You ask the agent for a new endpoint. It writes one. The tests pass, the types line up, the handler does what you described and nothing more. Then the review comes back. There is no rate limit. The errors come out in a shape your clients cannot parse. The migration it wrote takes a lock on a table that must not be locked before ten in the morning. Nothing it did contradicts anything in the repository. All three rules are real, and all three live in the head of the engineer who has been there since 2023. The task was never the difficult part.

The chart is accurate and still incomplete

Start with what a general model actually holds. It knows what an endpoint is, what a rate limit does, why locking a busy table during business hours is unwise in the abstract. That is the part of your world that made it into the weights, and the usual picture for it is an iceberg, where the interesting half is the submerged one. Underneath sits your current operating state, your local vocabulary — what your company means by “account”, which is not what the word means down the road — your quality bar, your procedures, the history of every exception someone granted, and who is allowed to approve what.

Think of bringing a large vessel into a harbour. The chart is public and accurate about the depths as of the last survey. The pilot who climbs aboard to take her in knows the channel silted over the winter, that the ebb runs hard two hours before low water, and which berth the crane cannot reach with a full load. None of that is on the chart. Nobody wrote it down, because everyone who works that harbour already knows it, and because it would have gone stale before the ink dried. The chart is not wrong. It is simply not the whole of what you need to come alongside without touching anything.

What you cannot state you cannot delegate

Every organisation carries a set of requirements that are not features: reliability, security, compatibility, maintainability, performance, operability, risk, and the quality your customers can feel. What makes them local is not the list, which is roughly the same everywhere. It is how you decided to satisfy them and how you decided to trade them off against each other when they pull in different directions. Those decisions were made in meetings, and they mostly stayed there.

Written into a document nobody opens, they change nothing. A model cannot read your intranet, and neither, on a busy Tuesday, can your colleague. The move that works is to turn each one into something the work runs into. If every public endpoint must be rate limited, that is a check that fails the build, one conforming endpoint sitting in the repository as the example to copy, and a line in the file the agent reads before it starts. Now the rule is retrievable, and it is also enforced, which are two different things and you want both.

This is where it stops being tidiness and starts being about autonomy. If a rule exists only in your review, then you are the part of the system that enforces it, and the agent can never be left alone with anything that touches it. Every unwritten requirement is a lap of rework you will pay for again next week.

No model has ever worked at your company

The tempting reading is that this is a gap in capability, closed by the next release. It is not. No amount of training on the public internet contains the fact that refunds over five hundred euro need finance to sign off, or that enterprise customers have had an exception to that since before you joined. This knowledge is private, it is yours, and it changes. A better model arrives knowing more about the world and exactly as much about you, which is nothing. Closing that gap is the work around the model, not inside it.

The good news is that these requirements are easy to find once you look for them, because they leave traces. They are the review comments that cite nothing. They are the sentences that start with “we always” and the ones that contain the word “obviously”. Each of those is a rule that was never written down, because between people it never had to be. That is the whole difficulty: the rules nobody had to write down are the ones your agent breaks first.