. Agents cover hours, not judgement. The decisions that matter still wait for the owner, and they still arrive at a civilised time of day. The roster does the volume; the human does the choosing. That division of labour is the point of everything in this part.
So here is Part VI in a sentence. One master, clear skills and firm gates can keep a business moving all night for the price of a subscription. The owner wakes up to a short report and one decision. Which is, when you think about it, how most people would like to wake up anyway.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Five Screens
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Business software has a habit of growing screens the way old houses grow extensions. Someone has a question, a screen appears to answer it, and ten years later nobody can find the kitchen. ABC takes the opposite view. An owner needs a small number of screens, opened often, and the rest of the house can be knocked down.
There are five, and they sit around one centre. In the middle is the cockpit , ABC OS itself, which is less a screen than a front door: the place you land, and the place every other screen is launched from. Around it sit the five screens an owner actually opens: the dashboard, the apps, the back office, the context and the agents. Everything the business does can be found from one of them in a click or two.
The dashboard is today. The goal, today's figure, what moved, what is stuck and what needs a yes. It is the screen you open first and, on a good day, the only one you need. The apps screen is the fleet: every app the business runs, its status, its live link and whether it is earning. The back office screen is the money and the records, every feed and every asset, with the honest empty boxes where a number is not yet connected.
The context screen is the memory of the business: the goal, the decisions, the people and the rules, written once in plain text. It is the screen you open least, because it changes least, and that is as it should be. The agents screen shows the roster: who is working, on what, in which thread, and what each has produced. It is where you look when you want to know not what happened, but who did it.
The discipline is in refusing a sixth. When a new question comes up, the first move is to answer it on one of the existing five: a new card on the dashboard, a new column in the apps table. Only when it truly cannot fit does a new screen get discussed, and even then it is usually a sign that the question was really two questions wearing a coat.
Five is also a number the head can hold. On the first day you learn where everything lives, and on the hundredth day it still lives there. Nothing moves because a designer had a quiet afternoon. Familiarity is underrated as a feature; it is the reason you can find the light switch in your own house in the dark.
The test is simple and slightly brutal. If an owner cannot find something in two clicks from the cockpit, either it is in the wrong place or it should not exist. Most of the time, on inspection, it should not exist, and the house is better for the demolition.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Cockpit
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every machine needs a place to stand while you look at it. In ABC that place is ABC OS , the cockpit, a small private web app that runs on port 25000 and does two jobs: it shows the state of the whole business, and it launches everything else. It is a status board and a front door, and it tries very hard not to be anything more.
The cockpit is built in layers, and the order of the layers is the design. At the top sit the stat cards: a handful of numbers, each in its own box, each fed from a real source. The goal and today's figure come first, because they always come first. Below the cards sits the apps registry table, one row per app in the fleet, with its status and its link. Below that sits the pillar status, one line each for Apps, Backoffice and Context, saying whether each pillar is healthy. At the bottom sit the launch links, the doors through to every other screen.
The cockpit leads with the numbers, and that is deliberate. Most software leads with navigation, as if the first thing you want on opening it is a menu. It is not. The first thing you want is to know whether things are all right. So the top of the screen answers that, in figures, before it offers you anywhere to go. If everything is fine, you can close it again in ten seconds, which is the best possible review of a status board.
The registry table is the quiet hero. It lists every app the business runs, from the small ones to the cockpit itself, and it is the single answer to the question "what do we actually have?" That question sounds trivial. In most businesses it takes an afternoon and two arguments. In ABC it is a table, kept current by the agents, with each row linking to its evidence.
The cockpit is also stubbornly ordinary underneath. It is a Next.js app with a plain data file behind it, built in the same house style as everything else, so any agent or developer who has seen one ABC app can find their way round this one. It is private, it is never pushed anywhere public, and it runs on a fixed port so that it is always where you left it.
There is a temptation to make a cockpit impressive: dials, animations, a map of the world with dots on it. Resist it. The cockpit is visited every day, briefly, by someone who wants to get on with their morning. Impressive wears off by Wednesday. Clear lasts.
Pilots do not admire their instruments. They glance at them, see that the needles are where they should be, and look back out of the window. That is the whole ambition of ABC OS: to be glanced at, believed and left alone, until the day a needle moves.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The 08:45 Status
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every weekday at 08:45 the machine sends the owner a short message. It has the same shape every day, it takes about a minute to read, and it is, without much competition, the most valuable thing ABC produces. Everything else in the system, the apps, the back office, the agents, is in some sense there to make this one message true. It is the heartbeat of the business.
The status runs in a fixed order, from ABC to the owner. First, yesterday: what moved, what shipped, what earned, each line linked to its evidence. Then today: what the agents are working on, what is stuck and why. And last, the one decision the owner needs to make, framed so that it takes a minute. The order matters. You learn what happened before you are asked to decide anything, so the decision is made with the facts already in your head.
The time is not arbitrary. Before quarter to nine, most owners are dealing with the morning: the school run, the commute, the overnight emails. After it, the day starts making demands of its own. The status lands in the gap, so the first decision of the working day is made with a clear picture rather than a vague sense of dread and an inbox.
The rules are strict, and the strictness is what makes it trustworthy. Every claim carries a link: a merged pull request, a live URL or a file. A line with no evidence does not go in. If a feed is not connected, the status says so, plainly. If nothing moved yesterday, the status says that too. A status that is allowed to be vague will become vague, and a vague status is worse than none, because it feels like information.
It is also deliberately short. The temptation, with a capable machine, is to report everything it did. That produces a document nobody reads by Thursday. The 08:45 status reports only what changed, what is stuck and what needs you. Detail is a click away for anyone who wants it, which, on most mornings, is nobody.
The status also keeps the machine honest about itself. It goes to the owner every weekday whether there is good news or not. A quiet week shows up as a thin status, and a thin status is a prompt to ask why. A machine that reports every morning cannot drift for long without someone noticing, and noticing early is most of what management is.
Over time the message becomes part of the furniture of the day, like the kettle. You stop thinking about it and start relying on it. A business that knows where it stands every morning no longer needs a meeting to find out, and for a company of one that is a meeting it never has to hold.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
One Decision a Day
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
If the agents do the work, what is left for the owner? The answer is the part that was always the owner's: choosing. In ABC the owner's job shrinks, on most days, to one decision . Not a pile of tasks, not a scroll of notifications, but a single choice, framed well enough to take about a minute.
The framing is the craft. A badly framed decision arrives as a question with no shape: "what do you want to do about the pricing page?" That kind of question costs twenty minutes of thinking before you can even begin to answer. A well framed one arrives with the work already done. It states the question in a line, offers two to four options, says what happens with each, and marks one as recommended, with a short reason. The owner's job is then to read, consider and choose.
The diagram makes the shape plain. There is one split, and it hangs on whether the recommendation is right. If it is, you approve, and the machine gets on with it. If it is not, you pick another option, and the machine gets on with that instead. Either way, the decision takes a minute and the work restarts at once. Most days, the recommended option is the right one, because the agent that framed it has read the context, the goal and every previous decision first.
That does not make approving a rubber stamp. The point of a recommendation is to save you thinking from scratch, not to save you thinking at all. Read the reason. If it does not move the goal, say so. The best owners approve most of the time and overrule occasionally, and the occasional overrule is worth more than all the approvals put together, because it teaches the machine something it did not know.
One decision a day is also a ceiling, not just a convenience. Owners who make twenty decisions a day make most of them badly, in a hurry, between other things. Narrowing the daily load to one, maybe two, means each gets proper attention. The rest are either reversible, and so left to the agents, or not urgent, and so left to tomorrow's status.
Decisions are written down once made. The choice, the reason and the date go into the thread where the work lives, and the important ones go into the context, so that no agent asks the same question twice and no future version of you reopens a settled matter on a tired afternoon.
Marcus Aurelius wrote that if you remove your judgement about a thing, you remove the distress. ABC is a little less ambitious. It does not remove your judgement. It simply hands it one clean question a morning, with a recommendation attached and a cup of tea within reach.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Checklist Card
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Reports are written to be read. Checklists are built to be glanced at. That difference sounds small and turns out to be most of what an owner needs. A report about a piece of work tells you what someone thinks of it. A checklist tells you where it is. On a busy morning, where it is will do nicely.
In ABC, progress is shown as the checklist card . Every piece of work in a thread carries one: a short header saying what the work is, then a line for each step. Each line is in one of three states: not started, in progress or done. That is the entire vocabulary. There is no "nearly done", no "blocked pending clarification", no percentage that has been ninety per cent for a fortnight. A step has either not begun, is happening now or has finished.
Every task moves left to right, visibly. A line starts as not started, becomes in progress when an agent picks it up and becomes done when its evidence exists. The card is replaced whole each time it changes, so what you see is always the current state, never a history you have to read to the bottom of to understand. The emphasis falls on done, because done is the only state that pays the bills.
The beauty of the card is how little it asks of the reader. You do not need to know the code, the agents or the plan. You look at the card, count the ticks and know where things stand in about three seconds. If most lines are done and one is in progress, the work is nearly there. If everything is still not started at lunchtime, something has gone wrong, and you know which thread to open.
The card also keeps the agents honest. A step can only be marked done when its output exists, a merged pull request, a live URL or a file, so the ticks mean something. An agent cannot write its way to a finished card the way it might write its way to an optimistic report. Before a result is handed over, every line is brought to its true end state, and the card and the result have to agree.
There is an old joke that a project is ninety per cent done for the second ninety per cent of its life. Checklists are the cure for that particular disease. A task broken into visible steps cannot hide in a percentage. Either the line is ticked or it is not, and everyone can see which.
