How we measure confidence to decide when an AI agent should ask for permission and when it should just act.