Automata from Agent Traces: Failure and Next-Step Prediction
A compact FSM built from agent traces predicted both next actions and failures across public datasets.
The paper collapses whole trace corpora into finite-state machines of 7 to 43 states. On twelve public datasets, those machines replayed held-out data with at least 0.997 fitness and similar topology across splits. FSM-state context beat Agent Workflow Memory for next-step prediction on every matched dataset. For failure prediction, per-state features reached held-out AUROC up to 0.94, and an online monitor ranked failing runs early enough to stop before completion. HF Daily Papers' note
The paper collapses whole trace corpora into finite-state machines of 7 to 43 states. On twelve public datasets, those machines replayed held-out data with at least 0.997 fitness and similar topology across splits. FSM-state context beat Agent Workflow Memory for next-step prediction on every matched dataset. For failure prediction, per-state features reached held-out AUROC up to 0.94, and an online monitor ranked failing runs early enough to stop before completion. HF Daily Papers' note
score 4