DCFA: Causal-inspired Failure Attribution for Multi-Agent Systems

DCFA is a training-free framework designed to identify the 'decisive error' in large language model multi-agent systems. It addresses shallow attribution and contextual degradation by tracing interactions to find the earliest action whose correction would reverse a system failure.