One bad call: how an agent's autonomy is withdrawn.

When we walk people through earned autonomy — report, then suggest, then do, each tier bought with clean approvals — the question that matters usually arrives last: what happens when it gets one wrong?

Here's exactly what happens. The human rejects the work and writes the reason in a line. The agent drops straight back to the bottom tier — not one rung down, all the way — and starts re-earning from zero. The rejection reason doesn't evaporate: it's distilled into a candidate rule, and if a human confirms it, it becomes part of how that agent works forever.

Why one rejection outweighs thirty approvals

Because trust is asymmetric, and pretending otherwise is how automation disasters happen. Thirty clean approvals tell you the agent handles the cases it has seen. One rejection tells you there's a class of case it doesn't handle — and until you understand that class, the safe assumption is that there are more. Dropping to the bottom tier isn't punishment; it's the honest statement that the track record needs rebuilding with the new rule in place.

It's the instinct any good operator already has about a process that has failed: stop, understand the class of case that broke it, and only then let it run again. What a system adds is that the correction lands in one place. Once a human confirms the rule, it applies to every case that agent handles from that point, and you can read it back to check.

Lose quickly, regain slowly

A system built to keep confidence is dangerous; a system built to lose it fast and rebuild it slowly is one you can leave running overnight. That asymmetry is the whole answer to "how do we stop it acting outside what we agreed". Not a smarter model, but a structure where one bad output withdraws the permission immediately, and every step of the way back is a decision a named person makes.

If you're evaluating any agent system — ours or anyone's — skip the demo of its best day. Ask what happens on its worst: who finds out, how fast, what it loses, and where the lesson goes. If the answers are vague, so is the governance.