Multi-agent failures come in dependent clusters, and modeling the dependency improves attribution

arXiv 2609.01360·medium signalXBlueskyLinkedInCopy link

EDGE builds an error dependency graph from observed error events in a multi-agent run, then validates a causal subset through counterfactual rollout before handing the graph to a two-stage LLM-as-judge detector. It improves category-level multi-error attribution across most models and settings on the TRAIL and MAST benchmarks, and the gain holds across different prompting strategies. The framing matters more than the framework: single root-cause attribution is the wrong model for agent failures that arrive as several related errors.