What autonomy levels do AI agents have?
AI agents work at five levels of autonomy: assist, prepare, act with approval, act within boundaries and autonomous. You set the level per agent and per type of case, and an agent only moves up a level when its results justify it.
- Level 1AssistAnswers and suggests, the person does the work
- Level 2PrepareCreates a draft, a person checks and sends it
- Level 3Act with approvalCarries out the steps, a person approves the result
- Level 4Act within boundariesHandles standard cases itself, people take the exceptions
- Level 5AutonomousRuns the whole process and directs other agents
Explanation
Autonomy works like a dial: you set it per agent and even per type of case. The same agent can send small credit notes itself (level 4) and have large ones approved (level 3).
Most first agents start at level 2 or 3. They move up a level when evaluations show they are ready. You record that decision in the agent passport.
The higher the level, the heavier the demands on the agentic harness: more evaluation, tighter permissions, better logging and an emergency stop that works.
What it looks like in practice
At an energy company, an agent for outage reports starts at level 2: it proposes a diagnosis and an employee decides. After three months of good results it may fix simple outages remotely itself (level 4) and calls in a technician for the rest.
