Paper: Automata from Agent Traces for Failure Prediction

rohanpaul_ai · x · 2026-09-02

The paper "Automata from Agent Traces" proposes collapsing entire LLM agent trace corpora into a single, compact Finite State Machine (FSM). Across 12 public datasets, these FSMs are tiny (7-43 states), build in milliseconds, and replay held-out data with >=0.997 fitness. The method outperforms Agent Workflow Memory for next-step prediction and enables online monitors to rank failing runs above passing ones from partial traces, offering a model-agnostic structural primitive for safety auditing.

Related event: Study: LLM Agent Behavior Collapses into Compact Finite State Machines(2 posts)→

Original post →

More from coding & agent

coding & agent channel →