Autonomy levels
Autonomy describes how much an agent may do without asking. It does not describe what it may do without a human release, because that answer is the same at every level.
#The levels
| Level | Means |
|---|---|
L0 | Off. The agent does not run. |
L1 | Read and propose. It can look and suggest; it takes no action. |
L2 | Act on low-consequence things within bounds - a task in a worklist, a factual status note. |
L3 | Act on consequential things, every one of them behind a human release gate. |
L4 | The widest configured level. Removes friction, never the gate. |
#Strictest wins
The effective level for a call is the lowest of four independent inputs:
- The principal's own configured level.
- The practice ceiling, from the config cascade.
- The external ceiling, if the principal arrived from outside. A practice can hold outside agents lower than its own.
- The agent type's ceiling, which is licensed by evals and expires.
A practice may lower any of these. The cascade's direction constraint stops it raising them past a fleet floor.
#Below the floor is unavailable, not denied
Each tool declares an autonomy_floor. Below it the tool is masked - it does not appear as something that was refused, because a denial invites a retry and an escalation request. It is simply not available at that level, and the denial names the effective level and where the cap came from so the reason is legible.
#Autonomy expires
An agent may act above L1 only while a recent eval run says it should. After fourteen days the licence lapses and the agent drops to proposing, whatever its configured ceiling says.