一句话：提出去中心化的智能体分诊框架 Comfey，让各团队轻量 agent 通过共享路由表协作，实现生产环境事故的高效分诊。
大规模云基础设施中事故频发，若不能及时分诊与缓解将造成严重服务中断。现有自动分诊系统假设事故报告完整、各 agent 质量一致，但生产环境中可观测性数据碎片化、团队成熟度参差不齐，这些假设难以成立。本文提出 Comfey：一个面向真实运维场景的去中心化 agentic 分诊框架。Azure 中每个团队部署一个轻量 Comfey triage agent，它分析本领域的特定数据，判断是否接受或拒绝某起事故，并通过共享路由表与其他 agent 协作完成路由。Comfey 已在 Azure 生产环境持续运行 22 个月，累计分诊约 19,500 起事故，准确率达 91.5%（由值班工程师验证）；相比人工流程，平均分诊时间缩短 6 倍，缓解时间缩短 4.3 倍，且部署配置要求极低。
Incidents occur frequently in large-scale cloud infrastructures and can cause significant service disruptions if not triaged and mitigated promptly. Existing automated triage systems rely on complete incident reports or uniform agent quality, which are often infeasible in production due to fragmented observability and heterogeneous team maturity. We present Comfey, a decentralized agentic triage framework designed for this operational reality. Each team in Azure deploys a lightweight Comfey triage agent that analyzes domain-specific data, decides whether to accept or reject an incident, and collaborates with other agents via a shared routing table. Comfey has been running in Azure production for 22 months, where it has triaged around 19,500 incidents with 91.5% accuracy (validated by on-call engineers). It reduces average triage time by 6 times and mitigation time by 4.3 times compared to manual processes, with minimal setup requirements.
@inproceedings{YaoJMVDHBLC26,
title = {An Agentic Framework for Triaging Incidents in Production Cloud Infrastructure},
author = {Yuhan Yao and Yuxuan Jiang and Minghua Ma and Madhura Vaidya and Jieren Deng and Yigong Hu and Chetan Bansal and Ze Li and Murali Chintalapati},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805228}
}