The owner, meanwhile, never has to ask "how is it going?", the question that interrupts more work than any other. The card already answers it, in silence, before anyone thought to ask. Which is roughly the most a good colleague can do.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Read It on a Phone
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Owners are rarely at desks. They are in car parks, kitchens, school queues and the quiet carriage of a train that is not, in fact, quiet. The desk is where the business gets imagined. The phone is where it actually gets run, usually with one thumb, often while holding a coffee in the other hand.
So ABC is built at phone width first and widened afterwards, never the other way round. Every page of the cockpit works on a small screen with no sideways scrolling. Tables fold into cards, the sidebar folds into a row of icons, and the buttons are big enough for a thumb that has not had its breakfast. The desktop view is simply the phone view given more room to breathe.
The diagram for this chapter is three circles. One is the phone. One is two minutes. One is enough. The useful part is where all three overlap: what you can read on a phone, in two minutes, that is enough to steer the business for the day. Not enough to do the work. The agents do the work. Enough to point it in the right direction and get on with your morning.
That overlap turns out to be a ruthless editor. A status that needs scrolling for a full minute is too long. A chart that only makes sense on a large monitor is a chart for its designer, not for its owner. A decision that needs three screens of background is not ready to be asked. If something cannot survive the phone, it gets cut, shortened or moved somewhere you can find it later when you choose to.
Approvals follow the same rule. When an agent needs a yes, the request arrives with the recommended option, the evidence one tap away and two buttons. Approve, or pick another. You can do it at a red light, though the highway code would rather you did not, and so would we.
None of this means the desk is obsolete. Specs still get written at a keyboard, the Friday review deserves a proper chair, and some thinking needs a big screen and a closed door. The point is that the daily business of steering should never wait until you get back to the office, because by then the day has usually made its own decisions without you.
Marcus Aurelius ran an empire from a tent on campaign and wrote his best notes to himself in short, plain lines that would fit very comfortably on a phone. He did not wait for a desk to think clearly. Neither should you. The business does not pause until you sit down, and its status should not either.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Weekly Review
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The weekdays are for shipping. Friday afternoon is for looking back. ABC closes every working week with the weekly review , a short, fixed-shape look at what happened, and it is the one moment in the week when the machine stops asking what is next and asks what was worth it.
The review has three questions, always in the same order. What shipped? What did it earn? What did we learn? Shipped means outputs with evidence: merged pull requests, live URLs, files. Earned means the movement on the goal, taken from the scoreboard, not from memory. Learnt means the one or two things that should change how next week runs. Three questions, answered honestly, take about fifteen minutes. Answered dishonestly they take all afternoon and produce a slide deck.
The diagram is a loop: ship, measure, review, and back to ship. The review is where the loop closes. Without it, the week is just a series of busy days that happen to sit next to each other. With it, the week becomes an experiment with a result, and next Monday starts from what was learnt rather than from whatever feels most urgent at 09:00.
This is also where the fleet verdicts start. Every app in the fleet has to earn its place, and Friday is when each one is looked at against its numbers. The verdict is one of three words: scale it, hold it or kill it. Most weeks most apps are held, which is fine, because holding is a decision too. The important thing is that no app drifts for months without anybody saying a word about it.
The machine prepares the review so you do not have to. By Friday lunchtime the agents have gathered the week's outputs, the scoreboard movement and any app whose numbers changed. Your job is to read it, agree or disagree, and write the one sentence that sets next week's focus. The Friday 08:45 status carries the summary, so the review starts from a page rather than a blank.
Seneca described ending each day by going over it in his head: what he had done well, where he had slipped, what he would do differently. He found it calming rather than harsh, because a fault you have named is a fault you have already half fixed. The weekly review is the same habit at the scale of a business, with better records than Seneca had.
Then the review ends, and so does the week. ABC runs on weekdays only, and Friday's look back is the signal that the machine has said what it needs to say. Whatever the verdicts were, they will still be true on Monday. The weekend is not on the agenda, which is perhaps the most useful thing the agenda does.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Quiet by Design
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A machine that works all night does a great many things. It opens branches, runs tests, retries a failed fetch, tidies a folder, updates a checklist, reads a file, writes a draft and reads it again. On a busy night the agents might log hundreds of small events. If each one became a ping on your phone, ABC would be the most efficient way ever invented of ruining your breakfast.
So ABC is quiet by design . It notifies you for three things only: a result, a blocker or a question. A result is an output that now exists, such as a merged pull request, a live page or a finished file. A blocker is something that has stopped and cannot restart without you. A question is a decision that only you can make. Everything else happens in silence, written down for anyone who wants to look, and announced to nobody.
The diagram is a funnel. At the top is everything that happens. In the middle is the much smaller set of things that genuinely need you. At the bottom is one message, usually the 08:45 status, which gathers those things into a few lines. The funnel is the product. Anyone can build a system that reports everything. It takes judgement to build one that reports almost nothing and still leaves nothing out that matters.
This shapes how the agents behave. Work is coordinated from one project chat, where each piece of work runs in its own thread with a checklist that updates as steps complete. Updating a checklist notifies no one. You can look at it whenever you like and see exactly where things stand, but it will never tap you on the shoulder to tell you that step three of nine has moved from not started to in progress. That information exists. It simply does not need to arrive.
Calm has a practical value beyond comfort. When notifications are rare, each one gets read. When a blocker arrives you know it is real, because the system does not cry wolf about retries and warnings it can handle itself. Noisy software trains you to ignore it, which is exactly the opposite of what you want from the thing that runs your business.
There is a reliability point too. A quiet system has to be confident about what it is not telling you. That confidence comes from evidence: the work was done, the checks passed, the log exists. Silence in ABC is not a gap in the reporting. It is a report that everything below the line is fine.
Most events never become a notification, and that is the whole idea. The best compliment anyone can pay a machine like this is that they forgot it was running, right up until it handed them something finished.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Book You Get
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most software projects begin with a meeting, then a proposal, then a contract, and then, some weeks later, a first look at something that may or may not resemble what was discussed. ABC begins with a book. Every client who finishes onboarding receives a V1 book of their own business : twenty-five pages about their goal, their workflow, their agents and their deal, written in plain English and readable in about twenty minutes.
It sounds like an unusual thing to hand someone before any code exists. It is meant to. The book is the cheapest possible version of the system. It describes what will be built, who it is for, what the agents will and will not be trusted with, what it costs and what happens first. Nothing has been built yet, which means nothing has to be unbuilt when the book turns out to be wrong about something.
And it will be wrong about something. Onboarding is quick by design, and every question has a sensible default, so some of the book is built from what you said and some from what we assumed. The assumptions are marked. That is the most useful feature of the whole document. It tells you exactly where to look first, and it turns the awkward question "did we understand you?" into a short list of sentences you can correct over a cup of tea.
The diagram is a chain: onboard, V1 book, correct, build. The book sits in the middle on purpose. It is the checkpoint between hearing and building, and it is where most of the expensive mistakes in software are caught for the price of an edit. Telling us the workflow has an extra step costs one sentence on page nine. Discovering it after the app is live costs a week.
The book also does something less obvious. It gives you language for your own business. Many owners have never seen their workflow written down in order, with each step named and each hand-off visible. Reading it back is often the first time they notice the step that hurts most, or the decision they have been quietly avoiding. Several have said the book was worth having even if nothing else followed. We prefer that something else follows, but it is a nice thing to hear.
The V1 is not the end of the story. When the system earns it, the book grows: a V2 of a hundred chapters, then a V3 of a thousand items, each one deeper and each one still about your business. The book you are reading now is a grown-up version of exactly that, written about ABC itself.
So before anything is built, you get something you can read, mark up and argue with. Correct it, and the build follows the paper. Most clients find that reassuring. The rest find it faster, which is also reassuring.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
It Feels Like Help
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
This part of the book has been about how ABC feels, which is an odd thing to write about software. Software is usually described by what it does. But owners do not live with features. They live with a feeling, every morning, about whether the thing helps or whether it is one more thing to manage. That feeling decides whether a system gets used for a year or abandoned by March.
The diagram is a simple grid. One axis is the effort a system asks of you. The other is the help it gives back. Most business software lands in the corner of high effort and modest help: a control panel , full of switches, that does a great deal once you have learnt which switch does what. ABC aims for the opposite corner. Little effort asked, much help given. Not a control panel. A capable colleague .
Look back at the chapters in this part and the pattern is plain. Five screens, not fifteen, because a colleague does not make you search for things. One cockpit that leads with the numbers, because a colleague tells you where you stand before you ask. The 08:45 status, short and the same shape every day, because a colleague reports rather than waits to be chased. One decision a day, framed with a recommendation, because a colleague does the thinking and leaves you the choosing.
Then the checklist card, which shows progress without asking you to read a report. The phone, because a colleague finds you where you are. The Friday review, because a colleague helps you learn from the week rather than simply survive it. Quiet by design, because a colleague does not interrupt you to say they are still working. And the book, because a good colleague checks they have understood before they start building.
None of these choices is technically difficult. Each one is a refusal: to add a screen, to send a notification, to ask a question that could have been answered, to show a chart nobody reads. The difficulty is holding the line when a new feature looks impressive in a demo and would quietly make the whole thing slightly more work to live with. ABC holds it, because effort asked is a cost you pay every single day.
There is a test for whether it is working, and it is not a metric. It is the moment an owner stops describing the system as a tool and starts describing it as though it were a person. "It told me the invoice was stuck." "It asked whether we should drop the second app." That shift in grammar is the whole of Part VII.
Software you operate is a job. Software that helps is a colleague. We know which one we would rather hire, and so, it turns out, does everyone who opens their phone at 08:45.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The House Stack
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Part VIII opens the bonnet. Before the agents, the pipelines and the evidence, there is the stack: the handful of technologies every ABC system is built on. It is short and it is deliberately boring. Next.js and React for the web app and its API. TypeScript for the code. Tailwind and shadcn/ui for the interface. SQLite for the data. Four layers, all well known, all widely documented, none of them likely to surprise anybody on a Tuesday.
