AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
The paper proposes checkpointing both an agent’s context and its environment so failed long tasks can be rewound and resumed.
AgentRewind is presented as a runtime recovery framework for LLM agents working over long execution horizons. The authors argue that early mistakes can contaminate later context and environment state in ways that ordinary plan refinement or safety checks do not repair. They also introduce MettleBench, a benchmark for long-horizon engineering assignments with related requirements and partial-progress scoring. Across tested tasks, models, execution strategies, and agent harnesses, the paper reports higher task success and average checklist progress than the baselines. ArXiv · AI/CL/LG's note
AgentRewind is presented as a runtime recovery framework for LLM agents working over long execution horizons. The authors argue that early mistakes can contaminate later context and environment state in ways that ordinary plan refinement or safety checks do not repair. They also introduce MettleBench, a benchmark for long-horizon engineering assignments with related requirements and partial-progress scoring. Across tested tasks, models, execution strategies, and agent harnesses, the paper reports higher task success and average checklist progress than the baselines. ArXiv · AI/CL/LG's note
score 5