Microsoft's OAT: Debugging Agents Using Only Success Trajectories

omarsar0 · x · 2026-07-15

Microsoft and collaborators have published a new paper proposing OAT, a method for debugging agent trajectories at scale.

The authors point out that when an agent fails, the hardest part is identifying exactly which step went wrong. Existing methods either feed the entire trajectory into expensive prompt pipelines for analysis or rely on step-by-step error annotations, which are hard to collect. OAT's approach is to train exclusively on successful trajectories. It uses Neural Controlled Differential Equations to model the dynamics of these successes and then identifies the exact step where a failed trajectory deviates from this "success manifold." This enables failure attribution without needing any failed samples or manual error labels.

Related event: Microsoft Proposes OAT for LLM Agent Failure Attribution(2 posts)→

Original post →

More from coding & agent

coding & agent channel →