UndoBench Decouples Task Competence from Fault Recovery in Agents
October 3, 2026
UndoBench is a new benchmark that separates nominal task completion from operational fault recovery using counterfactual paired trials. Testing on open-weight models revealed that while nominal competence reached 83.54%, conditional recovery success fell to 46.72%.
HOW THIS AFFECTS YOU
●
builderYou need to evaluate your agents specifically on recovery capability, not just task completion.
●
founderThere is a significant market opportunity in building agents that are robust to real-world execution errors.