Boring is the point. A fashionable stack is exciting for about six months and expensive for about six years. Every unusual choice is a future question that only one person can answer, and in a company of one that person is busy. A boring stack means any competent developer can open the code and recognise it, and, increasingly, any competent agent can too. The cockpit that runs the whole business, ABC OS on port 25000, is built on exactly this stack. So is every app in the fleet.
The diagram highlights one layer, and it is TypeScript. If the stack has a secret weapon, it is the type system. Types are contracts written into the code itself: this function takes a customer, returns an invoice, and will refuse to compile if anyone hands it a shoe. That matters more when agents write a large share of the code. An agent can be confidently wrong. The compiler cannot be talked round. Every change has to agree with every type before it is allowed anywhere near production, which turns a whole class of mistakes into red underlines rather than angry customers.
Tailwind and shadcn/ui give the interface a shared vocabulary. Buttons, cards, tables and sidebars come from one design system, so every screen looks like it belongs to the same business, and an agent building a new page does not invent a new shade of blue. Consistency is not decoration here. It is what makes five screens feel like one product.
At the bottom sits SQLite, a database that lives in a single file. That sounds modest until you notice what it means: the data can be copied, backed up and moved with ordinary tools, and a small app needs no separate database server to look after. Where a client genuinely needs more, such as hosted sign-in or storage in a particular region, a larger service can take its place. The default stays small until something proves it should not.
None of this will win an award for originality. It will, however, still be running, readable and fixable in three years' time, by someone who has never met us. For the foundations of a business, that is the only award worth having.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Spec First
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every behaviour in an ABC system starts life as a few sentences on paper. Not a sketch on a napkin or a thought in the shower, but a written spec : what the thing should do, who it is for, and how anyone will know that it works. The code comes later. The paper comes first, and the paper wins any argument.
The heart of a spec is its acceptance criteria . These are short statements a stranger could check. "When a payment fails, the customer receives one reminder after three days." "The scoreboard shows an empty box, not a zero, when a feed is not connected." Each one is either true of the finished software or it is not. There is no room for "broadly", "mostly" or "it depends how you look at it", which are the three most expensive phrases in software.
The diagram is a chain of four: spec, PBI, build, test. The spec becomes a PBI , a product backlog item with a number, such as the BL-076 that governs the scoreboard. The PBI is what gets scheduled, assigned to an agent and tracked on a checklist. The build is the agent writing the code. The test checks the code against the criteria in the spec, which closes the loop neatly: the paper that started the work is the same paper that decides when it is finished.
This order matters more with agents than it ever did with people. A human developer asked to build something vague will come back with questions, a raised eyebrow and possibly a counter-proposal. An agent will simply build something, quickly and with total conviction. If the request was vague, the result will be precisely, efficiently wrong. The spec is where vagueness gets caught, before it has been turned into a few hundred lines of very confident code.
Specs are also cheap to change, and that is half their value. Rewriting a sentence takes a minute. Rewriting a feature takes a day and usually breaks something nearby. So the arguments happen on paper, where they are fast, rather than in code, where they are slow and occasionally involve a rollback on a Friday afternoon. Most of the thinking in an ABC build happens before a single file is touched.
There is a quieter benefit too. A spec is documentation that was written at the moment anyone actually cared. Six months later, when someone asks why the reminder waits three days rather than two, the answer is not buried in a chat or in somebody's memory. It is in the spec, with the PBI number beside it.
So the rule is simple and holds for every feature, however small: no spec, no build. It sounds bureaucratic. In practice it is the fastest route to finished work, because nothing has to be built twice when it was understood once.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Pull Requests or Nothing
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
In an ABC system, code does not simply change. It arrives. Every change, whether written by a person or an agent, comes in as a pull request : a proposed edit, with a description, a list of the files it touches and a record of the checks it has passed. If a change is not a pull request, it does not reach the main branch. There is no side door, no quick fix pushed straight to production, and no "I will tidy it up later".
The diagram is a short sequence. An agent does the work on its own branch. The branch goes to GitHub. A pull request is opened, describing what changed and why, with a link back to the PBI it serves. The automated checks run: linting, type checks, tests and a build. Only when those pass, and when the change has been reviewed, is it merged. Five steps, the same every time, for a one-line typo and for a new feature alike.
This is also how ABC counts work. An output only exists when it can be pointed at, and a merged pull request is the clearest thing you can point at. It has a number, a date, a description, an author and a diff that shows, line by line, exactly what changed. "I improved the checkout" is a sentence. A merged pull request is evidence.
The rule matters more when agents are writing much of the code. Agents are quick and tireless, and occasionally they are quick and tireless in the wrong direction. The pull request is the moment where that gets noticed. The agent that builds the code never merges its own work, and anything that cannot be undone, from a deletion to a release, waits for a human yes. Speed is welcome. Unreviewed speed is not.
That line is the quiet benefit of the whole approach. Because every change came in through a pull request, the repository holds a complete, readable history of the system. When something breaks, you can find the change that broke it, read why it was made and reverse it cleanly. When someone new arrives, human or agent, they can read the history like a diary and understand not only what the code does but how it came to do it.
Without that history, every bug is a mystery and every fix is a guess. With it, problems become small, specific and dated. Most of the fear people feel about software comes from not knowing what changed. Pull requests remove that fear by making every change a document.
So the rule stands, without exceptions for urgency, seniority or confidence: pull requests or nothing. It costs a few minutes per change and saves whole days when something goes wrong. That is the kind of trade a small business should accept every time it is offered.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Green CI
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every ABC repository has a test suite, and every pull request runs it, along with the type checks, the linter and a full build. Together these are called continuous integration, or CI , and the result of each run is reported in one of two colours. Green means everything passed. Red means something did not. There is no amber, which is one of the more restful things about it.
The best way to think about CI is as the machine's immune system. Most of the time you do not notice it. It sits quietly in the background, checking every change against everything the system is supposed to do, and catching the small infections before they spread. A broken import, a renamed field, a function that now returns nothing: CI notices within minutes, long before a customer would.
The rule that goes with it is short. Main stays green every weekday, and no red run survives a working day. Red is allowed to happen. It happens all the time, because people and agents make mistakes and the whole point of CI is to catch them. What is not allowed is for red to become the normal state of affairs. A failure found in the morning is fixed, or the change that caused it is reversed, before the day ends.
The diagram is a loop: red, fix, green, with green highlighted. Green is the resting state. Red is a visit, not an address. The loop exists because a red build that lingers does a peculiar kind of damage. Once one failure is tolerated, a second one hides behind it, and soon nobody can tell which red matters. People start ignoring the colour altogether, and an immune system that is ignored is no longer an immune system. It is a light that flashes.
Agents make this rule easier to keep than it used to be. When a run fails overnight, an agent can read the log, find the cause and open a pull request with the fix before anyone has woken up. If the fix is not obvious, the failure goes into the morning status as a blocker, with a link to the run, and it becomes the first thing worth a decision that day. Either way, nobody discovers on Thursday that the build has been broken since Monday.
The discipline also protects everything else in this book. Pull requests are only safe if their checks mean something. Evidence is only trustworthy if the system that produced it is working. A scoreboard is only honest if the code feeding it passes its tests. Green CI is the floor the rest of the machine stands on.
So every weekday morning starts from green, and every weekday evening ends there. In between, red may come and go. It is welcome to visit. It is simply not invited to stay the night.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Evidence Files
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Chapter six set out the rule: an output counts only when you can point at it, as a merged pull request, a live URL or a file. This chapter is about where the pointing happens. Every ABC system keeps an evidence folder , a plain, dated record of what was claimed and the link, file or log that proves it. It is the least exciting folder in the repository and the one you will be gladdest to have.
The diagram is a funnel. At the top are all the claims made in a week: by agents in their threads, by checklists turning green, by reports and by the owner in a good mood. In the middle are the claims that come with a link. At the bottom is the trusted status, built only from those. A claim that never acquired a link simply stays at the top of the funnel. It is not deleted or punished. It just does not make it into anything anyone relies on.
What goes in the folder is deliberately mundane. A pull request number. A URL that returns a page. An export from the payment processor. A screenshot of a deploy. The log of a test run. Each entry has a date and says which PBI or goal it belongs to. Nobody needs to write prose about it. The evidence is the prose.
The folder turns out to be an excellent audit trail , which is the grown-up name for being able to answer "how do you know?" without going red. When a client asks whether the reminder feature shipped, the answer is a link. When the scoreboard moves, the receipts sit beneath it. When something goes wrong months later, the folder tells you what was true on the day and what it rested on. Memory is generous to itself. Folders are not.
This matters especially when agents do the work. An agent will report that it has finished, cheerfully and in good faith, and it will usually be right. Usually is not a standard you can run a business on. So agents are asked for the link along with the claim, and the status only includes claims that arrive with one. An agent that says "done" and attaches nothing has not finished. It has described an intention.
There is a nice side effect. Because evidence is collected as the work happens, nobody ever has to assemble it afterwards. The Friday review, the client update and the end-of-year look back are mostly a matter of reading the folder in order. The hard part was done one link at a time, on ordinary days, when it cost almost nothing.
So the folder grows quietly, entry by entry, and is mostly ignored. That is exactly right. Evidence is like insurance: dull to keep, and the only thing anybody wants to see when the story is in doubt.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Cheap to Run
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Software has a habit of costing more to run than anybody expected, slowly, in small increments, each one too modest to argue with. ABC learnt this the ordinary way. The fleet of small apps started out on Vercel, which is excellent, fast and pleasant to use. Then the bills crept up. The cause was server-side rendering : pages built fresh on the server for each visitor, which is useful for some apps and expensive for a fleet of small ones that mostly show the same content to everybody.
