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

Tanmay Sah and Dolly Sah look at a capability today's coding agents are missing: undoing what they've done. Drawing on their research projects UndoBench and EvoUndo, they show that an agent that completes a task can't necessarily put things back when something breaks, because Git doesn't cover external API calls, runtime config or notifications, and retrying doesn't reverse side effects. In EvoUndo, 197 of 478 changes that made an agent more capable (about 42%) also broke its ability to recover. They lay out an approach that captures the original state, picks the right response (retry, restore, compensate or escalate) and independently verifies the result. Tanmay Sah https://x.com/TanmayGradient Dolly Sah Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!