Trajectory-Centered Error Analysis in Long-Horizon LLM Agents
November 11, 2026
An empirical study of 95 failed runs across six multi-agent systems identifies that faults typically enter at the median 0.18 mark of a run. Most errors originate from early wrong decisions or fabrications regarding task state that propagate through the agent's trajectory.
HOW THIS AFFECTS YOU
●
builderYou should focus your debugging and control mechanisms on the very beginning of agent trajectories, where most errors originate.
●
researcherThis new annotation protocol allows you to map how a single fault propagates from its origin to a final failure symptom.