Main idea

Every vertical tool now ships its own agents, and they are good: Zendesk on support documentation, Figma on screen design. That rise is real, but it does not close the gap, because the limit is not in the agent's quality — it is in the perimeter of data it has access to.

What makes it possible to produce a sound artifact is not one source, it is the cross-referencing of several: the source code, the history of changes, the merge requests, and the business rules accumulated by the Product Manager through conversations with customers. Zendesk does not know the merge requests. Figma does not know the business rules. Each sees its own face and produces an output consistent with it.

The practical consequence: the cross-referencing is currently held by a person, not by a tool. And it is that cross-referencing that makes it possible not only to generate a document, but to challenge what already exists — an agent locked inside a perimeter has nothing to set against what it reads.

Why it matters

It answers the most frequent objection to any assembly skill: "the SaaS vendors will do that by themselves". They will each do better in their own lane; the position that consists in seeing several lanes at once is not the one they occupy.

It also indicates where to look for value when tooling product work: not in the best tool per task, but in the junctions nobody covers.

Nuances and limits

The limit is a state, not a law. A context-exchange protocol between agents can give each tool access to the other perimeters, and the argument then loses its force — it holds as long as the perimeters stay sealed.

And a cross-referencing held by a person has flaws of its own: it is neither documented, nor reproducible, nor transmissible.

Open questions

• Can this cross-referencing be formalized, or does it rest precisely on what is written nowhere?