None of the individual charges was alarming. That was the problem. A few pounds here and a few more there, multiplied across a growing fleet, turned into a line on the statement that no longer matched what the apps were earning. For a company of one, where each app is meant to earn its keep, a hosting bill that grows faster than the revenue is not a technical detail. It is a business decision someone forgot to make.
The answer was to step back and ask what the apps actually needed. Most of them needed very little: a few pages, a small amount of data and a steady place to live. So the alternative became a cheap VPS , a small rented server, with SQLite holding the data on the same machine. On that arrangement, roughly a hundred small sites can run for around five pounds a month. Pounds, not hundreds. The number is almost embarrassing, in the good way.
The diagram plots two things against each other: monthly cost and control. Managed platforms sit at higher cost with less control, because you are paying for convenience and accepting their way of doing things. The VPS sits in the opposite corner, low cost and high control. You decide what runs, where the data lives and how it is backed up. In exchange, you look after the server yourself, which for a fleet of simple sites, with agents to do the looking after, turns out to be a modest price.
This is not a sermon against managed hosting. Some apps genuinely need what a platform provides, and for those the bill is worth paying. The lesson is narrower and more useful: choose hosting per app, on purpose, rather than inheriting one default for everything. Static pages where static will do, server rendering only where it earns its cost, and a cheap box for the long tail.
The habit underneath is the same one the rest of the book keeps returning to. Look at the number. Hosting cost is a figure like any other, and figures that nobody looks at grow in the dark.
A small business does not need infrastructure that could survive a global launch. It needs infrastructure that survives a quiet Tuesday without eating the profit. Five pounds a month will do nicely.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Private by Default
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Everything in an ABC system starts life private. Code lives in private repositories . The context, the written memory of the business, is private too. Data, drafts, specs, reports and the evidence folder are all visible only to the people and agents who need them. Nothing goes public by default, and nothing goes public by accident. Going public is always a decision, made on purpose, by a human.
The diagram is a single question with two exits: is this meant to be shared? If yes, it is published, deliberately. If no, or if nobody is quite sure, it stays private. The highlighted branch is the second one. When in doubt, ABC keeps it in. This is not secrecy for its own sake. It is the recognition that the two choices are not symmetrical. Publishing takes a second. Taking something back off the internet, in any real sense, is not on offer.
The same thinking shapes how agents handle the most sensitive material. Mat's own context lives in a repository called context-2026, which holds the goals, decisions, people and rules of the business. Agents may read it when they need to, because context is what makes them useful. They may never write to it. It is read-only to agents , with no exceptions for helpfulness. An agent that wants something changed there proposes it, and a human decides. The memory of the business is not something a machine should be quietly rewriting overnight.
Publishing sits in the same category as payments, deletions and messages to customers: anything that cannot be undone needs a human yes. An agent can prepare a page, a post or a public repository, check it and put it in front of you with a recommendation. It will not press the button. That last step is short, but it is the one that matters, because the internet has an excellent memory and very little interest in your apology.
Private by default also makes the system calmer to work in. When everything is private until decided otherwise, nobody has to remember which drafts are safe to write in, which folder is visible, or whether a half-finished thought is about to appear in a search engine. People and agents can write freely, because the walls are already up. Ideas can be rough in private. They only need to be polished when they are going somewhere.
For clients the principle is identical. Their code sits in their repositories, their context stays theirs, and their data is never mixed with anybody else's. What they decide to share, they share. Everything else stays in the house.
There is an old line about keeping your own counsel, and it was written for people, not software. It works rather well for both. A business that publishes only what it chooses has nothing to explain later, which is a wonderfully restful way to run anything.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Budget Guard
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every small business has a story about the bill. The phone contract that quietly tripled, the subscription nobody remembered starting, the cloud invoice that arrived looking like a telephone number. Agents add a new chapter to this genre, because an agent can spend money far faster than a person can notice. A loop that calls a paid model ten thousand times overnight is not malicious. It is simply diligent, in the wrong direction, while you sleep.
So ABC has one rule about money, and it is not negotiable: every paid model or media call goes through the budget guard . There is exactly one door to anything that costs money, and the guard stands in it. An agent that wants to generate an image, transcribe a recording or ask a large model a long question does not call the provider directly. It asks the guard. The guard checks the request against a limit, lets it through or refuses it, and writes the result down.
The order matters, and the diagram shows it as a stack. At the top is the agent's request. Beneath it, the guard. Beneath the guard, the paid API. At the bottom, the usage ledger , where every call is recorded with what it cost, which agent made it and why. Nothing reaches the provider without passing the guard, and nothing passes the guard without leaving a line in the ledger. You can read it, add it up and find the culprit.
The limit itself is small and set on purpose. For a typical client the monthly AI budget starts at around fifty pounds, with an alert at eighty per cent, so the warning arrives while there is still room to act. When the limit is reached, the guard does not negotiate, apologise or try one more time. It says no, and the morning status says why. A refused call is an inconvenience. An unrefused one can be a small catastrophe with your card details attached.
There is a second benefit, less obvious than the first. Because every call is logged in one place, you can see which agent is expensive and whether it is worth it. Sometimes the costly one is earning its keep. Sometimes it has been rewriting the same paragraph in a slightly different mood for a fortnight. Either way, you now have the numbers to decide.
Seneca, who was very rich, wrote a great deal about not being owned by your money. The budget guard is a modest version of the same idea. It keeps the machine's spending small, visible and decided in advance, so that cost is a setting rather than a surprise.
The guard will never be the most exciting part of ABC. It will simply be the reason you open your invoices with a cup of tea instead of a sharp intake of breath.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Deploy in Five Moves
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There are many ways to get software from an idea to the internet. Most of them are improvised, a little different each time, and remembered only by whoever did it last. That person is usually on holiday when it matters. ABC removes the improvisation with one flow, the appai flow , which every app in the fleet follows, from the smallest single-page tool to the cockpit itself.
The flow has five moves. It begins with a PBI , a product backlog item: one written sentence of what should change and how we will know it has. Then build, preview, push and deploy. The diagram shows the last four as a chain, because the PBI is the ticket rather than the journey. Nothing gets built without one. If the change cannot be described in a backlog item, it is not ready to be a change. It is still a conversation.
Build is where the agent writes the code, against the spec and the tests, on its own branch. Preview is where the change runs somewhere real but harmless, at its own address, so that a human can click it rather than imagine it. This is the step that saves the most grief. A screen you have actually seen is worth a great many paragraphs explaining what the screen will look like.
Push is the pull request: the change, the evidence and the checks, arriving for review. Lint, types, tests and a build all run before anyone looks. If any of them go red, the pull request waits, and the agent fixes it. Nobody merges a red change, however small, however late in the day and however confident the agent sounds about it.
Deploy is the last move, and the shortest, because everything difficult has already happened. A human merges, the main branch updates, and production follows. No agent pushes straight to production, ever. That gate is the one place in the flow where speed is deliberately sacrificed, and it costs about thirty seconds of a person's attention. It is the cheapest insurance in the building.
The real value of five moves is not that they are clever. They are not. It is that they are the same every time. When every app ships the same way, a problem in one tells you where to look in all the others. A new agent learns the flow once and can ship anywhere in the fleet. A stranger can read the history of any app and understand it, because the history has the same shape everywhere.
Routine is often treated as the enemy of creativity. In deployment it is the opposite. Five dull moves, repeated faithfully, leave your attention free for the one interesting thing: what the app actually does for the goal.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Built to Be Handed Over
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Here is a test for any system you depend on. Imagine handing it, tomorrow morning, to a competent stranger. Not a genius, not the person who built it, just a sensible developer with a cup of coffee and no context. Could they run it by lunchtime? If the answer is yes, you own the system. If the answer is no, the system owns you, and possibly whoever built it as well.
Part VIII has been about passing that test. The house stack chose boring, well-known tools so the stranger already knows them. Spec first meant every behaviour was written down before it was built. Pull requests made the history readable. Green CI made the history safe. Evidence files backed every claim with a link, a log or a file. Cheap hosting kept the running costs small enough to survive a quiet month. Privacy by default meant nothing leaked while all this was happening. The budget guard kept the spending decided in advance, and five deploy moves meant every app shipped the same way.
Put all of that together and you arrive at the diagram: three overlapping circles. One is documented : the system explains itself in writing, in the repository, next to the code. One is tested : the system checks itself every time it changes, and the main branch is green every weekday. Where documented and tested overlap, a third quality appears, which is the one that matters commercially. The system is ownable .
Documented without tested is a beautiful manual for a machine that may not work. Tested without documented is a machine that works perfectly and that nobody dares touch. Only together do they produce something a stranger can pick up, change and trust, and that is the same thing as something you can keep.
This matters most on the days you would rather not think about. The developer moves on. The agency changes hands. You decide, for perfectly good reasons, to run the machine yourself, or to bring someone in to run it for you. A handover-ready system turns each of those from a crisis into an afternoon. Nobody needs to be begged for a password or a memory.
It also matters on ordinary days. A system built to be handed over is easier to change, because the next change is a small handover in its own right, from today's agent to tomorrow's. Every pull request is written for a reader who was not there. That reader is usually you, three months later, with no memory whatsoever of why that button is blue.
So that is how it is built: plainly, visibly and with the exit always open. Strangely, the easier a system is to leave, the less anyone wants to.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Start With Onboarding
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most consultancies begin with a discovery phase. It lasts six weeks, involves a great many workshops and ends with a slide deck that tells you, at some length, what you told them in the first meeting. ABC begins with onboarding , which is twenty questions and takes about five minutes. You can do it on a phone, on a train, with a coffee going cold beside you.
