Main idea

Evaluating probabilistic outputs, keeping a queryable product memory, writing documentation a machine can use as readily as a human, keeping a repository readable, stating crisp acceptance criteria, orchestrating chains of tools: none of this is new. That work existed, scattered, held by whoever had the time, and nobody lost much by neglecting it.

The moment a team puts agents to work, it stops being peripheral. The quality of the context provided determines the quality of what is produced, and a context nobody maintains degrades every output at once. What used to be a matter of care becomes a condition of operation.

These needs do not necessarily become positions. They become responsibilities that have to be assigned — and the question put to an equipped team is no longer "who does what?" but "who judges, who builds, who guarantees, who evaluates, who maintains the memory, who orchestrates?".

Why it matters

It explains why two teams with the same tools get very different results: the difference is not made on the tool, it is made on what nobody was in charge of holding.

And it gives a list of responsibilities to assign explicitly before their absence is noticed, rather than after a false output has reached production.

Nuances and limits

Making these responsibilities visible does not say how to resource them. A team that assigns them without freeing up the corresponding time gets a list of nominal owners and the same neglect as before.

And some of these loads are transitory: tool reliability improves, and an evaluation responsibility sized for the current state of the models may become disproportionate.

Open questions

• Which of these responsibilities gain from being held by a dedicated person, and which degrade the moment they are taken out of the team that produces?