OAT Performs Unsupervised Failure Attribution in Agents

OAT is a lightweight failure attribution model that identifies error steps during inference without requiring step-level supervision. It is trained exclusively on successful trajectories, treating failure identification as a one-class learning problem.