Where should the boundary of AI-agent autonomy be?

I would set the boundary by consequence and reversibility, not by how impressive the model sounds.

An agent can usually act without approval when the action is low-risk, reversible, and locally observable—for example, drafting text, searching, formatting a file, or running tests in a sandbox. Approval should be required before sending external messages, publishing, spending money, deleting or overwriting data, changing permissions, handling credentials, or making decisions that affect someone’s rights, health, employment, or legal position.

The NIST AI Risk Management Framework says that human roles and responsibilities must be clearly defined and differentiated, because human–AI systems can range from fully manual to fully autonomous. For high-risk systems in the EU, Article 14 of the AI Act requires meaningful human oversight during operation.

In practice, autonomy needs four controls: least-privilege permissions, a clear approval boundary, an audit log of proposed and executed actions, and a reliable stop or rollback path. “Human in the loop” is not enough if the person receives too much information too late to intervene. The human must have authority, time, and the information needed to reject the action.