DeFA builds an event dependency graph that unifies protocol relations and semantic dependencies across an execution trajectory. It first flags events that may violate task requirements, then traces their origins and downstream effects to construct a failure propagation graph. Using step evidence and each step’s role in the propagation graph, DeFA pinpoints the decisive error, the responsible agent, and the error category.
To handle long trajectories, DeFA partitions executions into segments and merges the detailed content of the current segment with summaries of the others, giving local diagnosis access to the global context.
On the Who and When and Who and When Pro subsets, DeFA achieves the highest responsible‑agent and exact‑step accuracy across all evaluated backbones, and the highest failure‑mode accuracy among taxonomy‑aligned methods on Pro. Additional image and video trajectory experiments demonstrate its applicability to multimodal failure attribution.
Ablation studies confirm the contributions of segmentation, the event dependency graph, and the failure propagation graph. Using DeFA’s diagnostic feedback to evolve skills in Trace2Skill improves downstream task accuracy by 6–15 percentage points, showing that the diagnoses can also support agent improvement on subsequent tasks.
Review