PrefixGuard: Online failure warning and trajectory-based diagnosis for LLM agents

PrefixGuard is a neurosymbolic method that provides online failure warning and trace-based diagnosis for LLM agents. A GRU predicts failure risk for a specific time horizon from learned events, and a DFA organizes event histories into paths with risk labels. Across all 24 benchmark-horizon settings, its mean test AUPRC exceeds the strongest baseline by 12.0-21.9 percentage points; under a 10% FAR budget, the average recall per benchmark is 56.1%-93.0%.