Tracing Agentic Failure from the Flow of Success

The paper introduces OAT, an unsupervised failure-attribution method for LLM-based agents that trains exclusively on successful trajectories. Using neural controlled differential equations, OAT learns the latent-space dynamics of successful execution flows. At inference time, it assigns anomaly scores to steps in a failed trajectory according to their deviation from those learned dynamics, then identifies likely error steps. With only 100 successful training trajectories, the reported experiments show that OAT is 200–5000 times faster than prompting-based baselines and improves F1 by 20% in-domain and 7% out-of-distribution.