Autonomy

On this page

Levels describe a particular arrangement

Feng, McDonald and Zhang’s 2025 research proposes five levels based on the user’s role. This is a paraphrase of that framework, not an AE rating or an adopted industry standard.

Level	User’s role	How the user participates
L1	Operator	Directs the work and invokes assistance
L2	Collaborator	Shares planning and execution through frequent interaction
L3	Consultant	Provides direction and feedback while the agent leads
L4	Approver	Resolves blockers or gives required sign-offs
L5	Observer	Monitors and can use an emergency stop, but cannot otherwise steer the work

Higher autonomy is not a recommendation. Identify the framework, task and operating conditions when using a level; do not label an entire company “Level 4”.

The same model, different permissions

A constructed example: one repair business lets its system collect missing information but requires approval before sending customer messages. Another permits a defined set of information requests to be sent automatically. The model could be identical in both arrangements.

Neither arrangement grants permission to order parts or diagnose faults. Those decisions need their own authority. A successful test of message drafting also says little about the quality of a proposed repair.

Observe what actually happens

Anthropic’s deployment research treats autonomy as dependent on the product and user oversight. Measurements of individual tool calls can miss human review elsewhere in the workflow.

For an operating business, record where people intervene and why: incomplete information, conflicting records, uncertain results or a decision reserved for a person. That evidence is more useful than assuming a longer run time means a better agent.