AI駆動開発 #3: セッション×サブエージェントで作るAIプロジェクト体制 — 19体から5〜7体へ絞って分かった「任せる仕事・任せない仕事」This article documents a hands-on experiment reducing an AI agent team from 19…

The third installment in a series on AI-driven development tackles a practical question that many teams face once they start automating with large language models: how many specialized AI agents does a project actually need? Written for the Qiita community and centered on Anthropic's Claude, the article describes an experiment in which the author scaled a working team of AI agents down from 19 to roughly five to seven, and uses that reduction to draw conclusions about which tasks are safe to delegate and which still benefit from human oversight.

The core architecture combines two building blocks that Claude Code users will recognize: sessions and sub-agents. A session is the persistent working context in which a conversation and its accumulated state unfold, while a sub-agent is a scoped worker with its own instructions, its own context window, and often a restricted set of tools. In this model a primary or orchestrating agent breaks a larger goal into smaller pieces and hands them to sub-agents, each of which reports back without polluting the main conversation's context. The appeal is clear: separating concerns keeps the top-level reasoning focused and can reduce the token bloat that tends to degrade performance in long-running tasks.

The headline finding is that more agents did not translate into better results. The initial roster of 19 agents, apparently organized around narrowly defined roles, proved difficult to coordinate and maintain. Overlapping responsibilities, ambiguous handoffs, and the overhead of routing work between many specialists appear to have outweighed the benefits of fine-grained specialization. Consolidating to a smaller set of five to seven broader agents is presented as the more sustainable configuration, where each agent covers a coherent area of responsibility rather than a single micro-task.

Much of the article's practical value lies in its attempt to codify what should and should not be handed off. Tasks described as good candidates for delegation tend to be well-bounded and verifiable, the kind of work where the definition of done is explicit and the output can be checked mechanically or by tests. Work that is retained under closer human control, by contrast, involves ambiguous requirements, architectural decisions with long-term consequences, or judgment calls where an incorrect autonomous action would be costly to reverse. This framing echoes a broader theme in the current agent-tooling conversation: autonomy is most reliable when the surrounding guardrails, acceptance criteria, and feedback loops are strong.

The discussion sits within a fast-moving ecosystem. Anthropic has been building out Claude Code as a terminal-based coding assistant, and sub-agents are among several features, alongside memory files, custom commands, and hooks, aimed at making longer and more structured workflows manageable. Comparable orchestration ideas appear in frameworks such as LangGraph, Microsoft's AutoGen, CrewAI, and OpenAI's Assistants and Agents tooling, all of which grapple with the same design tension between a single capable agent and a coordinated team of narrower ones. The recurring lesson across these tools is that multi-agent designs add coordination cost, so the number and granularity of agents should be justified by a real gain in reliability or clarity.

For readers new to these concepts, a few prerequisites help. The context window is the finite amount of text a model can consider at once, which is why isolating sub-agent work matters. Prompt engineering, the careful specification of each agent's role, constraints, and output format, remains central, because a sub-agent is only as dependable as the instructions and tools it is given. And the orchestrator pattern, in which one agent plans and delegates while others execute, is the structural backbone that the article's session-plus-sub-agent setup embodies.

As a single practitioner's account rather than a controlled study, the results should be read as informed guidance rather than definitive benchmarks; the optimal number of agents is likely to vary with project size, domain, and tolerance for error. Still, the underlying principle is a useful counterweight to the assumption that adding more specialized agents automatically improves an AI-assisted workflow. The more durable takeaway is to match delegation to task verifiability and to keep the human in the loop where consequences are hard to undo.