Autonomy levels

The four levels of how far the factory may go on a repository, how the tenant maximum, the default and per-repository values combine, and the limits on concurrency, retries and waiting.

Autonomy is how far the factory may go on a repository without a person. It is a ladder: each level allows everything below it.

Level	Console	The factory may
`observe`	Observe	Read, decide and record. It writes nothing to GitHub.
`suggest`	Suggest	Comment and post reviews.
`draft_pr`	Draft PR	Open draft pull requests on its own branches.
`auto_merge_low_risk`	Approve when ready	Approve a pull request once it is ready. It never merges.

How the level is chosen

Three values combine:

autonomy.max (Maximum autonomy) caps everything. No repository and no workflow can exceed it. Default draft_pr.
autonomy.default (Autonomy of a repository with no settings of its own) applies to every repository in a managed org that has no settings row. Default draft_pr.
A repository’s own autonomy, set on its panel on the Repos page.

The effective level of a repository is its own value if it has one, otherwise autonomy.default, and in both cases no higher than autonomy.max.

Lowering autonomy.max below a repository’s stored value lowers that repository’s effective level at once, but keeps the stored value. Raising the maximum later restores it. Saving a repository’s autonomy above autonomy.max is refused with autonomy_above_max.

Each default workflow needs a minimum level. pr-review needs suggest, the fix and sweep defaults need draft_pr, and approve-ready needs auto_merge_low_risk. A default skips a repository below its level and records a trigger.skipped event, at most once an hour per repository.

Plan approval

autonomy.plan_approval (Plan approval required for belai:feature) is on by default. A run that dispatches the belai:feature crew, or the issue-to-pr default, stops after the planner and waits for a person to approve the plan before any code is written. Switch it off to let those runs go straight from plan to build.

Concurrency, retries and waiting

These settings sit in the Autonomy section of Settings:

Setting	Console label	Range	Default	Effect
`autonomy.concurrency`	Max concurrent runs	1 to 64	4	Runs of the tenant that may be `running` at once. Others queue.
`autonomy.parked_concurrency`	Max runs waiting for a machine	1 to 1000	50	Runs that only wait for a host do not use a concurrent slot. This caps how many may wait.
`autonomy.no_host_ttl_hours`	Hours a step may wait for a machine	1 to 168	6	A crew step with no host to run on waits this long, then the run ends with `no_host_timeout`.
`autonomy.claim_ttl_minutes`	Minutes a dispatched card may go unclaimed	1 to 1440	60	A crew that starts but never opens a session for its card fails the attempt with `card_unclaimed` and retries, so the run does not hold a slot for hours.
`autonomy.max_attempts`	Stop the line after failed attempts	1 to 10	3	How many times a failed step is retried before the run stops.

The same section also holds two Belai keys that every factory session reads: belai.mode (Default agent mode: goal, agent, plan or auto) and belai.ask_permission. See Agents.

Managed orgs and repositories to set a repository’s own level.
Default workflows for the level each default needs.
Settings reference for every key.