🕵️ The cause of an agent's failure may be findable from the "flow of success," not from failures. Title: Tracing Agentic Failure from the Flow of Success URL: 🧩 Overview Oat is trained only on successful trajectories and pinpoints which steps of a failed trajectory caused the failure, with no labels. Accepted to NeurIPS 2026. 🎯 The problem Existing failure attribution needs either expensive LLM prompting pipelines or failure data with step-level error labels. Both are costly. 🔧 Method ・Represent each step with LLM hidden states ・Learn the dynamics of successful runs with a Neural CDE ・A gated control path improves robustness to distribution shift ・At inference, steps that deviate from the learned flow get high anomaly scores (top-k or conformal prediction) Only about 100 successful trajectories are used for training. 📊 Results ・In-domain F1: Oat 0.435 vs. GPT-4o 0.212 and GPT-5 0.181 ・About +7 F1 points out of distribution too ・About 7 ms per trajectory with zero tokens, vs. about 39.6 s for GPT-5 ・200-5000x faster on scoring alone #AIAgents# #LLM#