Main idea

What a product is made of — its objects, the states they go through, what is forbidden and from what point — lives in the code and in a few people's heads. That arrangement held as long as you could knock on the right colleague's door: the missing rule took three minutes to obtain, and nobody needed to write it down.

An agent handed a change knocks on no doors. It receives a few sentences of instruction and whatever it manages to read in the code; everything else it fills in — plausibly, coherently, and sometimes wrongly. It does not flag that it filled anything in, because it does not know that it did.

Hence an unexpected measurement: the gap between what we wanted and what we get back is exactly the part of the mental model we had not written down. Every rework after a delegated change points to a rule that stayed spoken.

Why it matters

It changes the justification for writing the product model down. The classic argument — corporate memory, the departure of an expert — has always lost against urgency, because its due date is distant. Here the due date is next week: whatever is not written is paid for in immediate, visible, attributable rework.

It also gives a free diagnostic instrument. Rather than guessing what ought to be documented, you let an agent work and write down what it invented. The list of holes builds itself.

And the size of the artifact counts: a few hundred object cards fit in a context where thousands of code files do not — the agent reads what it needs instead of rummaging.

Nuances and limits

Writing the model down does not make the agent smarter: it only stops it from inventing where a rule already exists. In areas where no rule was ever laid down, it will go on filling in, and what is then missing is a product decision, not a line of documentation.

None of this is measured until two batches of requests have been compared, handled with and without the reference model in context: the mechanism is solid, the gain remains an expectation.

Open questions

• How do you recognize, in an agent's output, what has been filled in — without rereading the area concerned as closely as if you had written it yourself?