The questions are not clever. They are the ones that matter. What is the one goal, with its number and its date? Who are your customers, and what do they buy? Which step of your week hurts the most, and how many hours does it eat? Where does your data live now, honestly? What would you be proud of on the last day of the year? Each question has a sensible default already filled in, so if you do not know or do not care, "ok" is a perfectly good answer. Nobody is marking you.
What happens next is the part people do not expect. The diagram is a short chain: questions, answers, and then a V1 book . Your answers are not filed in a CRM to be looked at by somebody next Thursday. They are read straight away and turned into a book about your business: twenty-five pages, written plainly, with your goal at the centre and a diagram on every page. It covers why the business exists, what the machine would do, how it would feel to use, how it would be built and what the deal would be.
Every assumption in that book is marked. Where you answered, the book uses your words. Where you accepted a default, the book says so, so you know exactly which pages to argue with first. This is deliberate. A document that hides its guesses invites you to trust it. A document that shows them invites you to correct it, which is far more useful.
There is a reason the first output is a book rather than a demo or a quote. A demo shows you what software can do in general. A quote tells you what it costs before you know what it is. A book about your own business shows you whether we have understood you, and it is the cheapest possible place to find out that we have not. Rewriting a paragraph costs nothing. Rebuilding an app costs rather more.
Once you have read it and marked it up, the corrected book becomes the specification for what gets built. The paper comes first, the code follows, and everyone is working from the same twenty-five pages.
So the deal does not start with a price, a meeting or a signature. It starts with twenty questions and the small, pleasant surprise of reading about your own business, and finding it makes rather more sense on paper than it did in your head.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
V1, Then V2, Then V3
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Software projects have a habit of starting large. The first meeting produces a vision, the second a roadmap, and by the third someone has drawn a diagram with forty boxes, all of them labelled "phase one". Eighteen months later phase one is still being delivered, and nobody can quite remember what the vision was for.
ABC starts small on purpose, and grows in three tiers. The tiers are named after books, because in ABC the book always comes before the build. V1 is twenty-five pages: five sections, five items each, one diagram per item. V2 is a hundred chapters. V3 runs to a thousand items. The diagram stacks them with V1 at the bottom, because V1 is the foundation the others stand on, and nothing is ever built from the top down.
V1 is the smallest complete version of the idea. Not a prototype, not a sketch, but the whole proposition in miniature: the goal, the workflow, the agents, the deal. You receive it after onboarding, correct it, and the corrected version becomes the specification for the V1 product. That product is deliberately modest. One core workflow, one master agent, one back office, one morning status. Enough to move the number, and no more.
V2 arrives only when V1 has earned it, which means it is live and the scoreboard says it is working. The V2 book goes to a hundred chapters, the full design of the product and the business around it: a second workflow, integrations, perhaps a client portal. Each addition is written, diagrammed and agreed before a line of it is built. If it cannot survive being written down, it will not survive being built either.
V3 is a different animal. At a thousand items, it is less a book than the operating manual of a company. Very few businesses need it in their first year, and that is fine. It exists so you can see where the road goes, not so you feel obliged to drive the whole of it before lunch. The book you are reading now is a V2, for what it is worth. Its V1 was twenty-five pages and fitted comfortably on a phone.
The point of tiers is not to sell you three things. It is to make each step cheap enough to stop at. If V1 moves the goal, you grow. If it does not, you have spent a modest sum learning something true, and you stop, with a working system and a clear book of what it does. Either outcome is a good one. The only bad outcome is the eighteen-month phase one, and the tiers exist precisely to make that impossible.
Every engagement starts at V1, including the large ones. Especially the large ones.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Three Ways In
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
People come to ABC from different places. Some want the machine built and running and would prefer never to see a pull request in their lives. Some want to learn how it works while it is being built. Some already write code and want the method, the skills and a second pair of eyes. One size would fit none of them well, so there are three ways in , and the diagram draws them as spokes around the same hub.
The first is done for you . You do onboarding, read your V1 book, mark it up, and then we build the machine. You approve what needs approving, read the morning status and make the occasional decision. This suits the owner whose time is the scarcest thing in the business, which, in our experience, is most owners. You are buying the outcome, not the education, and that is a perfectly respectable thing to buy.
The second is done with you . The same machine, built alongside you. You sit in on the decisions, watch the backlog move and learn why things are arranged as they are. By the end you could run most of it yourself, and you understand the parts you choose not to. This suits the owner who wants to keep a hand on the wheel, or who has someone in the team who will eventually look after it.
The third is done by you . You get the method, the skills from skills-2026, the structure and the review, and you do the building yourself, with your own agents. This suits the technical founder who would rather pay for direction than for labour, and who enjoys this sort of thing in the evenings, which is more people than you might think.
What does not change between them is the destination. The same one goal sits at the centre. The same three pillars hold it up. The same rules apply: humans approve anything irreversible, the main branch stays green, outputs count only as a merged pull request, a live URL or a file. Whichever door you use, you end up in the same building, and the building is yours. Nor is the choice permanent. Plenty of people start with done for you and drift towards done with you once they see how the parts fit.
Separately from the way in, there is the size of the build: Lite, Core or Full. The way in is about how involved you want to be. The size is about how much the first version needs to do. The two are independent, and the onboarding answers suggest both.
Marcus Aurelius advised doing what is yours to do and leaving the rest. Each way in is simply a different line drawn around what is yours. Draw it honestly, and the machine takes care of the other side.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Setup and Monthly
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Pricing is where most propositions start to mumble. The numbers arrive late, wrapped in phrases like "it depends" and "let's get on a call", and by the time you see one you have already invested three meetings in not knowing. ABC prefers to say it early and plainly. You pay in two parts: a setup fee to build the V1, and a monthly retainer to run it.
The setup fee is fixed for the agreed scope, and it comes in three bands. Lite, built in one to three weeks, is usually under two thousand pounds. Core, in four to eight weeks, usually sits between two and ten thousand. Full, in eight to sixteen weeks, starts above ten thousand. These are bands, not quotes. The written proposal pins the number to your scope, and once it is pinned, it stays pinned.
The retainer is sized to a small business, typically a few hundred pounds a month. It pays for the machine to keep running: the morning status every weekday, the agents doing their daily work, the fixes, the small improvements and someone watching the costs and the checks. It is deliberately modest, because a system you cannot afford to run is a system you will eventually stop running, and then the setup fee was wasted.
Why split it at all? Because the two parts pay for genuinely different things. The setup fee pays for building something that did not exist before, a one-off piece of work with a beginning and an end. The retainer pays for keeping a living thing alive, which never ends and should never pretend to. Bundling them would hide one inside the other. A huge setup fee with no retainer means the builder has no reason to care after launch. A tiny setup fee with a large retainer means you are renting your own business back.
Notice what is not on the bill: hours. Hours are a poor measure of agent work. An agent can spend four seconds on something that used to take you four hours, and charging for the four hours would be cheeky, while charging for the four seconds would bankrupt the builder. So you pay for outcomes you can see: a workflow that runs, a dashboard that is true, a status that arrives every weekday at 08:45.
The diagram puts the two circles side by side and finds the fair deal where they overlap. The builder is paid properly for building and properly for running, and has a continuing reason to make the machine better rather than merely bigger. You know what the first version costs and what each month costs, and neither number moves without a conversation.
Cheap is what you pay once. Value is what you stop paying for every week, which, conveniently, is the subject of the next chapter.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Payback Maths
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every purchase has a moment, usually late at night, when you ask whether it was worth it. Most software never answers. It simply continues to exist on your card statement, faintly reproachful, like a gym membership with a login. ABC tries to answer the question before you buy, with arithmetic simple enough to do on the back of a receipt.
The diagram is a funnel with three layers. At the top: hours saved a month . In the middle: those hours times your hourly worth . At the bottom, where the funnel narrows to a point: the number of weeks until the machine has paid for itself. That bottom figure is the payback date , and it is the one number in the whole proposition that you should feel entitled to hold us to.
Start with the hours. Onboarding asks which step of your week hurts most and how long it takes. That answer is the raw material. If the step is invoicing, or chasing, or copying the same details between three systems, you already know roughly how long it eats, because you resent every minute of it. The machine will not save all of those hours. It will save most of them, and the proposal says which, step by step, so you can disagree with any line.
Then the hourly worth, which is the figure people most often get wrong, and almost always too low. It is not what you would pay someone else to do the dull task. It is what an hour of your attention is worth when it is spent on the one goal instead. For most owners that is considerably more than they are comfortable writing down. Write it down anyway. Modesty is a virtue, but not in a spreadsheet.
Multiply the two, compare the result with the setup fee and the retainer, and the payback date drops out. The arithmetic is short enough that nobody can hide in it. That is the point. If the date is a few weeks away, the decision is easy. If it is years away, the honest answer is that ABC is not right for you yet, and we would rather say so now than discover it together at month nine. This is also why the recommendation can step up a level when the most painful step costs more than ten hours a week: at that size, the sums tend to settle the matter quickly.
There is a small Stoic comfort in all this. You cannot control whether the machine is a good idea in general. You can control the inputs to one honest sum. Do the sum, look at the date, decide.
The best investments are rarely exciting. They simply stop costing you something every week, quietly, until one day you notice the evenings are yours again.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Plan, M0 to M4
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
A project without milestones is a long corridor with no doors. Everyone is walking, everyone is busy, and nobody can say how far along they are, except in phrases like "nearly there" and "the hard part is behind us". ABC replaces the corridor with five rooms, named M0 to M4 , and each room has a door you can see.
M0 comes first, in the first week: the repository, the checks and the landing page. Something real exists on the internet almost immediately. It is not much, but it is not a slide either, and that changes the tone of everything that follows. M1 brings sign-in and the data model, with your spreadsheet imported, so the system holds your truth rather than a sample. M2 is the core workflow and the master agent, the moment the machine stops being furniture and starts doing work.
