VERDICT: Verification-based Diagnosis via Causal Topology for Multi-Agent Failure Attribution

Figuring out which agent caused a multi-agent system (MAS) to fail and exactly when it happened is still a major challenge. Current methods usually ask an LLM to review execution logs and guess the culprit, but they don't provide formal guarantees or estimates of the model's confidence. The real issue is that these approaches rely on pattern matching rather than calculating what would have happened if a specific agent had acted differently. We developed VERDICT (VERification-based DIagnosis via Causal Topology), a new framework that brings structured reasoning to this problem. Instead of trying to learn causal relationships from scratch, VERDICT extracts a formal model directly from the system's execution protocol. It introduces two main ideas: first, it leverages the system's known design to construct a causal graph, thereby avoiding the need for causal discovery from massive amounts of data; and second, it uses Halpern–Pearl (HP)-structured blame diagnostics with bootstrap stability intervals to provide reproducible diagnoses with explicit uncertainty metadata. Extensive experiments demonstrate that VERDICT achieves competitive results without incurring any LLM calls. Notably, on the Who&When benchmark, VERDICT significantly outperforms the published leading LLM-judge baseline, raising step-level accuracy from 16.5% to 29.3% and agent-level accuracy from 53.2% to 64.1%. These gains are achieved with reproducible causal diagnostics in less than 40 seconds on a single CPU.