feature · autonomy-levels
Extrapolated — control
5 tags · 39 out · 39 in · 3 artifacts · 0 gaps · 1 corpus docs
Autonomy levels are an explicit spectrum of how much an agent may plan, decide, and act without a human in the loop. The MIT 2025 AI Agent Index operationalizes this roughly as L1–L5: from turn-based collaboration (human drives each step) through mixed initiative to long-running browser/enterprise agents that proceed with little mid-execution intervention.
This feature is the declared ceiling, not the gate UI. It answers: for this agent, in this mode, what is allowed unsupervised? Chat agents often sit at L1–L3; browser agents at L4–L5; enterprise builders may be configured at L1–L2 yet deploy event-triggered agents at L3–L5 — a split the OS must not hide.
Without declared ceilings, every capability (tools, computer use, payments) defaults to whatever the vendor shipped. Canon requires company shape — authority and thresholds — declared as reviewed diffs. Autonomy levels are how those thresholds are expressed for agents. They also make Stage-1 “stalled vs slow” meaningful: a high-autonomy path that never hits a gate cannot be measured as waiting.
Graph edges for this feature — technologies, category lens, companies, DM.
features/autonomy-levels/README.md
No open gaps for this entity.
Autonomy Levels · ok
features/autonomy-levels/README.md
Research path: features/autonomy-levels