Your Coding Agent Can't Undo Its Own Mistakes - Tanmay Sah & Dolly Sah

Tanmay Sah and Dolly Sah presented UndoBench and EvoUndo at AI Engineer, two research projects on whether coding agents can reverse their own actions. Their answer is that finishing a task and undoing it are separate abilities. Version control covers files in a repository, but not external API calls, runtime configuration or notifications already sent, and retrying a failed step does not reverse the side effects of the first attempt. In EvoUndo, 197 of 478 changes that improved an agent's capability, about 42%, also broke its ability to recover. The speakers propose capturing the original state before acting, choosing a response among retry, restore, compensate or escalate, and verifying the result independently of the agent that made the change. The finding means a capability gain measured on task success needs a separate recovery check before it ships.