How Large‑Model Agents Transform AIOps: From Automation to Self‑Healing Operations

At the 22nd GOPS Global Operations Conference, ByteDance STE’s Ren Zhiqiang presented the practical use of large‑model agents in AIOps, highlighting their ability to automate repetitive SRE tasks, improve human efficiency, and enable autonomous problem solving.

Why use large‑model agents? Large models excel at conversational interaction; agents combine multi‑step dialogue, planning, reflection, and tool usage to achieve goal‑driven autonomy, significantly boosting AI capabilities for AIOps scenarios such as inspections, fault detection, and knowledge queries.

Agent deployment modes include single‑agent, multi‑agent, and human‑in‑the‑loop approaches. Single agents handle complex tasks via task decomposition and divide‑and‑conquer. Multi‑agents collaborate through defined roles, knowledge, and tools, enhancing efficiency and innovation. Human interaction supplements agents when model reasoning or knowledge is insufficient.

Key AIOps enhancements provided by agents:

Anomaly detection : Transform multimodal data into unified vectors and use pre‑trained models or agent‑driven prompts for comprehensive detection.

Fault diagnosis : Agents plan and execute investigation workflows, leveraging data retrieval, anomaly checks, and root‑cause analysis tools.

Fault repair : Agents drive code execution and tool usage to mitigate, stop loss, and even achieve self‑healing.

Alert convergence : Agents parse alerts, apply rule‑based and semantic merging to reduce noise.

ChatOps : Direct intent recognition and tool invocation enhance conversational operations and knowledge‑base retrieval (RAG).

Agent architecture consists of model (brain), planning, memory, tools, environment interaction, and role definition. Planning uses reflection methods such as ReAct, Self‑Ask, ReWoo, Tree‑of‑Thought, or Graph‑of‑Thought, often guided by SOPs. Memory combines long‑term retrieval via RAG and short‑term prompt context, while tools are accessed through function‑calling capabilities.

Multi‑agent collaboration faces challenges in role differentiation and coordination; practical solutions include a moderator agent, limited collaboration rounds, and simplified interaction strategies.

Practical fault‑triage workflow (range definition → investigation → summary) reduces omission, reuses historical cases, and accelerates analysis using parallel reasoning (ReAct, ToT, GoT). Summaries are generated via RAG‑augmented generation, stored, and reviewed to enrich knowledge bases.

In on‑call scenarios, agents enable intelligent Q&A to intercept issues early, employing multi‑source retrieval, fine‑tuned embeddings, reflective result optimization, multimodal knowledge parsing, and nl2sql‑style data queries.

Future outlook emphasizes scaling agent autonomy, fine‑tuning models for specific domains, reinforcement‑learning‑based role collaboration, and designing multimodal models for comprehensive anomaly analysis.