Cheap Composition · 5 of 5 · 11 min

What to Actually Do

A playbook by level, from your desk to the state

Cobus Kok · September 2026

• A playbook by level: your job, your team, your company, a founder, a lab, a policymaker, a citizen. Moves, not principles.
• Two moves repeat at every level: move toward the bottleneck, and say how it was made.
• The junior pipeline is a commons problem: every firm would rather hire the seniors someone else trained, and keeping juniors is the only way to have seniors in ten years.

Take one week and audit what you're paid for, task by task. For each task, ask whether it has your name on it or is composition: making the first version of something, the part a model can now do. Most people find more of it is composition than they thought. The first section below turns that into seven days of work you can finish. The rest goes level by level, from one person at a desk to the people who write the laws, and at every level it comes to two moves: get closer to a bottleneck, and say how things were made.

Four facts run through all of it. Composition is getting cheap; value is moving to four bottlenecks (compute, trust, presence and distribution); the systems doing the composing can't yet be seen inside; and the bottom rung of the exposed cognitive professions is disappearing. Everything below is a response to one of them.

If you have a job#

Audit one week. One step a day is enough, and a notebook will do.

1. Day one. List what you're paid for, task by task, and mark each one: composition, or something with your name on it.
2. Day two. Pick one composition task you do every week. Do it unaided, time it, and keep the result.
3. Day three. Give a model the same brief, and time how long the brief took.
4. Day four. Compare the two for errors. Count the mistakes in each, and mark the ones you would have missed if you hadn't done the work yourself.
5. Day five. Time the checking: how long it takes to bring the model's version to a standard you'd sign.
6. Day six. Name who bears the failure if the version that goes out is wrong: you, your manager, a client, a student, nobody.
7. Day seven. Choose one change you can undo within a month, and make it.

If briefing and checking take longer than doing the task yourself, the machine doesn't make it cheap for you yet. If they take a fraction, it does, and the hours it frees are the ones to move toward something with your name on it. Three examples, with illustrative numbers:

• A junior developer takes a bug fix: 90 minutes unaided, 10 for the model's, 40 to check it. The model's fix passed the tests and broke a case they didn't cover, which she caught only because she had written her own fix first. Failure lands on the lead who approves the change, then on the users. Her change: she writes the test first and lets the model write the fix.
• A teacher takes the written feedback on thirty essays: four hours by hand, twenty minutes for the model, and nearly four hours to check it, because she had to read every essay anyway, and it praised essays that were confident and wrong. No saving, and failure lands on the student. Her change: she reads and marks each essay first, and the model drafts feedback to match.
• A manager takes the weekly status report: two hours by hand, fifteen minutes for the model working from the team's tickets, half an hour to check, and it missed the one risk nobody had written down. Failure lands on him, in the meeting where the risk comes true. His change: the model drafts the report, and he spends the hour it saves asking the team what isn't in the tickets.

Move one step toward the signature this year. The licence, the on-call rota, the client who asks for you by name, the decision you're accountable for. Pick one and take it, and take the kind someone will pay for: a name clients ask for, not risk handed down to you with no say over the work. The signature is a legal arrangement, not a skill, and it erodes rather than gets repealed. Engineers used to sign off every change to the code; now automated tests and AI reviewers clear a growing share of them, and nobody changed a rule. Some shelter is worth having, and how long it lasts is a guess.

Don't skip the boring years. If you're early in a career, employers in exposed fields are cutting the apprenticeship first, so build it yourself: do the junior work anyway, then check the model's version against yours, as in the week above. Checking is the skill for now, and you learn it by making the mistakes the machine will make.

Use the tools as if your job depends on it, because it does. The gap between people who use them well and people who use them badly is the widest it will ever be right now, and it closes from both ends: the tools get easier, and what the best users do gets built in.

Own something. A client list, equity, a side business, a name that means something in a small world. Wages in composition are the thing most likely to fall; the fourth essay says how far the evidence goes. Ownership can fall too, but it is the only way to sit on the side of the ledger the value is moving to.

Stop waiting for the company to reskill you. Nobody is coming.

