Glossary
Autonomy levels
Autonomy levels are graduated settings for how much an AI agent may do without asking, so the boundary can be moved per action rather than switched on or off for everything.
Why a single switch fails
One on-off control forces a bad choice: either the agent asks about everything, which is unusable, or it asks about nothing, which is unacceptable. Neither survives contact with real work, and the usual outcome is that the control gets turned off.
Graduated instead
Levels let reading run freely while sending waits, and let a category be widened once you have evidence. Typical progression: read-only, then internal writes that are easy to correct, then outbound and irreversible actions last — each moved independently.
Widen from evidence
Start stricter than feels necessary and loosen where the record supports it. Widening is a two-minute change; rebuilding trust after an incident is not.
Related terms
FAQ
What are autonomy levels?
Graduated settings for how much an AI agent may do without asking, so the boundary moves per action rather than being switched on or off for everything. Typically reading runs freely, easily-corrected internal writes come next, and outbound or irreversible actions stay gated longest.
Why not just turn autonomy on or off?
Because a single switch forces a bad choice: ask about everything, which is unusable, or ask about nothing, which is unacceptable. Neither survives real work, and the usual outcome is that the control gets disabled entirely.