What is Autonomy Level?
Also called level of autonomy, agent permissions.
Autonomy level describes how much an agent may do without human involvement, from suggesting an action, to acting with approval, to acting freely within limits and reporting afterward. There is no single industry standard scale, so specific levels are defined per system. Levels are usually assigned per action type rather than to the agent as a whole.
A widely used informal ladder runs roughly: observe and report, recommend an action, act only after explicit approval, act and notify, and act freely within a defined boundary. Various vendors and analysts publish their own numbered versions, sometimes by analogy with vehicle autonomy levels, but none of these is a recognized standard and the numbers are not comparable across sources.
Assigning one level to a whole agent is a design mistake. The same agent may read data freely, draft external messages only for review, and never issue a refund without approval. Modeling autonomy at the action level keeps the fast paths fast while gating exactly the operations that warrant it, and it makes the policy legible to the people relying on it.
Levels are normally earned rather than chosen. A team begins with approval on everything, measures how often approvals are rejected and why, and relaxes gates only where the record shows a high approval rate and low consequences. Moving in the opposite direction, granting broad autonomy first and tightening after an incident, is the more common and more expensive path.
Autonomy level is the policy expression of ideas covered elsewhere: human review supplies the pause mechanism, tool permissions supply the boundary, and the trajectory record supplies the evidence used to check whether a level is set correctly. The level should be visible to the people relying on the agent, because expectation mismatch causes more trouble than the setting itself.
Key points
- Ranges from recommend only to acting freely within limits.
- No standard scale exists; each system defines its own.
- Assign levels per action type, not to the whole agent.
- Raise levels from measured approval history, not optimism.
In practice
A finance assistant is configured three ways at once. Reading transactions and building reports needs no approval. Drafting a payment run is allowed, but the run sits until a person releases it. Changing a supplier's bank details is never permitted at all, regardless of instruction, because that action is both irreversible and a classic target of fraud.