AgentDebugX Toolkit Enhances LLM Agent Debugging and Recovery

Debugging failures in Large Language Model (LLM) agents is notoriously difficult because the visible error often stems from an earlier, hidden cause. Existing tools primarily offer execution trace replays but lack robust support for identifying root causes or translating diagnoses into effective recovery actions. A new open-source framework, AgentDebugX, aims to solve this by structuring debugging as a continuous cycle of detection, attribution, recovery, and rerun. Central to this is DeepDebug, a multi-turn diagnosis system that understands global trajectories, investigates with structure-guided methods, and cross-examines to pinpoint the exact agent and step of failure. AgentDebugX has shown significant improvements in attribution accuracy and task repair rates on benchmarks like Who and When, and GAIA. It is available as a Python library, CLI, web console, and an agentic skill, also featuring an optional Error Hub for sharing and reusing debugging insights.