TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories

TL;DR - TrajDebug is a framework for finding the earliest critical error in failed long-horizon LLM agent trajectories, paired with TrajErrBench, a 486-trajectory human-annotated benchmark. It matters because cascading errors in agentic systems are hard to debug, and pinpointing the root-cause step enables targeted fixes.

• Frames the task as "error-lifecycle tracing": beyond flagging local errors, it tracks each error's resolution status and terminal impact to decide which one actually caused the final failure.
• Two techniques address long trajectories: multi-granularity history compression (evidence may be scattered across distant instructions, observations, and context) and evidence-based error identification.
• TrajErrBench contains 486 manually annotated failed trajectories drawn from Tau2Bench (tool use) and SWE-Bench Pro (coding).
• Reported to achieve best overall performance vs. existing baselines, with application studies showing its diagnoses give actionable feedback that improves downstream agent success; code and data to be released.