個人開発を「擬似会社」にした — Claude Codeでマルチエージェント開発の仕組みを作った話 This post shows how one developer built a multi-agent 'pseudo-company' using Claude Code, …

A developer has published a detailed account of restructuring their solo software projects around a team of cooperating AI agents, using Anthropic's Claude Code to build what the author calls a "pseudo-company." The write-up, posted on the Japanese engineering community Zenn, matters because it extends a fast-growing practice—agentic coding—from a single assistant into a coordinated group of specialists, an approach that could help individual developers organize and scale work that would normally require a team.

The central premise is to stop treating Claude Code as one general-purpose helper and instead split its work across multiple agents, each with a clearly defined role. In the author's framing, these roles echo the structure of an actual organization: planning, implementation, review, and quality assurance, among others. Each agent is given its own instructions, scope, and responsibilities, so that a feature request flows through the system much as a ticket would move through departments at a company. The result, the author argues, is development that is more scalable and more systematically organized than a single, monolithic prompt-and-response loop.

Technically, this kind of setup builds on capabilities that Claude Code already exposes. The tool, which runs in the terminal and can read, write, and execute code in a project, supports configuration files such as CLAUDE.md that persist project context and conventions across sessions. It also offers a subagents mechanism that lets a primary agent delegate focused tasks to separate agent definitions, each with its own system prompt and tool permissions. Building a "pseudo-company" appears to lean on these primitives: defining role-specific agents, scripting how work is handed off between them, and using automation so that one agent's output becomes another's input. The post emphasizes workflow automation as a key benefit, suggesting that much of the value comes from the orchestration around the agents rather than any single model call.

The motivation reflects a broader tension in solo development. One person can only hold so much context at once, and switching between high-level planning, detailed coding, and careful review is cognitively expensive. By assigning these modes to distinct agents, a developer can, in principle, keep each agent narrowly focused and reduce the chance that a single conversation drifts or loses track of earlier decisions. Specialized agents with constrained roles are also easier to reason about and debug than one assistant asked to do everything.

The approach sits within a wider industry movement toward multi-agent systems. Frameworks such as Microsoft's AutoGen, CrewAI, and LangGraph have popularized patterns in which several language-model agents negotiate, delegate, and check one another's work, and Anthropic has itself written about multi-agent research systems in which a lead agent coordinates subagents working in parallel. Competing agentic coding tools, including Cursor, GitHub Copilot's agent features, and autonomous systems like Devin, illustrate how quickly the category is maturing. The "pseudo-company" metaphor is essentially a developer-friendly way of describing orchestration concepts that these tools and frameworks formalize.

It is worth noting the practical caveats that such setups tend to carry, even where the post is optimistic. Running multiple agents typically increases token consumption and cost, since each role may reprocess shared context. Coordination can introduce its own failures, such as agents duplicating effort, disagreeing, or propagating an early mistake downstream, and the quality of results still depends heavily on how prompts and handoffs are designed. Human oversight generally remains necessary, particularly for review steps where an automated "QA" agent may miss issues a person would catch.

For readers, the post functions less as a finished product and more as a pattern others can adapt. The takeaways—define clear roles, persist shared context, automate the handoffs, and let each agent specialize—are transferable to other agentic tools, not just Claude Code. As models grow more capable of sustained, tool-using work, accounts like this one suggest that the near-term advantage for individual developers may come less from any single model and more from the systems and processes built around it. Whether a "pseudo-company" of agents reliably outperforms a well-managed single assistant is still an open question, but the experiment reflects where a meaningful part of the developer community appears to be heading.