一句话：提出以知识为锚的智能体化 AIOps 愿景：把运维知识作为一等资产，支撑 LLM 智能体进行跨服务故障响应。
本文是一篇立场论文，针对云系统规模化运维中的长期瓶颈：理解“正在发生什么、为何发生、下一步该做什么”所需的运维知识，仍然碎片化地散落于团队、工具、仪表盘、文档与个人经验之间。作者主张 LLM 时代能够带来阶跃式变革，但前提是把知识视为一等运维资产，而非工具链的附带产物。为此提出“以知识为锚的智能体化 AIOps”（Knowledge-Anchored Agentic AIOps）：以标准化、可治理的知识锚点为基础，将知识附着于跨服务层级的运维信号之上，形成全面、可行动、可共享且持续演化的跨服务知识织物，并据此构建 LLM 智能体。论文以事件管理为具体场景，介绍智能体化排障指南自动化框架 StepFly 作为该愿景的体现，并讨论其在 AIOps 其他场景及软件开发生命周期中的延伸意义。
Cloud Intelligence/AIOps applies AI/ML to help design, build, and operate complex cloud systems at scale effectively and efficiently. Despite substantial progress, a persistent bottleneck remains. The operational knowledge, needed to understand what is happening, why it is happening, and what to do next, largely remains fragmented across teams, tools, dashboards, documents, and individual experience. In this position paper, we argue that the LLM era enables a step-function change, but only if we treat knowledge as a first-class operational asset rather than an incidental byproduct of tooling. We propose Knowledge-Anchored Agentic AIOps focusing on LLM-based agents grounded in standardized, governed knowledge anchors that attach to operational signals across service hierarchy, forming a cross-service knowledge fabric that becomes comprehensive, actionable, shareable, and continuously evolving. Using incident management as a concrete scenario, StepFly, an agentic troubleshooting guide automation framework, is discussed as a manifestation of this vision in knowledge-anchored agentic incident response. We also discuss implications beyond incident management to more scenarios in AIOps and the software development lifecycle.
@inproceedings{ZhangLQLCSCZDKR26,
title = {Cloud Intelligence/AIOps 2.0: Knowledge-Anchored Agentic AIOps},
author = {Dongmei Zhang and Qingwei Lin and Si Qin and Liqun Li and Lianbin Chi and Dawei Song and Biao Cheng and Chaoyun Zhang and Yingnong Dang and Samia Khalid and Saravan Rajmohan and Sitaram Lanka},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3806091}
}