AIの専門家4人にAzure障害を調査させる「インシデント調査チーム」を作った — GitHub Copilot のカスタムエージェント×スキル実践A practical walkthrough showing how to build a multi-agent incident…

GitHub Copilot is increasingly being used for tasks well beyond code completion, and a recent walkthrough published on Qiita demonstrates how its custom agent and skills capabilities can be assembled into a coordinated "incident investigation team" for analyzing Azure outages. The example is worth attention because incident response—triaging alerts, correlating logs, and pinpointing root causes—remains one of the most time-consuming and stressful parts of operating cloud infrastructure, and it is a domain where structured automation could meaningfully reduce operational toil.

The central idea is to define four specialized AI roles, each responsible for a distinct slice of an investigation, and to have them collaborate rather than depend on a single general-purpose assistant. Splitting the work this way mirrors how human incident bridges often operate, where different engineers focus on networking, application logs, infrastructure metrics, and communications. By narrowing each agent's scope, the author appears to aim for more focused reasoning and clearer division of responsibility, which can make the overall output easier to audit than a monolithic prompt that tries to do everything at once.

Two GitHub Copilot features underpin the setup. Custom agents let developers configure purpose-built assistants with their own instructions, personas, and tool access, so an agent can be tuned to behave like, for instance, a network specialist or a log analyst. Custom skills extend what those agents can actually do by connecting them to external capabilities and data sources. In practice this means an agent is not limited to reasoning over text it was given, but can invoke defined actions to fetch telemetry, query services, or run analyses relevant to an Azure environment.

The article's tags point to the Model Context Protocol, or MCP, as a key enabling layer. MCP is an open standard, originally introduced by Anthropic and since adopted across a growing number of tools, that defines how AI assistants connect to external systems through servers exposing tools and resources. An MCP server can wrap an API—such as Azure's monitoring and management endpoints—so that a Copilot agent can request logs, metrics, or resource status in a consistent, structured way. This is the mechanism that likely lets the specialized agents gather real signals about an outage rather than speculate, grounding their conclusions in actual system data.

Coordinating multiple agents introduces its own design considerations. The walkthrough describes the agents dividing the investigation and working together, which suggests some form of orchestration where findings are passed between roles or aggregated into a combined assessment. Multi-agent systems can improve coverage and parallelize analysis, but they also raise questions about how conflicting conclusions are reconciled, how much context each agent shares, and how to prevent redundant or contradictory recommendations. Readers evaluating the approach should treat the output as a draft assessment to be verified by engineers rather than an authoritative root-cause determination.

This project sits within a broader industry shift toward agentic AI and AIOps. Vendors including Microsoft, with its Copilot lineup and Azure-native monitoring, have been moving toward AI-assisted operations, and site reliability engineering teams have long used runbooks and automation to standardize incident handling. The multi-agent pattern shown here can be seen as a natural extension of that trajectory, applying large language models to the interpretation step that traditionally required experienced on-call engineers. Comparable frameworks such as Microsoft's AutoGen, LangGraph, and CrewAI have popularized similar role-based agent collaboration, so the concepts should feel familiar to teams already experimenting in this space.

For readers who want to try something similar, a few prerequisites are worth noting. Familiarity with GitHub Copilot's configuration for custom agents and skills is essential, as is a working understanding of MCP servers and how to expose Azure data securely. Access controls matter, since agents that can query production telemetry need appropriately scoped credentials. It is also prudent to test such a system against past incidents with known outcomes before relying on it during a live outage.

Ultimately, the walkthrough is best read as a concrete, reproducible pattern rather than a finished product. It illustrates how currently available Copilot features, combined with MCP-based tooling, can be composed into a collaborative investigation workflow, while leaving the harder questions of accuracy, trust, and human oversight to the teams that adopt it.