Definition
Risk-scoped autonomy gives an agent different levels of discretion and authority according to the task's risk. The decision should consider how easily the output can be verified, whether effects can be reversed, what sensitive data or systems are involved, and how far a failure could spread.
Autonomy is therefore a set of operating modes, not one permanent score for an agent. The same coding agent might work independently in a disposable branch, require approval before merging, and be prohibited from changing production access controls. As evidence improves and controls mature, a particular task can move to a less supervised mode without granting broad authority everywhere.
Each mode needs an explicit action boundary, credential scope, approval rule, monitoring requirement, and rollback plan. The system should choose the mode from task and environment evidence, not from the model's confidence alone.
Distinguish it from nearby terms
Human-in-the-loop describes where a person participates. Risk-scoped autonomy explains why that participation or another control is required for a particular class of work. Least privilege limits permission within every mode.
Check your understanding
An agent may autonomously rewrite a test fixture in an ephemeral environment because the change is isolated and easy to verify. The same agent should not autonomously rotate production credentials merely because both tasks involve editing text.