一句话：提出免运行手册的多智能体根因分析框架，借助工具质量管理提升跨模态诊断的准确率与可迁移性。
根因分析（RCA）对云服务可靠性至关重要，需要关联 traces、logs、metrics 等多模态可观测数据以定位故障成因。现有基于 LLM 的 RCA 方法存在两大局限：过度依赖系统特有的产物（如 runbook），导致迁移性差；以及难以保证 LLM 生成的跨模态诊断工具的质量与复用。本文提出 OpsLens，一个将按模态分工的多智能体协作与以质量为中心的工具管理相结合的 RCA 框架，其中包含反馈驱动的工具精化机制，以及带记分板的工具图（tool graph with scoreboard）。在包含三个生产系统 335 个真实故障的 OpenRCA 基准上，OpsLens 取得 0.80 的得分率，优于基线 1.79–4.75 倍。更重要的是，OpsLens 自 2025 年 6 月起在华为 Platform A 上测试，在两个工业数据集上达到 80% 以上的 RCA 准确率，并将诊断时间从 1 小时以上缩短到 5 分钟以内。
Root Cause Analysis (RCA) is critical for cloud service reliability, requiring correlation of multi-modal observability data (i.e., traces, logs, and metrics) to identify failure causes. Current LLM-based RCA approaches face two key limitations: over-reliance on system-specific artifacts (e.g., runbooks), limiting transferability; and difficulty ensuring the quality and reuse of LLM-generated diagnostic tools across modalities. To address these issues, we propose OpsLens, an RCA framework that combines modality-specialized multi-agent collaboration with quality-focused tool management , including feedback-driven refinement and a tool graph with scoreboard. On OpenRCA (335 real-world incidents from three production systems), OpsLens achieves a score rate of 0.80, outperforming baselines by 1.79–4.75×. More importantly, OpsLens has been tested on Platform A in Huawei since June 2025, achieving over 80% RCA accuracy across two industrial datasets and reducing diagnosis time from over 1 hour to under 5 minutes.
@inproceedings{paperbot12218,
title = {OpsLens: Runbook-Free Root Cause Analysis via Multi-agent Collaboration and Tool Management},
author = {Haiyu Huang and Jiewei Lyu and Gou Tan and Pengfei Chen and Zhihan Jiang and Guangba Yu and Penghui Mi and Fuqiang Wang and Jian Han and Chongkang Tan and Michael R. Lyu},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834501}
}