When One AI Isn't Enough, Attach a Team of Role-Based "Assistants" with /agents

When building a big feature, have you had one conversation become hopelessly scattered because you had to search code and write tests at the same time as the main work? Just as one person doing everything is slow and confusing, leaving everything to one AI is too. The way to attach a "team" then is /agents. You divide side tasks among small Claudes (subagents) that take on specific roles.

Definition

/agents is a command that manages assistant worker (subagent) settings. A subagent is a "small Claude" that takes on a specific role and handles side tasks in your place.

For example, if you set roles like "code review specialist" or "research specialist," the main work proceeds in the main body while side jobs are entrusted to the specialist worker. In human terms, it is attaching role-based team members to a big project. You can also run multiple workers at once, finishing big work quickly.

How to Use (by difficulty)

Basics — Opening the Management Screen

Type this into the chat box.

/agents

Once you enter it, a screen for creating and managing subagents appears. Here you can create a new worker, modify an existing worker's role, or delete one you no longer need.

Advanced — Creating Them by Role

'코드리뷰 전담' 에이전트 만들기

Instructing it like this in words lets you create an agent that takes on a specific role. For repeatedly needed work like code review, keeping a dedicated worker separately handles just that role cleanly without disturbing the main work.

Deeper — Running Multiple Workers in Parallel

리서치 에이전트와 테스트 에이전트 동시 운용

You can run multiple specialist workers at once. Proceeding in parallel, with one searching and organizing code while the other writes tests, makes big work much faster.

Common Pitfall — Best for Side and Read Tasks

Using /agents for the first time, it is easy to think "if I chop every task finely and hand it to agents, it'll be fast," but that is not so. Subagents are best suited to tasks with a clear role and a defined output format, especially read-focused side tasks.

For example, exploration and research tasks like "find every place that calls this function" or "summarize what this subsystem (submodule) does" are perfect. Such tasks confine the exploration cost inside the agent, keeping the main conversation clean. Conversely, blindly offloading even the writing of core logic makes management more complex, so start by entrusting side tasks.

Real-World Example

While building a big feature, I had to grasp the code structure and also write tests. Doing it all in the main conversation made the context scattered. So I kept a "research specialist" and a "test-writing specialist" agent separately.

리서치 에이전트와 테스트 에이전트 동시 운용

While the research agent organized the existing code structure, the test agent wrote tests, and I could focus only on the main work. Dividing in parallel the work I used to do all alone was much faster and my head was less cluttered.

Using It Further

• Role-based agents: Keep specialist workers separately, like "code review specialist" or "research specialist."
• Parallel division: Run multiple agents at once to finish big work quickly.
• How to create: Create them on the /agents screen, or write them directly as a Markdown (.md) file.
• Combine with large-scale work: Bundle with /batch (batch — a command that divides and processes the codebase at once) to divide up big work.

Tip: A reviewer-role agent works better the more you give it a clear checklist, limited tools, and a defined output format. Define the role narrowly and clearly, like "check only these items and report the result in this format." Collections of settings to reference before creating are also published, so it is good to look them over.

How Others Use It

• Reviewer agents are especially suited to a clear checklist, limited tools, and a defined output format. — Claude Code Subagents (Builder.io)
• Optimal for read-focused research like "find all callers of a function" or "summarize a subsystem's behavior." — Best practices for subagents (PubNub)
• A collection of 100+ ready-to-use subagent settings (for reference before creating). — awesome-claude-code-subagents (VoltAgent)

Summary

The core fits in one sentence. Do not do big work alone; divide it among specialist workers. Create role-based subagents with /agents to entrust side and research tasks, and running multiple workers in parallel speeds up big work. The clearer the role and output format you give, the better it works.

Based on: Claude Code v2.1.154 (2026.05)