The paper addresses automated root cause analysis (RCA) with large language models. Existing approaches either use general-purpose agents (like Codex or Claude Code) or build specialized RCA agents from scratch. The authors find that modern general agents often outperform custom-built RCA agents, but still fall short of production needs. They argue the gap stems from the external harness, the layer that adapts the agent to the task.
They introduce OpsHarness, a self-evolving harness that accumulates system-specific experience from past diagnoses. It uses a data plane with layered operational knowledge and an idea-card tool library, and a control plane that coordinates setup, diagnosis, evolution, and verification. Evolution contrasts successful and failed diagnosis trajectories, and updates are admitted only after dual-gate verification to prevent overfitting. Evaluated on two public benchmarks and an industrial deployment, OpsHarness achieves 59.0% top-1 accuracy, a 63.4% improvement over a bare general agent and 4.02x improvement over baseline RCA agents.