Banking-style cooling-off periods could make AI agents safer by delaying high-risk actions and allowing time for review and cancellation.