M3 adds the back office and the reports. It is in the plan but missing from the diagram, which jumps from M2 to M4 for the same reason the plan does not dwell on it: it is the solid, quiet middle that makes the launch possible rather than a moment anyone will remember. M4 is launch, with the morning status running every weekday. That is the end of the build and the start of the machine.
Each milestone ends with an output you can point at : a live URL, a merged pull request or a file. Not a percentage, not a feeling, not a status called "in progress" that has been in progress for a month. If you cannot click it, open it or read it, it is not done, however nearly done it is. This rule is mildly irritating to builders and enormously reassuring to everyone else.
The default timeline for Core is eight weeks from kick-off. The date is fixed and the scope is flexible. If something has to give, a nice-to-have moves to V2 rather than the launch moving to spring. This sounds harsh until you have lived through the alternative, in which everything is kept and nothing is shipped, and the project quietly becomes a hobby.
Along the way you are never waiting for a progress meeting. The morning status tells you which milestone you are in, what moved yesterday and what is waiting for you, which is usually an approval and occasionally a question. You do not need to chase the plan. The plan comes to you at 08:45 with a short report and, on good days, a link.
By M4 you have something specific: a running machine, moving your one goal, reporting every weekday. Not a launch party, though you are welcome to have one. Just a system that does on Tuesday what it promised on the first Monday, which is the most quietly impressive thing software can manage.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What Could Go Wrong
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every proposition has a section on risks, and most of them are written by someone who would rather you skipped it. The risks listed are suspiciously gentle, like a job applicant whose greatest weakness is caring too much. ABC tries to do this one properly, because the risks are real and, handled early, quite manageable.
Projects like this rarely fail on technology. The code is the predictable part. They fail on people, data and enthusiasm, roughly in that order, and the diagram asks one question of each: was the risk seen early ? If it was, it is mitigated, quietly, inside the plan. If it was not, it has to be escalated, which is louder, slower and more expensive for everyone.
The first risk is adoption , and it usually looks like this: the owner does not have time to review. Approvals pile up, questions wait, and the machine sits ready but unblessed, like a new kitchen nobody cooks in. The mitigation is built into the product. Approvals are batched into a queue you can clear in about ten minutes, the morning status shows exactly what is waiting for you, and defaults are pre-filled so you only change what is wrong.
The second is data quality . The machine is only as good as the records it reads, and most small businesses keep their truth in four places that disagree. V1 brings the important records into one clean database, with a single import from the spreadsheet that holds the most truth. Where a figure is unknown, the scoreboard shows an honest empty box rather than a guess. Null beats invented, every time.
The third is over-automation : the temptation, once agents are working, to let them do everything. ABC holds a hard line here. Humans approve anything irreversible: sending money, publishing in public, deleting data, merging to production. No agent merges its own work. The budget guard caps the spending. The aim is a machine that does the dull work and asks before the consequential work, not one that surprises you on a Monday with something it was very proud of.
There is a fourth risk, rarer but bigger: building the right thing for the wrong market. So we agree a stop rule at the start. The default is that if there is no paying customer within three months of launch, we pause and rethink rather than pour in more money. Pausing is not failure. It is the cheapest lesson a project can buy.
Epictetus suggested rehearsing misfortune in advance so that it arrives as an expected guest rather than an intruder. A risk register is the business version, with fewer philosophers and more columns. List what could go wrong, decide early what you will do about each, and most of them will turn up politely, if at all, and be shown straight to their seat.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
How We Recommend
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ask most advisers which option you should choose and you will receive a beautifully balanced table. Three columns, each with strengths and weaknesses, and a closing line about how "the right choice depends on your priorities". This is very fair and almost entirely useless. You did not ask for a table. You asked for an opinion, from someone who knows more about this than you do.
ABC always names one option. Not because we are sure we are right, but because a recommendation is a service, and declining to give one is a way of handing the work back to you while invoicing for it. The proposal shows all three sizes, Lite, Core and Full, so you can trade up or down with your eyes open. Then it says, in a sentence, which one we would choose in your shoes, and why.
The why is the important part. We do not choose by instinct or by what would be most fun to build. We score four signals from your onboarding answers. Budget : what you said you could spend on the build. Timeline : when you need it live. Team size : how many people will use it. Data maturity : whether your records live in one sheet, a few tools or several systems with real volume. Each signal points at one option. The option with the most points wins, and ties go to the smaller one.
There is exactly one override. If the step that hurts most is costing you more than ten hours a week, we may recommend one level up, and we will say so in writing, with the sum. That is the only place judgement overrules the scoring, and it is always written down where you can check it.
The diagram draws this as a quadrant, with risk on one axis and value on the other. Low value and high risk is the corner nobody should be in. High value and high risk is tempting but premature. Low risk and low value is safe and pointless. The recommended option lives in the remaining corner: low risk, high value. In practice that is almost always the smallest option that genuinely moves your one number, which is why ties go to the smaller one.
This makes the recommendation slightly boring, which is exactly what you want. You can see how it was reached, you can disagree with a signal, and if you change an answer, the recommendation changes in front of you. Nothing is hidden in the adviser's gut, which is not, in fairness, an organ known for its record keeping.
You are, of course, free to choose differently. It is your business and your goal. But you will always know what we would do, and why, and that turns a choice from a guessing game into a decision you can explain to anyone, including yourself.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What You Own
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
There is an old arrangement in software that nobody admits to. The client pays for the system, but the builder keeps the keys. The code lives in the agency's account, the data sits in a platform you cannot export from, and the knowledge of how it all fits together lives in the head of a developer called Dave, who has just taken a job in Lisbon. You paid for it, but you do not, in any practical sense, own it.
ABC turns this round completely. Everything built is yours, in your own repositories , from the first day. The diagram puts "you own" at the hub, with four spokes: code, data, context and the book.
The code lives in a private repository under your account, not ours. Every change arrives as a pull request with a readable history, so you own not only the system but the story of how it got that way. Because the stack is boring and well documented, any competent developer can pick it up, including one you hire after we have gone.
The data lives in your database, with daily backups and a CSV export of everything. If you leave, you take it all with you in an afternoon. No exit fees, no hostage negotiations, no polite emails asking whether it might be possible to have your own customer list back.
The context is the written memory of your business: the goal, the decisions, the people, the rules and the skills the agents follow. It is perhaps the most valuable of the four, because it is the part that took longest to produce and would be hardest to rebuild. It lives in your repository too, in plain text, readable without any special software.
The book is the fourth spoke and the easiest to forget. The V1 book from onboarding, and every version after it, describes your business, your machine and your deal in plain words. It is yours to keep, to edit, to hand to a new hire or a potential buyer. When people ask what your business does and how, you will have a better answer than most companies ten times your size.
All of this is private by default. Nothing is published without a decision, and nothing of yours is used to train or sell anything else. Ownership is not a clause added at the end of the contract. It is how the machine is built, from the first commit. The Stoics were firm that very little is truly ours: our judgements, our choices, not much more. They had not, in fairness, met a well-run Git repository. Here at least, the line is easy to draw. If we built it for you, it is yours. Dave can go to Lisbon with our blessing.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Next Step
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Nine parts in, you know the shape of the deal. You start with onboarding. You get a V1 book about your own business, then a V2 and a V3 if the first one earns it. There are three ways in, a setup fee and a small monthly, a payback sum you can do on the back of an envelope, a plan from M0 to M4, an honest list of what could go wrong, one named recommendation and a clear answer about who owns what. The answer, in case you skimmed, is you.
That is a lot of reading. The next step is much shorter. It is five minutes of onboarding , and it costs nothing but the five minutes.
The sequence is deliberately dull. You arrive. ABC asks its questions, each one with a sensible default already filled in, so "ok" is a perfectly respectable answer. You finish. A short book comes back: your goal at the centre, your workflow, your agents, the option we would pick for you and every assumption marked so you know what to correct first. Then the next move is yours. That is the whole funnel, and at no point in it does anyone ring you.
People often expect a catch here. Surely the free part is a lure, and the real ask arrives on page two in a slightly smaller font. It does not. The book is useful whether or not you ever build anything. Plenty of owners read their V1, fix three assumptions, and discover that the most valuable thing in it was the single sentence at the top: the goal, written down plainly for the first time. If that is all you take away, the five minutes were well spent.
There is a reason the step is so small. Big first steps are where good intentions go to die. A discovery call needs a diary, a proposal needs a meeting, and a meeting needs a second meeting to discuss the first. Each of those is a polite argument with Tuesday, and Tuesday usually wins. Five minutes, on your own, at whatever hour suits you, is a step Tuesday cannot really object to.
The Stoics were fond of reminding themselves that the only time anyone acts is now. Seneca wrote that while we postpone, life speeds by. He was not selling software, but he would have recognised the whiteboard with twelve goals on it, and he would have had views.
So here is the whole of Part IX in a sentence: the deal is fair, the work is yours, and the first step fits in a coffee break. You do not need to decide anything about Lite, Core or Full today. You only need to answer some questions, most of which have already been answered for you. Then read your book, and make your call. The kettle will barely have boiled.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Machine Runs Overnight
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The most noticeable thing about an ABC business is that it keeps working after the owner has gone to bed. Not dramatically. There are no lights blinking in a basement. Somewhere, quietly, a set of agents is drafting, sorting, checking, building and tidying, and by morning there is a pile of finished work waiting to be looked at.
This is the night shift , and it changes the shape of a day more than any single feature. A small business used to have one working day, roughly the length of the owner's patience. Now it has two. The second one runs from about midnight to breakfast, and nobody needs to be awake for it.
