Knowledge-Based Zero-Replay Debugging of Multi-Agent LLM Traces

The article presents a novel approach to debugging multi-agent large language model (LLM) systems by framing trace debugging as a knowledge-based decision-support problem. It introduces a zero-replay counterfactual-effect prediction method that utilizes a structured event knowledge graph to identify high-impact events without the need for costly replay, achieving a significant improvement in trace localization (Branch Recall@5 from 0.73 to 0.93) using a gradient-boosted predictor. This method offers practitioners a cost-efficient and auditable solution for enhancing AI reliability, particularly in scenarios where traditional counterfactual replay is infeasible.