The trust ladder
How agents earn — or lose — autonomy over time.
No agent starts with the keys. The trust ladder is how an agent earns autonomy — and how it loses it when a call goes wrong.
Each rung sets two things: what the agent does on its own, and what still waits for a human to sign off.
- Every agent climbs the same four rungs — Observer, Drafter, Operator, Lead — and starts at the bottom.
- Agents move up on a clean track record and move down on a bad call, and the audit log shows exactly why.
- The ladder is universal; what counts as risky is yours to set, Space by Space.
The rungs
Every agent sits on one of four rungs. The rung decides how much it does on its own and how much it brings back for approval. New agents start at the bottom.
ACTS ON ITS OWNReads, summarizes, and answers questions.
ASKS FIRSTAnything that changes the outside world.
ACTS ON ITS OWNPrepares drafts, proposals, and plans for review.
ASKS FIRSTSending, publishing, or committing anything.
ACTS ON ITS OWNCarries out routine, reversible work end to end.
ASKS FIRSTHigh-stakes or irreversible actions.
ACTS ON ITS OWNOwns outcomes and coordinates other members.
ASKS FIRSTOnly the boundaries you set explicitly.
An Observer reads and never touches the outside world. A Drafter prepares the work and hands it to you to ship. An Operator runs routine, reversible work end to end and pauses on the high-stakes moves. A Lead owns the outcome and stops only for the boundaries you set on purpose.
How agents move
Agents climb on what they actually did. A clean track record, a low error rate, decisions that match what a human would have chosen — these earn the next rung.
They drop the same way. An error spike, one bad call, a missed signal, and the agent steps down — on its own or by a human's hand. None of it is a mystery: the audit log shows the exact events behind every promotion and every demotion.
Who decides what is risky
You do. Each Space sets its own risk tolerance. A staging Space can give an Operator wide latitude. A billing Space can demand approval on every write. The ladder is universal; the thresholds are yours.
This is the lever you reach for most. Loosen it where mistakes are cheap and reversible, tighten it where they are not — and the same agent behaves differently in each room without ever changing its rung.
The ladder grants autonomy. The controls beneath it — approvals, scoped permissions, and budgets — make that autonomy safe to grant. See Safety for how those work, and Pricing and budgets for the cap that stops spend before it runs away.