New Microsoft + Tsinghua Univ paper finds that agent failures are far easier to diagnose when the LLM gets a structured view of the run.

Long agent runs are messy. An early mistake can create later symptoms, so an LLM reading the whole history may blame the wrong step.

Give the LLM a structured run instead of the raw conversation; that raised GPT-5.1 exact localization from 3.63% to 31.35%.

better agent debugging depends heavily on how the run is represented, not just how strong the judge model is.

This is not solved yet; exact localization is still only 31.35%, but make structured traces and explicit failure checks part of the system before asking an LLM to explain what went wrong.