Automated Root-Cause Attribution as a Search Problem for Long-Horizon Agents

This research addresses the difficulty of diagnosing failures in long-horizon agent execution logs by treating root-cause attribution (RCA) as a search problem. It aims to prevent LLM judges from settling on early, incorrect diagnoses by better navigating sparse, distributed evidence in large traces.