Autonomy in Brainis is not a switch; it is a ladder. Every module and agent operates at a level you choose, and the level decides what happens between an AI's conclusion and an action.
The ladder
| Level | Name | What the AI may do |
|---|---|---|
| 0 | Manual | Nothing. AI features off for this scope |
| 1 | Assistive | Answer and summarize when asked |
| 2 | Supervised | Draft and suggest; every action needs explicit approval |
| 3 | Conditional | Execute low-risk actions; everything else needs approval |
| 4 | Delegated | Execute within an explicit verb allow-list |
| 5 | Managed | Run scheduled work within contract; report by digest |
| 6 | Proactive | Initiate work from signals, within contract |
| 7 | Autonomous | Operate the scope end to end; approvals only for flagged risks |
| 8 | Self-Optimizing | Additionally tune its own rules within bounds |
Each level carries a default risk ceiling - the maximum action risk executable without approval - which you can tighten further per contract.
Levels are per-scope
You set autonomy globally, per module, and per agent, and the strictest applicable bound wins. A common shape: level 2-3 globally, level 4-5 for a proven agent in a low-stakes module, level 0-1 for Finance until trust is established.
Choosing a level
Start lower than feels efficient. The cost of a level too low is some approval clicks; the cost of a level too high is cleaning up actions you did not want. Raise one level at a time, per scope, after reviewing what the AI proposed at the current level for a week or two.
Important: Regardless of level, some actions always require approval - those flagged approval-always in the action registry (typically destructive or high-monetary-impact verbs). Level 8 does not mean unaccountable; everything still lands in the audit trail.