Draft, copilot and autonomous
Three rungs, not a switch
Section titled “Three rungs, not a switch”Every published agent starts in draft and paused. Leaving each rung is somebody’s decision, and it is recorded.
| Stage | What it may do |
|---|---|
| Draft | Opens no real run. Rehearsal only |
| Copilot | Runs for real, and every write waits for a person (the work-item claim aside) |
| Autonomous | Runs and writes on its own, within what policy and classification already allow |
Notice what does not change between them: the Gate evaluates the same in all three. Autonomous is not “no brakes” — it is “no human wait on what was already cleared”. A destructive tool is still blocked, taint still stops a write, a ceiling still holds.
The agent exists, has a published version, and opens no real run.
This is where you write, rehearse and correct. A rehearsal runs against situations you choose with the tools dry — nothing is sent to or changed in an external system.
Leave when: the rehearsal passes the situations that matter, and the ones it marked as needs a look you have understood and resolved.
Copilot
Section titled “Copilot”Runs for real. Every write stops and waits for a person to approve that specific call. The one exception is the work-item claim ($fuseone.work_item.claim), which only records in the platform which card a run holds and passes without approval, so that two runs never wait on a person to reserve the same card.
This is the rung where you learn what the agent actually does with real data, which is never exactly what the rehearsal showed.
Leave when: approvals have become routine without surprises. There is no magic count, but there are three questions:
- Are the approvals all the same? If you approve the same thing every time without thinking, it has stopped being a decision.
- Has any refusal surprised you? If so, the agent is still learning the work — or the policy is wrong.
- Would you read one of these runs’ trails and be comfortable? If that depends on who was looking, not yet.
Autonomous
Section titled “Autonomous”Runs and writes without waiting for approval, within what was already permitted.
What still stops it: destructive and financial effects, content marked as coming from outside, a policy that escalates or denies, and any ceiling.
Which means promoting to autonomous is not the moment to relax policy — it is the moment policy becomes the only thing holding. Before, a human reviewed every write and would have caught what policy let through. Now nobody does.
Promoting safely
Section titled “Promoting safely”The order that works:
- Rehearse until the situations that matter pass.
- Publish in draft and unpause.
- Copilot, and approve properly — reading the arguments rather than clicking from habit.
- Before going autonomous, write the policies that would hold whatever worries you, and turn them on in monitor mode. See whether they would match.
- Switch those policies to enforcing, and only then promote.
Step 4 is the one usually skipped, and it is what separates promoting from gambling.
Going back is normal
Section titled “Going back is normal”Dropping a rung is not a failure — it is the mechanism working. An agent that started getting things wrong after a change on the other side goes back to copilot until somebody understands what changed.
What a policy decides is in Policies, and what the Gate stops regardless of stage is in What the platform stops before it happens.