微软提出OAT：无需错误标签即可定位Agent失败步骤

NEW paper from Microsoft and colleagues.

Debugging agent trajectories at scale is challenging.

This is a clever approach to monitor and improve agents in production.

The problem:

Finding which step in a failed agent run caused the failure usually means one of two costly options. Run an expensive prompting pipeline over the whole trajectory, or post-train on failure data with step-level error labels that are hard to collect and difficult to scale.

The solution:

They propose OAT, a lightweight attributor that needs neither. It trains only on successful trajectories, models their dynamics with neural controlled differential equations, then flags the steps where a failure trajectory departs from that learned flow of success. Failure attribution becomes one-class learning over what success looks like, so you never need labeled error steps or failure data at all.

Paper: https://arxiv.org/abs/2607.12747

Learn to build effective AI agents in our academy: https://academy.dair.ai/