Researchers adapt grounded theory, a decades-old qualitative research method from the social sciences, into AutoTraceGT, the first automated multi-agent pipeline for analyzing LLM agent trajectories at scale. Applied across six trajectory corpora, AutoTraceGT’s generated codebooks recover 73-91% of failure modes found in human-annotated taxonomies while surfacing additional patterns those taxonomies missed, and outperform zero-shot and few-shot LLM baselines when used as a feature space for predicting agent failures.
