Levels of autonomy
A five-level scale of how much an agent acts without the user, named by the role the user plays, and why autonomy is a design choice rather than a by-product.
Autonomy is not capability
How much an agent should do alone is often treated as a consequence of how capable it is. Feng, McDonald and Zhang argue otherwise: an agent’s level of autonomy can be a deliberate design decision, separate from its capability and its operating environment1. They define autonomy as how far an agent is designed to act without user involvement1.
Diagram: my own, restating the separation argued in 1.
The practical consequence is that a highly capable agent can still be deployed at a low autonomy level if the stakes call for it.
Five levels, named by the user’s role
The framework is built around five roles a user can take when interacting with an agent: operator, collaborator, consultant, approver and observer1.
Diagram: my own rendering of the levels in 1. Level names and one-line descriptions follow the paper’s summary figure.
| Level | User role | What it means (per the paper’s summary) |
|---|---|---|
| L1 | Operator | The user directs the agent’s actions |
| L2 | Collaborator | User and agent collaboratively plan, delegate and execute |
| L3 | Consultant | The agent takes the lead but consults the user for expertise or preferences |
| L4 | Approver | The agent involves the user only in risky or pre-specified scenarios |
| L5 | Observer | The agent operates with full autonomy under user monitoring |
For each level the paper also describes how the user can exert control and raises open questions about designing the user-agent interaction1. It proposes “autonomy certificates” as a way to communicate the intended level to stakeholders1.
Using the scale
Some ways to apply this, as my own working notes rather than claims from the paper:
- Pick the level per task or action, not once per product. Reading a file and sending an email do not need the same level.
- Move the boundary with evidence. Raise autonomy only after evals show the agent handling that class of task reliably.
- Make the level visible. Users should know whether the agent will ask before it acts.
- Decide what triggers an approval. The approver level depends on a clear rule for what counts as risky. See Human in the loop and Permissions.
Anthropic’s guidance points the same way from the engineering side: autonomy raises cost and the potential for compounding errors, so agents should be tested extensively in sandboxed settings and run with appropriate guardrails2.
Related
- The loop being governed: What is an agent.
- Hooks as an enforcement point for these levels: Hooks.
Sources
- [1]Levels of Autonomy for AI Agents — K. J. Kevin Feng, David W. McDonald, Amy X. Zhang · arXiv · 2025PARAPHRASED
- [2]Building effective agents — Erik Schluntz and Barry Zhang · Anthropic · 2024PARAPHRASED