Structural retrieval shows promise over basic RAG for agent failure prediction on trajectory snapshots

Researchers evaluated a structure-aware retrieval engine for predicting agent trajectory failures from snapshot data.

Using KNN voting, it achieved about 0.705 ROC AUC from step 60 onward on held-out trajectories, versus roughly 0.60 for basic RAG.

The test set was small—220 trajectories from a single repository—so confidence intervals overlapped and the gain is still preliminary.

The result suggests trajectory systems may need structure-aware retrieval for monitoring, reflection, and other agent infrastructure tasks.