PERMISSIONS AND ACCESS CONTROL

HUMAN APPROVAL FOR HIGH-RISK ACTIONS Some actions are risky enough that no automatic rule should be trusted to decide them alone. These include irreversible actions such as deleting data, sending money, publishing publicly, and sending messages on someone's behalf, and actions with large scale such as modifying many files at once. Human approval means the harness pauses and requires a person to confirm before the action proceeds. DESIGN RULES FOR APPROVALS Define clearly which actions require approval, using objective triggers such as a threshold: more than ten files, more than a set amount of money, or any action outside a defined scope. Show the approver exactly what will happen, in plain terms, including the specific targets. Make approval independent of the agent's confidence. An agent can sound extremely sure and still be wrong or manipulated, so the trigger must be based on the action, not on how confident the model claims to be. Avoid approval fatigue by reserving approvals for the genuinely risky and by using other controls for the rest. Record who approved what and when. Allow the human to deny, and make sure denial is handled gracefully by the agent, which should explain and offer alternatives rather than retrying. Consider time limits, since a stale approval should expire.