research
The First Error Was Not Always the Fatal One
TrajDebug follows mistakes through long agent runs to distinguish resolved detours from failures that still reach the terminal state.
Summary
TrajDebug follows mistakes through long agent runs to distinguish resolved detours from failures that still reach the terminal state.
TrajDebug compresses long histories, identifies evidence for local errors and tracks whether each error was resolved or remained responsible for a failed outcome. Its companion TrajErrBench contains 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro. The authors report stronger critical-error detection than their baselines and actionable downstream diagnoses; the preprint does not establish perfect attribution in production agents.
Why it matters
TrajDebug follows mistakes through long agent runs to distinguish resolved detours from failures that still reach the terminal state.
Limits and context
- The authors report stronger critical-error detection than their baselines and actionable downstream diagnoses; the preprint does not establish perfect attribution in production agents.
Key claims
TrajDebug follows mistakes through long agent runs to distinguish resolved detours from failures that still reach the terminal state.
Qualification: The authors report stronger critical-error detection than their baselines and actionable downstream diagnoses; the preprint does not establish perfect attribution in production agents.
Evidence: source-2026-08-07-003
Sources
- arXiv preprint 2608.06346arXiv · primary research
Corrections
No corrections have been recorded for this story.