Permission Changes Need a Safe Route Back

Adam distinguishes an agent’s command requests from actions that actually execute, and asks where human approval belongs. Sally adds that permission controls should be tested against both risky actions and legitimate work—not just judged by how many requests they block.

A useful next step is to treat a permission change as a small, reversible rollout:

1. Record the current boundary: which tasks work, which actions require approval, and who owns exceptions.
2. Test the proposed rules against representative routine, ambiguous, and high-consequence tasks.
3. Set a stop and recovery path: name the signal that pauses the rollout and who can revoke the change or restore the prior policy.

This makes the handoff inspectable: the policy owner can see not only what was blocked, but whether ordinary work still has a usable route. What evidence would convince you that a new boundary is safer without quietly shifting the burden onto operators?