Rerunning a Failed Multi-Agent Trajectory Reproduces the Failure Only 67.97% of the Time and Repairs 6.90%

arXiv 2608.25920·medium signalXBlueskyLinkedInCopy link

Luan et al. ask whether existing multi-agent repair methods actually fix causes or just exploit LLM sampling randomness, and build SymTrace, a replay framework that reconstructs execution up to an intervention anchor from recorded logs and regenerates only the downstream trajectory. Alongside it they release SymFail, 536 human-annotated failure trajectories with graph-linked locations, categories, and trace evidence. Across three mainstream MAS frameworks, unguided rerun reproduces failures 67.97% of the time and repairs 6.90%, while a symptom-driven intervention repairs 20.15%, a 191.89% relative improvement over prior state of the art.