What they did
The authors developed a two-stage failure-attribution system. It first transforms a failed multi-agent execution trace into a structured graph representing relevant components, actions, and dependencies, then uses an agent-directed traversal process to search for the critical error.
They evaluated this approach on Who&When, a benchmark for locating the source and timing of failures in multi-agent systems, and compared different trace representations and navigation strategies across multiple language models.
Key findings
- Richer, structured representations of execution traces consistently improved failure attribution across the models tested.
- Adaptive graph construction produced better results than less structured ways of presenting traces to the diagnosing model.
- Combining adaptive construction with agent-directed graph traversal performed best and established a new state of the art on Who&When.
- The results indicate that diagnostic accuracy depends on both the underlying model and how execution evidence is organized and explored.
Why it matters
Multi-agent failures can span several components and interactions, making raw logs difficult to inspect manually or with a single-pass model. A graph-based representation could make automated debugging more targeted by exposing causal or influential dependencies and allowing diagnostic agents to inspect only the most relevant parts of a run.
Caveats
The reported evidence centers on the Who&When benchmark, so it remains unclear how well the method transfers to noisy production traces, unseen agent architectures, or failures with several interacting causes. The provided abstract does not report absolute scores, effect sizes, graph-construction costs, or added inference latency.