autonomyLevel
autonomyLevel records how much human oversight an agent works under. It is information, not a permission.
| Defined in | draft-kushwaha-scim-agent-governance-00 |
|---|---|
| Type | string |
| Values | autonomous, semiAutonomous, supervised |
| Required | No |
| Mutability | readWrite |
The three values
autonomous: the agent starts and finishes actions without a human in the loop.semiAutonomous: the agent acts on its own inside a bounded task, but a human starts it or approves checkpoints.supervised: a human reviews or approves the agent's individual actions.
What it is not
The draft is firm here. The attribute is classification metadata recorded at provisioning time. It is not an authorization control, it must not be treated as authoritative unless local policy or another trusted process validates it, and it must not be the only input to an access decision.
Who might use it
The draft says regulatory frameworks that classify AI systems partly by the degree of human oversight, such as the EU AI Act, are an anticipated consumer of this value. The value set is protocol neutral, so it could also appear as a claim value in other protocols.
Audit
Changes to the attribute should be audited, because a quiet change from supervised to autonomous is the kind of thing a reviewer wants to see.
Example
Value used in the draft's own example.
Source: draft-kushwaha-scim-agent-governance-00 (datatracker). Drafts change. Read on 18 October 2026. Independent site, not affiliated with the IETF.
Own the name for this category.
agentscim.com is for sale. If you build identity, security or agent tooling, write to me. Serious offers get a reply within 24 hours.
Make an offer