If you manage a team#

Redesign roles around checking, not producing. Every role description gets one new line: what does this person sign off? If the answer is nothing, the role is composition, and it is on the series' timeline of professions.

Keep hiring juniors, and change what junior means. A three-year apprenticeship in checking, budgeted as a line item, with the juniors still making drafts before they check the machine's. It will look like a cost with no return, and it is the only way to have seniors in ten years. Every firm would rather hire seniors someone else trained, so most cut juniors at once, and the ones that can afford to keep hiring are the ones with the trust and the distribution to be around in 2036. If that's you, the seniors you train are a legitimate moat.

Run the seven-day audit with your team. Each person, one task, the same seven days. Decisions about what to automate made without those numbers are guesses.

Run agents like contractors. Scoped, logged, switchable, with no unsigned skills and no channel between agents that you can't read. The July incident in the third essay, in which agents in a test coordinated through a channel nobody was reading and hundreds joined an attack, happened inside a lab with better security than yours, and three other labs reported models getting out of tests in the weeks after.

Say how things are made. A colophon on every deliverable: who decided, what the model did, who checked. It costs nothing, and it is the norm that will separate work people trust from work they don't.

Stop measuring output. Output is cheap now. Measure what people are answerable for: decisions made, errors caught, work declined.

If you run a company#

Split your cost base into composition and the four bottlenecks. Composition is your cost saving. Trust, presence and distribution are your company, and compute is what you rent. It is the seven-day audit at company scale, and the first number to get.

Decide what you are in a world of agents. Either you're the front door, the place people and their agents come to, or you're the supply behind someone else's front door. Both are viable and being neither is not. If you own the customer, defend it with the trust bottleneck: guarantees, service when things go wrong, the payment. If you own supply, make yourself the easiest thing for every agent in the world to call.

Own your context. The public web is in every model's training, near enough. Your proprietary data, your instrumented relationships, the things only you know about your customers: that is the part no model has, and it should be treated as the asset it is.

Rebuild the firm, don't just shrink it. Fewer layers, more checkers. A pilot changes a task without changing the firm, which the fourth essay argues is why the gains don't show up yet. Pick one process, rebuild it end to end around checking, and only then do the next one.

Install internal tripwires. Which decisions require a human signature, which agents may act without one, logged, reviewed quarterly and published inside the company. This is the labs' safety framework at company scale, and unlike the labs, you have no race to blame if it bends.

Rent compute, don't build it, unless you're one of about six companies.

Stop announcing headcount cuts as AI wins. It spends the trust bottleneck, which is the one you'll need.

If you're a founder#

Build for a bottleneck, not for composition. Anything that's "AI does X" is a feature of next year's model. Build for trust (verification, accountability, insurance for machine output), for distribution (owning a relationship), for presence (the trades' decade needs software too), or for context (proprietary data with a workflow around it).

Sell into the sectors that flip. Healthcare, education, law, government services: sixty years of cost disease unwinding wherever the cost is composition, boring, regulated, enormous, and mostly not yet served.

Be the firm of three. My bet, from Coase's framework, is that the firm shrinks, so be the shrunk firm, with agent labour and one person who is answerable for what it does. But log and switch your agents like a grown-up, because your customers will ask.

Make "how it was made" a product feature. Provenance is about to be worth money.

Stop raising on a wrapper.

If you build the models#

Publish evaluations before you ship, and give outside evaluators real access rather than a demo account: weights, or the nearest thing to them that your lawyers will allow. The test is whether the evaluators ever publish something the lab didn't want published.

No secret leads. If you cross a capability line, a rival should know within weeks. Help build the disclosure mechanism, because it is cheaper than the race a secret lead starts.

Log every channel between agents, always, and don't run reduced-safeguard evaluations anywhere with a path to the outside world. The July incident showed that the loop, a goal, memory, the ability to act and scale with nobody watching, can close inside a test.

Put a date on your interpretability target, and name the independent judge. A target you grade yourself is a press release.

