The loop keeps you able to notice

Reaching a person is not the safeguard. A person asked to approve what the machine has already handled stops evaluating and starts confirming. Mitchell, Ghosh, and Passi call that approval fatigue.¹ Bainbridge named the other half forty years earlier: the more a system handles, the less practiced its supervisor, and the moment that needs a person is the moment that person is least ready.²

Corrective competence is what remains when the agent is gone. This is what has to be present while the agent is running. A run of approvals can pass while the skill the exception needs has already left.

The test is the one Manual on purpose uses. Does this prompt ask the person to notice something they would not otherwise have seen, or to confirm a pattern they have already learned to trust? The first is a rep. The second is a toll, the same test friction uses.

Enforced by

• A waiting or permission event the queue already holds does not raise another desktop notification: the open row is the block, and a second toast asks for a yes on it.
• The console shows a permission's close control only after that row is opened, so the paths the event named, and the message, have been on screen.
• session dispose stays a typed id.
• A failure still notifies, because it is news and news does not enter the queue.

Good

What raises a desktop notification, and what does not:

a failure                                     -> notifies; it is news
a waiting or permission event already queued  -> no second toast; the open row is the block

Bad

Agent wants to run `magus run test .`. Allow? [y/N]

Asked a hundred times a day about a command that is always fine, the person learns to press y, and presses it the one time it was not fine.

See also

• Corrective competence
• Human in the loop