LLMs, GenAI, RAG, agents, evaluation · insight-1572
Label failed agent runs with a shared failure taxonomy at the first wrong step
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Applied AI, Data scientist
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A team reviews a pile of failed agent traces and writes a different free-text diagnosis for each.
Action
Define a small taxonomy such as bad plan, wrong tool, bad arguments, misread result, missing verification and premature stop, annotate each failed trace at the first step where it went wrong, and rank causes by frequency and cost.
Reason
A study of over 1,600 annotated multi-agent traces organised failures into a 14-mode taxonomy in three categories, including a verification category (a preprint on multi-agent systems; the labels here are a labelling aid, not that taxonomy), which shows that labelled categories can be tallied across traces.
Boundary
Any agent with enough failed traces to review.
Symptom or risk
Failures cannot be counted by cause, so effort goes to the loudest example rather than the most common mechanism, and the final symptom hides the earlier error.
Exceptions
Below a few dozen failures, read them all instead.
Trade-offs
Annotation takes time and labels are subjective; check agreement between two reviewers.
Pitfalls
Labelling the last visible symptom rather than the first divergence.
Lessons and concepts
Claims and their sources
Cemri et al. report a taxonomy of 14 failure modes in three categories (system design, inter-agent misalignment, task verification) built from more than 1,600 annotated traces of multi-agent systems (preprint).