An agent never asks what I mean

Managing agents — published 4 September 2026

I manage thirty engineers. And on a personal project in progress, outside my professional activity, I have several positions held by agents.

These are two management experiences, not one experience and a set of tools. The positions have scopes, deliverables, acceptance criteria. What changes is what happens when a request is poorly framed.

An engineer comes back. An agent produces.

The mechanism, in one sentence

When a request is vague, an engineer gets stuck on it. Either the question comes back, or something wrong is delivered and the disagreement surfaces at review. Either way the vagueness is revealed, and it is revealed early.

An agent never gets stuck. It produces something plausible for any request, including a badly posed one. Vagueness no longer blocks: it passes through.

This is not a failure of intelligence. It follows from what a generated answer is — there is always one.

An example, taken from this system

I asked for a taxonomy of the cases the system does not catch. Analytical work, on real data, with an expected output.

Two sentences of the request should not have to be there.

The first: derive the categories from the data, do not start from the list I am giving you as an example. Without it, you get a classification that hugs the suggestion. It looks correct. It is confirmatory.

The second: if a category holds fewer than five cases, say so. Without it, you get five neatly aligned percentages, two of which rest on a handful of cases, and nothing flags it.

An engineer given that task would probably have told me on their own that two categories were too small. Or they would not have said it the first time, I would have caught it at review, and they would never have forgotten it again.

That is the whole difference.

Framing no longer amortizes

With a team, a rule set once holds for years. You explain once why a proportion over twelve cases does not get published; it is acquired. The person knows it, passes it on to the next hires, and the initial cost spreads across hundreds of tasks.

With agents, that rule is paid in full on every request. It is never acquired. It does not carry from one position to another. It does not improve on its own.

The cost of framing becomes a variable cost, where it used to be an investment.

That is the first thing that surprised me, because it is the opposite of what is announced everywhere. Less execution is promised. More specification is what arrives.

What moves

In a team, the asset is the person. You invest in someone, and that investment produces for years, including on subjects nobody anticipated.

Here, the asset is the written request. What I accumulate is not a position, it is the text that defines it — and the list of everything that has already gone wrong.

In practice, my requests have grown longer. They now carry explicit prohibitions that have nothing to do with the task: do not start from my suggestion, flag small counts, tell me the cause before working around it, write “not measured” rather than an estimate.

Each of those sentences is there because once, it was missing.

What this changes for a manager

Three things, none of them comfortable.

Framing work moves upstream. It can no longer be delegated to iteration, since there is no longer anyone to ask the question that reveals the misunderstanding.

Review becomes the only control point. With a team there are several: the question asked, the disagreement in a meeting, a peer’s remark. Here there is one, and it sits downstream.

And nothing grows in competence. A position held by an agent is as good on the hundredth piece of work as on the first. That is comfortable — no learning curve — and it is also a hard limit: the only possible progress is that of my own requests.

What I cannot conclude

This note describes an organization of a few positions on a personal project, not a production team. I do not know what holds at that other scale.

I have measured neither the time spent framing nor how it evolves. The impression that my requests are growing longer is an impression, not a reading — and I spend enough time documenting my measurement errors not to present it as anything else.

This is the first note in a series. It describes what I observe after a few months, not a conclusion.

A note from the Managing agents series, from a hands-on manager’s point of view.