AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents
AgentDebugX turns agent debugging into a detect-attribute-recover-rerun loop, with root-cause tracing at its center.
The paper says agent errors often show up later than the step that caused them, making trace replay insufficient. Its DeepDebug component uses multi-turn diagnosis to identify both the responsible agent and step. On the Who and When benchmark, it reports 28.8% exact agent-and-step accuracy with qwen3.5-9b, ahead of a 21.7% single-pass baseline. On GAIA, it repaired 13 of 73 failed tasks in one rerun and raised overall accuracy from 55.8% to 63.6%. HF Daily Papers' note
The paper says agent errors often show up later than the step that caused them, making trace replay insufficient. Its DeepDebug component uses multi-turn diagnosis to identify both the responsible agent and step. On the Who and When benchmark, it reports 28.8% exact agent-and-step accuracy with qwen3.5-9b, ahead of a 21.7% single-pass baseline. On GAIA, it repaired 13 of 73 failed tasks in one rerun and raised overall accuracy from 55.8% to 63.6%. HF Daily Papers' note
score 5