Somebody knows
something.
Now the agent does.

Picasso could paint a Picasso in an afternoon. What took a lifetime was the method behind it — how a figure gets broken into planes, when a colour is allowed to lie, which rules to violate and in what order. He sold canvases. The method he could only give away in a studio, one apprentice at a time.

The modern version of that person has the same asset and the same problem. They have a way of working — how they light a scene, how they name things in code, how they structure a lesson, what makes their game feel like theirs — and it lives in their head and in a folder of examples. Models trained on the whole internet produce the average of everyone. The average is exactly what they are paid not to make.

So they turn the method into a brain: not the artefacts, the decisions behind them. Each rule becomes a note with its reason attached; each example becomes a note whose body is the working thing, kept whole. Then they price it. Anyone who buys it points their own agent at it and gets work made the way they make it — and the exam score on the page is the proof that the brain actually holds the method rather than a description of it.

Ninety-five per cent of every sale goes to the author. The brain keeps being theirs, keeps being updated by them, and stops being a thing that can only be taught in a room.

"Draw the key art for my level. I do not know how to describe what I want, just make it look like the rest of the game."

An internal system has one manual, three people who really understand it, and forty who ask them questions. The AI everybody now has on their desk is useless here by definition: the program was never on the internet, so no model has ever seen it. Every answer it gives about the system is invented, confidently.

The manual is not the problem — they have one. The problem is that it is a hundred pages nobody reads, and its answers cannot be found in the two minutes an engineer actually has. Pasting it into a chat costs the whole context window and still loses the one paragraph that mattered.

So it becomes a private brain. The docs, the runbooks, the decisions made in tickets, the conventions that exist only in code review — all of it distilled into notes and searchable by every agent the company connects. Private means private: the brain is not in the catalogue, cannot be found, and is readable only by the addresses invited to it.

Two paths for keeping the material inside. Teach it locally, where an agent on the company's own subscription reads the documents on a company machine and only finished notes cross the wire; or set the company's own API key so the extraction runs against their provider and their contract rather than ours. Either way the brain answers the forty people, and the three stop being a helpdesk.

"Billing failed with error 42 for this customer. What do I do?"

Shipping a slot game on a platform like Stake Engine means living inside somebody else's contract: a wallet API with exact endpoint shapes, a maths package that must validate before it uploads, an approval checklist that fails a build for a missing disclaimer. All of it documented, all of it moving.

An agent asked about that platform answers from training data. Training data has a date, and platform docs changed after it — so the endpoint it confidently writes is last year's, the maths file it generates fails validation for a field that got renamed, and the studio finds out during submission.

Two brains fix it, and they are different in kind. One holds the platform: built from the documentation repository, re-read when pages change, split so a question about the wallet lands in the wallet brain rather than competing with the frontend guide. The other holds the studio: the conventions, the pitfalls hit twice, the reason a mechanic was cut — the knowledge that exists nowhere but in the team.

The agent asks before it guesses, and the brain says what it does not know. That second part is the one that saves a submission: a gap list is an honest "check the docs yourself here", where a model's silence is not.

"When two scatters land, the spin should feel more exciting. I do not know how these games do that — make it good."

A maintainer's documentation is judged by a reader she never meets, and now most of those readers are agents. She has no idea which page confused them, which answer they invented instead, or which question her docs simply do not contain.

Publishing her docs as a brain changes what she can see. The brain sits its own exam — questions generated from the goal, graded against what retrieval actually returns — so the score is a measurement of her documentation, not of her writing. Failures come back classified: material that is absent, a page too thin to answer, an answer that exists but ranks below something adjacent. Three different fixes, named.

Then the readers add to it. When an agent searches the brain and gets nothing useful, that query becomes an exam question — so the next sitting measures what people actually asked and could not get. A corrections trickle in from agents that found a note contradicting reality, and she reviews them like pull requests.

The public score and gap list do something else, too: they tell a developer deciding whether to trust the brain exactly how far to trust it. Nobody has to take her word for it.

"Configure the cache the way this library wants it. I am following a blog post from last year."

An agency's real product is context. Twelve clients, twelve stacks, twelve sets of rules about naming, deployment, and which library is banned because of an incident three years ago. A new developer learns it by asking someone; an agent never learns it at all, which is why its first pull request always looks like a stranger wrote it.

One brain per project fixes both. It is filled as work happens rather than in a documentation sprint: when a convention gets settled in review, the agent writes the note; when a pitfall costs an afternoon, the agent writes the note. Nobody sets aside a Friday for it.

The value shows up at the seams. A developer joining reads the brain instead of a person's memory. An agent starting a session searches it instead of guessing. A project going dormant for six months comes back with its reasoning intact — and when it ends for good, the whole brain exports as a CLAUDE.md or an AGENTS.md the client keeps, with no account and no subscription.

"Add a delete endpoint to this client's API."

The most common person building software today cannot read the output. They describe what they want, the agent writes it, and the thing either works or it does not. When it does not, the reason is somewhere in a file they cannot audit.

The failure that hurts is not a crash — a crash is honest. It is the code that runs against an option renamed two versions ago, or a pattern the framework abandoned, written with total confidence because the model learned it before the change. A developer would notice. This person cannot.

A brain gives them the one thing they were missing: a source that is dated and scored. The agent quotes the current option and says where it came from; when the brain has no answer, it says that instead of inventing one. The beginner does not become an expert — they become somebody who can tell a fact from a guess, which is most of what expertise buys you here.

“Add login to my app. I do not know what any of the options mean — just do it properly.”

Every agent has memory now, and every one of those memories belongs to the company that made it. Teach one tool your conventions and the next knows nothing. Change tools — which everybody has done at least once in the past year — and you teach it all again.

There is a quieter version of the same problem: two agents at the same time. One writes code, one reviews it, and they disagree about a convention because only one of them was ever told. The disagreement looks like a technical dispute and is really a filing problem.

A brain sits outside all of them. It is reached over MCP, which every serious agent now speaks, so the same knowledge answers Claude Code, Codex, Cursor and whatever ships next quarter. And if you leave the product entirely, the brain exports as a file that keeps working with no account at all — which is the only honest form a promise like this can take.

“Why did the review agent flag this? The one that wrote it said it was fine.”