AgentRewind: Recoverable Execution for Long-Horizon LLM Agents

TL;DR - AgentRewind is a runtime recovery framework that checkpoints both agent context and environment state, enabling long-horizon LLM agents to roll back after errors and retry with knowledge from failed attempts.

• Aligns checkpoints across the agent’s context and its controlled environment.
• Addresses error propagation after failures occur, rather than relying only on prevention.
• Introduces MettleBench for measuring completion and partial progress on multi-requirement engineering tasks.
• Improves task success and checklist progress across multiple models, execution strategies, and agent harnesses versus tested baselines.