What happens overnight is the dull middle of the work. Data gets reconciled. Reports are drafted. A new fleet app is built from its dataset and sits in a preview, waiting. Chapters of a book appear. Follow-up emails are written and placed, unsent, in the approvals queue. The machine is busy in the way a good kitchen porter is busy: steadily, without fuss, and without anyone needing to watch.
What does not happen overnight matters just as much. Nothing irreversible leaves the building while you sleep. No payment is made. Nothing is pushed to production. No message goes out to a customer. Those are written into the system as human gates, and the night shift treats them the way a sensible house guest treats a locked door: it notices, it does not try the handle, and it leaves a note.
That note is the morning status. At 08:45 on a weekday it arrives with the night's results, what got stuck, and a short list of things that need your yes. The night shift ends there, and the day decision begins. You read, you approve, you change a word, you say no to one thing. Then the work you approved goes out, in daylight, with your name on it.
There is an old worry about automation, that it will quietly do something terrible at three in the morning and you will find out from a customer. The design answers that worry with a simple rule rather than a clever one. Anything that cannot be undone waits for a human. Anything that can be undone is fair game. It is not a sophisticated rule. It does not need to be.
The real change is psychological. You stop going to bed with the day's leftovers in your head, because the leftovers have a place to go. You wake up to progress rather than to a list. The cycle runs night, morning, decision, and then night again, and over a few weeks it starts to feel less like software and more like weather. Reliable, mostly pleasant, and somebody else's job to arrange.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Prodgame
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most people treat a business year as a long walk uphill in the rain. It begins with resolutions, sags in February, rallies briefly in spring and ends in a December blur of things that were nearly done. It is grind, and grind has a well-known weakness: people stop.
ABC treats the year differently. It treats it as a game, and the game is called Prodgame . The year has levels. Each level is made of runs. Every run produces a score. Enough score, and you level up. That is the whole chain, and it sounds frivolous until you notice that it is the only way most humans have ever voluntarily done the same difficult thing a thousand times.
A run is a single, bounded attempt at something: build an app, ship a feature, write a book, close a week. It has a start, an end and an outcome you can check. Crucially, it is allowed to fail. A failed run is not a disaster. It is a run that scored nothing, which is information, and the next run starts with that information already in hand. Grind treats failure as a moral problem. Games treat it as Tuesday.
The score is not a feeling. It comes from the same place as everything else in ABC: real evidence. An app that went live, a pound that arrived, a pull request that merged, a chapter that was written. The machine counts these, and the morning status reports them. Nobody gets points for being busy, which is a relief to everybody who has ever been rewarded for looking busy and found it oddly hollow.
Levels give the year a shape. Instead of one enormous goal sitting at the far end of December, there is a sequence of smaller thresholds, each one visible, each one reachable in weeks rather than months. You always know which level you are on and roughly how many runs it will take to clear it. The one goal is still the one goal. Prodgame simply cuts the road to it into stretches you can see the end of.
There is a deeper reason this works, and it is not about fun, although fun helps. Games make progress visible and failure cheap. Those are precisely the two conditions under which people keep going. Remove either one and you get either despair or recklessness. Keep both and you get persistence, which is the quality most businesses actually die for lack of.
Marcus Aurelius told himself to do each task as if it were the last of his life, without fuss or pretence. A run is that idea with a scoreboard attached. One thing, done properly, counted honestly, then the next. It turns out that a year of runs adds up to rather more than a year of resolutions, and it is considerably more cheerful to play.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Book a Day
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Writing a book used to be the kind of project people mentioned at dinner parties for a decade. It required a sabbatical, a shed, or both. ABC takes a less romantic view. A book is an output, like an app or a report, and the same machine that ships one can ship the other.
The process is called Book a Day , and it lives in the machine as a single skill, abc-bad. You give it a topic. It returns a complete illustrated non-fiction book: ten parts, ten chapters each, a hundred chapters in all, one editorial diagram per chapter, in the funny, stoic voice you have presumably noticed by now. It arrives as a folder of markdown and a print-ready A5 PDF. You may have guessed where this is going. You are holding one.
The diagram in the middle of the hub has four spokes, and each one is a stage. First the outline : a hundred titles, each with a one-line synopsis, arranged so the book argues something rather than merely lists things. Then the plates: one diagram per chapter, specified as a shape and a handful of labels, so every chapter has a picture of its idea before it has a single sentence. Then the prose, written from the diagram rather than the other way round, which keeps the words honest. Finally the PDF, typeset and ready for a printer or an inbox.
Writing from the diagram is the quiet trick. A chapter that has to match a picture cannot wander off. If the plate shows three boxes in a chain, the prose must explain three things and how one leads to the next. That constraint does more for clarity than any style guide, and it means the illustrations are never decoration bolted on afterwards.
Why would a business want a book a day? Partly because books are the most persuasive format ever invented for an idea that needs more than a slogan. Every ABC client receives one about their own business, the V1, and the V2 grows to a hundred chapters. Partly because a topic turned into a book can be turned into many other things: an app, a talk, a season of short clips. One dataset, many outputs.
The human part does not vanish. Someone chooses the topic, and choosing is the hard bit. Someone reads the draft, corrects the claims, cuts the jokes that did not land and approves it before it goes anywhere. The machine is fast and tireless. It is not, on its own, a good judge of whether a book should exist.
There is something faintly comic about a machine that writes a book a day being used to write a book about the machine. It is also the most honest demonstration available. If you have read this far, the factory works. Whether the book was any good is, as ever, your call.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
From Client to Partner
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
The ABC sales spine has seven gates, and you have met most of them already. Lead, Prospect, Workspace, Demo, SaaS, Customer, Partner. Each gate is a stage a relationship passes through, and each has a test for moving on. This chapter is about the last two gates, and in particular about the last one, which is the one most businesses forget exists.
A client is someone you are building for. They have onboarded, read their book, chosen an option and paid for a build. A customer is a client whose machine is running: the morning status arrives, the agents are doing real work, and the retainer quietly pays for its own upkeep. Most sales processes stop here, with a happy customer, and declare victory.
ABC has one more gate. A partner is a customer who runs their own machine so well that they start sending the next one. Not because they were offered a commission, although that conversation may come, but because they have a working system and a neighbour who keeps saying, at eleven at night, "I just need to know where everything is." The partner hands over the link. The neighbour does five minutes of onboarding. The spine starts again.
This matters for a simple reason. The cheapest, most trusted sales channel a small business has ever had is a satisfied owner telling another owner what happened. No advert can match it. A partner is not a salesperson. They are evidence, walking about, with a morning status on their phone they are happy to show you.
Getting there is less about selling and more about handing over properly. A customer becomes a partner when they understand their own machine: what it does overnight, what waits for their yes, where the code lives, how to read the evidence. Everything built is theirs, in their repositories. You cannot recommend a thing with confidence if you suspect it would collapse the moment the builder stopped answering emails.
There is a nice symmetry here. The spine begins with a stranger who knows nothing about ABC and ends with someone who could explain it better than the brochure. The middle gates are where the work happens, but the final gate is where the work proves it was worth doing. A machine that only ever produces customers is a business. A machine that produces partners is something closer to a movement, although nobody involved would use a word that grand.
So the measure of a good engagement is not the invoice, or even the renewal. It is whether, six months later, a new lead arrives with a short note that says someone told them to get in touch. That note is worth more than a five-star review, and it is considerably harder to fake.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
A Hundred Small Companies
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Here is an idea that would have sounded ridiculous ten years ago: one person, running not one small business but many, each modest, each mostly looking after itself, all of them sharing a single machine. ABC calls this the portfolio life , and the arithmetic behind it is gentler than it sounds.
The diagram is a funnel. At the top, many small ventures: apps, books, a client practice, a niche newsletter, a booking site for something nobody else thought to make a booking site for. In the middle, one shared machine: the same back office, the same context, the same agents, the same skills. At the bottom, one calm owner, who looks at one morning status rather than a dozen dashboards.
The fleet apps are the clearest example. Each one is built quickly from a single dataset and set a modest target: about £1 a day . That number is not a typing error. One pound a day is not a business. It is a signal. An app that clears it has found an audience and earns its hosting with change to spare. An app that does not is held, fixed or quietly retired at a weekly review. Enough of the ones that clear it, and the pounds stop looking small.
The reason this works for one person, when it would once have needed a holding company and an accountant with a nervous twitch, is the shared machine. The tenth venture does not need a tenth back office. It needs a new row in the existing one. It does not need its own memory. It reads the same context, with a section of its own. The build, the deployment, the monitoring and the status are already there. Each new company costs a little attention to start and almost none to keep.
That last point is the whole game. Most people who try to run several things at once fail not for lack of ideas but for lack of attention. Every venture leaks a little of it, every day, in small checks and small worries. A shared machine plugs most of the leaks. What remains is a short list of decisions, gathered into one place, at one time of day.
This is not a recipe for frantic empire building. A hundred small companies is a figure of speech, not a target, and the one goal still sits above all of them. Every venture in the portfolio exists because it moves that goal or teaches something useful about how to. The ones that do neither are dropped, kindly, with a note.
Epicurus thought a good life needed surprisingly little: friends, freedom and time to think. He would have been suspicious of a hundred companies. He might have been reassured, though, by an owner who runs them all and still has the afternoon free for a long lunch with friends.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
What Stays Human
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
After ninety-five chapters about what a machine can do, it is worth being precise about what it cannot, or rather what ABC is designed never to take. The list is short. It is also the most important list in the book, because everything else is arranged around it.
The diagram is a Venn with two circles. One is taste : the sense of what is good, what fits, what is worth making at all. The other is judgement: the sense of what is right, what is wise, what to do when the evidence points two ways at once. Where they overlap is you. Not a role, not a seat on an org chart. You.
