From Confident Closing to Silent Failure: Characterizing False Success in LLM Agents

LLM agents can fail silently by asserting task completion even when the underlying environment state shows the action did not occur. We study this failure mode, \emph{false success}, on 9,876 tau2-bench trajectories across 8 model families and 3 customer-service domains. We introduce a three-class decomposition of failed trajectories---false success, honest failure, and ambiguous---validated against human annotation (91.5\% agreement, $\kappa=0.86$). False success is common but environment-dependent: it accounts for 45--48\% of failures in single-control airline and retail domains, but only 3\% in dual-control telecom. LLM-as-judge evaluation systematically misses this failure mode: across 5 judges, 5 prompt strategies, and a strong baseline with full task specifications, no judge exceeds AUROC 0.65. Judges anchor on confident closing language, scoring assertion-rich traces 0.27--0.36 more ``completed'' regardless of tool evidence. Lightweight TF-IDF detectors instead achieve task-disjoint AUROC 0.83 with sub-millisecond inference. Although zero-shot cross-domain transfer is limited (AUROC 0.69), few-shot target-domain calibration recovers performance, reaching AUROC 0.79 with 50 examples per class and 0.83 with 100 examples per class. These results suggest that production monitoring should treat lightweight detectors as efficient, domain-calibrated triage filters rather than relying on LLM judges as primary detectors for false success.