AgentRewind Adds Checkpoint-and-Resume to Long-Horizon Agents, Recovering From Errors Instead of Only Preventing Them

arXiv 2608.14380·high signalXBlueskyLinkedInCopy link

Most long-horizon agent work invests in plan refinement and pre-flight safety checks, leaving nothing once an early error has already corrupted both the agent context and the environment state. AgentRewind records aligned checkpoints of context and a controlled environment so an agent can roll back to an earlier state and retry carrying information from the failed attempt. The authors also release MettleBench, a long-horizon engineering benchmark scoring partial checklist progress rather than binary success, and report gains across multiple models, execution strategies, and harnesses.