万能エージェント1体をやめる — GitHub Copilot のカスタムエージェントと自作 MCP サーバーで「分業」を設計するThis article details a multi-agent architecture for GitHub Copilot where one…

A recent post on Zenn walks through a hands-on attempt to stop relying on a single, do-everything AI assistant and instead split the work across several GitHub Copilot custom agents. The topic matters because it speaks to a broader question facing teams adopting agentic coding tools: as tasks grow more complex, is it better to hand everything to one general-purpose "persona," or to design a division of labor where narrowly scoped agents each own a step? The author frames the article for readers who want to push custom agents beyond a single convenient persona, and for those who want to build their own MCP server to connect Copilot to in-house business systems.

The concrete goal is automating updates to a knowledge base—specifically, help-center articles. Instead of asking one agent to carry out the entire workflow, the design places an orchestrator agent at the top, which delegates to three specialist agents. This "one supervisor plus three specialists" structure is the core idea: the orchestrator coordinates the overall flow and hands off discrete responsibilities, while each specialist focuses on a narrower part of the update process. The stated rationale is that separating concerns can make each agent's behavior easier to reason about and maintain than packing every instruction into a single prompt or configuration.

To let those agents act on systems outside the editor, the author operates external SaaS through a custom-built MCP server. The Model Context Protocol is an open standard, introduced by Anthropic in late 2024 and since adopted across a range of editors and AI tools, that defines a consistent way for AI applications to connect to external data and services. In practice, an MCP server exposes tools and resources that a client—here, GitHub Copilot—can call, which means teams can wire an agent to their own ticketing systems, content platforms, or databases without inventing a bespoke integration format each time. Building the server in-house, rather than using an off-the-shelf connector, appears to be a deliberate choice so the agents can operate the specific help-center SaaS the team uses.

The work is grounded in VS Code's GitHub Copilot Chat and its custom agent capability. Custom agents let users define specialized configurations—persona, instructions, and available tools—so that Copilot behaves differently depending on the task at hand. Combined with MCP, this gives the setup two complementary layers: custom agents shape how each agent reasons and what role it plays, while the MCP server governs what real-world actions it can take. The article positions this combination as the mechanism that makes a agent">multi-agent division of labor feasible in a normal editor-based workflow.

Notably, the post does not present the approach as a finished solution. Alongside the design and implementation, the author records what became clear only after running the system, as well as problems that remain unresolved. That candid framing is useful context: agent">multi-agent orchestration is still an emerging practice, and handoffs between agents, error handling, and the reliability of tool calls are common friction points that a single-agent setup does not surface in the same way. Readers evaluating a similar architecture should likely treat it as an experiment to learn from rather than a turnkey blueprint.

For background, the pattern reflects a wider industry shift. Vendors including GitHub, Anthropic, and others have been expanding agentic features and standardizing tool access, and MCP in particular has gained traction as a common connective layer across competing assistants. agent">Multi-agent orchestration—where a coordinator delegates to specialists—is a recurring theme in frameworks and research aimed at handling tasks too large or varied for one prompt. Against that backdrop, this article's contribution is a specific, real-world example: applying the supervisor-and-specialists model to routine documentation maintenance, and connecting it to production SaaS through a self-authored MCP server. It offers a template teams can adapt, along with an honest account of where the approach still needs work.