Taste is the thing agents imitate most convincingly and possess least. A machine can produce a hundred versions of a landing page, a hundred chapter titles, a hundred names for an app. It cannot tell you which one makes you slightly proud. It can rank them by a rule, but someone wrote the rule, and someone has to notice when the rule produces something technically correct and faintly embarrassing. That noticing is taste, and it does not automate.
Judgement is harder still. The machine is very good when the question has an answer. It is much less good when the question has a trade-off: the client who is late paying but sent you three others, the app that earns well but makes you uneasy, the opportunity that moves the goal and costs a friendship. Those decisions need someone who will live with the consequences. A system that pretended to make them for you would be lying about who carries the risk.
Then there are relationships, which the machine can support but never replace. It can remind you that a customer has gone quiet, draft the note and find the history. It should not pretend to be you. People can tell, and they mind. The approvals queue exists partly so that every message going out under your name has passed through your hands.
And finally there is the final yes . Every irreversible step in ABC, a payment, a release to production, a message out of the building, waits for a person. That is not a safety catch reluctantly bolted on. It is the design. The machine does the preparing and the checking and the drafting so thoroughly that your yes takes seconds. But the seconds are yours, and they always will be.
There is a quiet comfort in this. Owners sometimes worry that the more they automate, the less the business is theirs. The opposite turns out to be true. Strip away the admin and what remains is almost pure taste and judgement, the parts that made you start the business in the first place. The machine does not hollow out your role. It finally gives you time to do it.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Unbothered Operator
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Epictetus began his handbook with a sentence that has outlived empires: some things are within our power, and some are not. He had been a slave, and he had a bad leg, and he was not given to decoration. If he had run a small business, he would have pinned that sentence above the desk and then got on with the invoices.
This chapter borrows his stance for a company of one, and gives it a name: the Unbothered Operator . It is not indifference. An unbothered operator cares a great deal. They simply refuse to spend their care on things it cannot change.
The diagram is a split with a single question at the top: is it in your control? If yes, decide it. Make the call, write it down, move on. If no, let it go. The algorithm, the weather, the competitor's launch, the customer who will not reply, the economy, the mood of whoever reads your proposal: all of these are real, and none of them respond to worry.
ABC adds a third path the Stoics did not have, because they did not have agents. Some things are in your control but not worth your attention. Reconciling the bank feed. Chasing the second reminder. Rebuilding the report every Friday. These are not for deciding or for letting go. They are for automating . The machine takes them, does them properly, and puts the result in the morning status. The operator's day then contains only the first kind of thing: decisions that genuinely need a person.
This is what makes the stance practical rather than merely admirable. It is easy to tell an exhausted owner to accept what they cannot change. It is unkind to say so while they are drowning in things they could change but have no time for. Clear those away and the Stoic advice becomes surprisingly easy to follow. There is room, at last, to notice which worries are useful and which are just noise.
Marcus Aurelius, who had an empire to run and a plague to manage, kept reminding himself that he could wipe away a disturbing impression as easily as a smudge. Most operators will not be managing a plague. They will be managing a late payment and a broken checkout. The method is the same: see it clearly, ask the question, take the path. Then put the kettle on.
The goal of all this is not serenity for its own sake. It is better work. An operator who is not rattled decides faster and more fairly, notices more, and is pleasanter company on a Thursday. Being unbothered is not the absence of trouble. It is knowing, quickly and calmly, which troubles are yours.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
Compounding Calm
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Every automated chore buys back a little attention. Ten minutes here, a stray worry there, the half hour you used to lose finding a file. On any single day the saving looks trivial, which is exactly why most people never bother. This chapter argues that they have the maths wrong, and that attention behaves less like small change and more like money in a savings account.
The idea is compounding calm . The diagram is a simple quadrant: years along the bottom, calm up the side, and a curve that starts almost flat and then bends upward in a way that looks suspicious until you have seen it happen.
Here is how it works. Automate one chore and you get back a sliver of attention. Spend that sliver well, on the next chore worth automating, and you get back a slightly larger sliver. Do it again. Each round of freed attention pays for the next round of automation, and the next round frees a little more. In the first months it feels like nothing much. By the end of the year the owner has noticeably fewer things in their head. By the third year they have the kind of calm that colleagues mistake for a personality trait.
Compounding also works the other way, which is the part people tend to miss. Every unautomated chore is a small loan taken out against tomorrow. You do it by hand today, and again tomorrow, and the interest is the faint dread of knowing you will do it again next week. Over years, that dread compounds just as reliably as calm does. Businesses do not usually fail in a single crisis. They fail slowly, under a pile of small things nobody had time to fix.
What makes ABC suited to compounding is that its improvements stay improved. A chore handed to an agent is not handed back next month. The skill that does it lives in a repository, is versioned, gets better from its own logs and survives your holidays. Calm built on habit erodes when you are tired. Calm built on a machine waits patiently for you to come back.
The Stoics practised a daily review, looking back each evening at what had gone well and what had not. Seneca recommended it warmly. ABC's weekly review is the business version, and it is where the compounding is steered. Each Friday asks which chore cost the most attention this week. That chore becomes a candidate for the machine, and the curve bends a little further.
So the right question about automation is never whether it saves time today. Usually it saves very little today. The question is what the next three years look like with that saving reinvested, every week, without fail. The answer, for most owners, is quieter, which turns out to be the most expensive thing they never knew they could buy.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
The Machine Improves Itself
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Most software is as good on the day it is installed as it will ever be. After that it ages, slowly, like a sofa. Somebody files a complaint, somebody else writes a workaround, and five years later nobody remembers why the third button exists. ABC is built on a different assumption: that a system which does real work every day should get a little better every week, and should do most of the improving itself.
The diagram is a cycle with three stops. Run , Measure , Improve skill . Then run again. It looks simple because it is. The cleverness is not in the loop. It is in actually closing it, every time, instead of letting the measurements pile up in a folder nobody opens.
A run is any piece of work the machine does: building an app, drafting a status, writing a chapter, reconciling a feed. Each run leaves a trail. Usage tracking records what was used, how long it took and what it cost. Evals check the output against a written standard, the way a teacher marks against a rubric rather than a mood. Did the status link every claim to evidence? Did the chapter match its diagram? Did the app pass its checks? The answers are recorded, not remembered.
Then comes the weekly review, and this is where the loop closes. Friday looks at the week's runs and asks a plain question: which skill let us down, and what would fix it? A skill, remember, is a written playbook that an agent follows. When it falls short, the playbook is edited. A rule is added, an example is improved, a step is clarified. The change is made as a reviewed pull request in one repository, skills-2026 , which is the single source of truth for every skill on every surface the machine uses.
That one repository matters more than it sounds. An improvement made once is picked up everywhere. Fix the way the status handles a quiet day, and every client's status handles quiet days better on Monday. Fix the way a book chapter lands its last paragraph, and every future book benefits. The machine does not merely learn. It learns in a place where the lesson cannot be lost.
None of this is autonomous in the alarming sense. The machine proposes improvements. A human reviews them and merges them. The cycle is fast, but it is not unsupervised, and the skills only ever change through the same front door as the code. That is what keeps the improvements honest and the history readable.
Epictetus said that progress belongs to the person who studies their own conduct rather than other people's. He meant people, but it suits machines remarkably well. Run, measure, improve. It is not glamorous. It is simply how things that last tend to get that way.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.
One Goal, One Machine
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
It started, if you remember, with a whiteboard. Twelve goals, written in January in a confident hand; by spring eleven of them had become part of the wallpaper. Nobody abandoned them. They simply lost, one by one, to the steady pressure of ordinary weeks. A hundred chapters later, it is time to rub the board clean and write one line in the middle.
Here is the whole book, stated plainly. Choose one goal, build one machine to move it, and let that machine carry the rest. Everything else has been detail, and most of it was there to make those three short sentences believable.
The diagram is a stack, and it reads from the top. One goal: a number with a date, written where every agent can read it. Beneath it, one machine: not a pile of tools but a single system with one memory, one back office and one voice. Beneath that, the three pillars that hold the machine up, Apps, Backoffice and Context, each earning its place by moving the goal. And at the bottom, holding the whole thing honest, daily evidence : the 08:45 status, every weekday, with a link under every claim.
The bottom layer is the one people underestimate. Goals are easy to write and machines are fun to build. Evidence is the dull part, and it is the only part that tells you whether the rest is working. A goal without evidence is a hope. A machine without evidence is a hobby. With it, both become something you can steer, a little, every morning, before the day has a chance to argue.
What the machine does is considerable. It works overnight, plays the year as a game of runs, writes books, builds and retires apps, carries clients along the spine until some of them become partners, and quietly improves its own skills each week. What it does not do is choose. The goal is yours. So are taste, judgement, relationships and the final yes. The machine was never meant to replace the person at the centre of the business. It was built to clear enough space around them that they could finally see the goal clearly.
There is a line, usually credited to the Stoics, that we suffer more in imagination than in reality. Most of the suffering in a small business is imaginary in exactly that sense: the dread of things forgotten, the fog of not knowing where everything is. A machine that remembers, reports and handles the dull middle removes a remarkable amount of it. What is left is real work, which turns out to be much lighter than the worry ever was.
So, the whiteboard. Wipe off the eleven. Thank them, if you like; they were good ideas, and they are safe in the parking file. Write the one that remains, with its number and its date. Then build the machine, starting with five minutes of onboarding, and let it run. Tomorrow at 08:45 it will tell you how far you have to go. The distance may be long. That is fine. For the first time all year, you will know exactly how long it is, and you will not be walking it alone.
One Goal, One Machine — The ABC proposition: apps, a back office and context, run by agents
Ready when you are.
Tap below to open the full chapter in the book reader. Anchored books jump to this chapter automatically.