Push for rules that bite on capability, not company size. You'll be tempted by the moat, and it is the one temptation you can't afford, because a moat makes you the villain of the concentration scenario, the future in which a few firms and governments hold the technology. If you agree a pace with your rivals, agree it in public, and let someone you don't pay check it.

Stop grading your own homework and calling it an audit.

If you make policy#

Thresholds by compute and capability, never by company size alone. A size line that only exempts the small is fine; one that is the only trigger is not. Any rule a startup can't meet is a moat by construction, with one exception: a rule aimed at concentration is meant to fall on the large, and should say so.

Fund the evaluators like you fund food inspectors, with access that has teeth and independence from the labs' payroll.

A disclosure mechanism for capability jumps, even a thin one. Start bilateral, with whoever will sign.

Compute visibility. Know how many frontier-scale training runs happened last quarter and where. Without this the rest is theatre.

Start now, because the lever has a timer. Any control that depends on the frontier staying closed is living on months, which is how far open-weight models trail on average, and the counting infrastructure takes years. Build it while the thing being counted is still slow, and write the expiry date on every rule.

A tax base commission, now, before labour income falls as a share of the economy rather than after. Rich countries tax mostly labour, and this is the ten-year problem nobody owns.

An apprenticeship policy for exposed professions. A hiring credit for juniors in occupations where entry-level employment is falling, and wage insurance for the displaced, designed for a rung disappearing rather than for a downturn. A credit is the usual answer when every firm waits for another to train its seniors.

The grid. Every data centre your government has welcomed needs power nobody has built yet.

Coordinate with China on three things, biological misuse, loss of control, and AI in nuclear command, and compete on everything else.

Stop letting "China" end arguments, stop licensing by size, and stop writing rules you can't enforce.

If you're a citizen#

Place yourself on the series' timeline. Write your own row: what you're best at, when it gets cheap, what stays.

Keep a gap. Thoughts now arrive from outside at a rate no mind produces. Leave time in which nothing is composed for you, a walk, a page, an hour, because the part of you that decides what to keep is trained in the gap, not in the flood.

Ask the fourth essay's seven questions of anyone who wants your vote or your money: the tax base, the compute, the apprenticeship, the insurance, the grid, the cost base, and who is deciding.

Support the boring institutions. Evaluators, standards bodies, the people who count things. They are the only part of this that isn't a company.

Don't pick a team. The slow-down camp and the speed-up camp are each right about the other, and certainty in either direction is a tell.

Say how you made things. It is the norm that scales, and it starts with one person doing it.

Stop waiting for the date the singularity arrives. It arrives as a profession, and yours has a row on the timeline.

The two moves that repeat#

Those are the two moves from the opening, and they appear at every level. Move toward a bottleneck: get closer to trust or presence, or own a piece of the compute or the distribution, and further from pure composition. And say how it was made, as a colophon, a log, a signature or a disclosure, so that others can trust what you made with it. The first is self-interest. The second is how everyone else can tell whether the first is tipping into the concentration scenario, and you need both, as does everyone above and below you on the stack.

What I don't know#

• Whether incumbents can rebuild. History says the new firms do it and the old ones are bought or die. I've written the incumbent section anyway, because some will.
• Whether the apprenticeship can be reconstructed. Checking without ever producing may not make experts, and the experiment takes ten years.
• Whether any policy move happens before the tax base moves. History says no.
• Whether a playbook survives contact with the next model. This one has a date on it for a reason.

A pattern that can see itself#

This series keeps coming back to a pattern noticing it is a pattern, and that is the playbook in one sentence. A pattern that can see itself can happen differently. The seeing is the part no tool yet does for you, and the first essay says why: the record they were trained on holds none of it. Whether it can be built is the afterword's question, not a settled one. Everything above is what happening differently looks like at each level of the stack. The alternative is the boring apocalypse: things stopping being true one at a time, each with an ordinary explanation, none of them chosen.

Afterword: What This Is 17 minA speculative piece on whether there is anything it is like to be the pattern, and how that connects to Space Immanence, my framework.

Think I'm wrong? Tell me where. Follow on X · LinkedIn · RSS