Supervised until proven: how an agent earns autonomy

No WeBuild agent gets the keys on day one. Every build ends with a proving run: the agent does the work, and a human approves every output before it moves. The agent drafts the reply; your team sends it. The agent stages the update; your team commits it.

This sounds slow. It's the opposite — it's the fastest safe path to autonomy, because of what happens with every correction.

Every miss becomes a fix

When your team edits a draft, that edit is information: a rule that wasn't written down, a tone the business actually uses, an edge case the spec missed. During the proving run those corrections get folded back into the agent's playbook. The agent you approve in week four is materially better than the one that started week one — and it got better on real work, not toy examples.

Autonomy is granted per category — never in general

When the agent stops needing edits on a category of work — say, routine first-touch replies — that category gets autonomy, on your sign-off. Novel enquiries might stay draft-only for months. And some categories never earn it, by design: anything touching money, bulk data, or permissions keeps a human approval forever. Those are gate classes in The Cage, our published security spec.

Why we won't skip it

Clients occasionally ask to go straight to autonomous — the demo looked great, why wait? Because the demo ran on the cases we chose, and production runs on the cases reality chooses. The proving run is where the gap between those two closes. An agent that embarrasses you in front of one customer costs more trust than a month of approvals — we'd rather lose the impatient sale than build you that story.

The full engagement arc — scope, build, prove, own — is on the Method page.