Subagents & Power Moves

What Are Subagents?

Until now, you’ve been working with Claude as a single assistant — one conversation, one task at a time. But some jobs are too big or too specialized for one agent to handle alone.

Subagents are specialized AI assistants that Claude can delegate work to. Think of it like a manager who hires specialists: instead of doing everything themselves, they assign the security review to a security expert, the documentation to a technical writer, and the testing to a QA engineer.

Claude Code comes with built-in subagents and lets you create your own custom ones.

Built-In Subagents

Claude Code ships with two powerful built-in subagents:

Explore — A research agent that reads files, searches codebases, and gathers information without modifying anything. It’s read-only, which makes it safe for investigation.

Plan — A planning agent that analyzes a problem, considers trade-offs, and creates a step-by-step implementation plan. It doesn’t write code — it strategizes.

You can invoke them directly:

Use the Explore agent to map out every API endpoint in this project. List each endpoint, its HTTP method, what it does, and which file it's defined in.

Use the Plan agent to design how we should add multi-language support to this app. Consider the file structure, translation approach, and which parts of the UI need changes. Don't write any code yet.

Custom Agents

The real power comes from creating your own agents. Custom agents live in the .claude/agents/ directory of your project. Each one is a Markdown file with a name, description, and detailed instructions.

Here’s what a custom agent looks like:

Now you can use this agent anytime:

Ask the security-reviewer agent to review the src/auth/ directory and all API routes for vulnerabilities.

Pro Tip

Notice that the security reviewer only has Read, Grep, Glob, WebSearch as allowed tools — it cannot edit files. This is by design. A reviewer should find problems, not fix them. You decide what to fix.

Parallel Sessions with Worktrees

Here’s where things get really interesting. Normally, Claude works on one thing at a time. But what if you need to build three features simultaneously?

Git worktrees let Claude run multiple sessions in parallel, each working on a separate copy of your code. Think of it like having three desks in your office — each one has a different project spread out, and you can work on any of them without shuffling papers.

While that session works on authentication, you can open another terminal and start a parallel session:

Both sessions run at the same time, working on separate branches. When they’re done, you merge the results. It’s like having two developers working in parallel — except both of them are Claude.

Batch Processing with /batch

What if you need to make the same change across dozens of files? The /batch command processes multiple items in sequence or parallel:

/batch For each React component in src/components/, add TypeScript prop types and convert any class components to functional components with hooks.

This is the “mass operation” tool. Instead of asking Claude to update files one by one, you describe the pattern once, and Claude applies it everywhere.

Other examples:

/batch For each utility function in src/utils/, generate a comprehensive test file with at least 5 test cases covering edge cases. Save tests to src/__tests__/.

Non-Interactive Mode

Sometimes you want Claude to run a task without any conversation — just execute and report back. That’s what non-interactive mode is for:

The -p flag stands for “prompt.” Claude runs the task, prints the result, and exits. No interactive session, no back-and-forth. This is essential for automation — you can use it in scripts, cron jobs, and CI/CD pipelines.

Pro Tip

Combine non-interactive mode with scheduled tasks for powerful automation. For example: claude -p "Check for security vulnerabilities and email me a report" running every night at midnight via cron.

The Fan-Out Pattern

The most advanced technique combines everything you’ve learned: fan-out. This is when you split a large task into subtasks, assign each to a subagent or worktree, run them in parallel, then combine the results.

Here’s the pattern:

I need to add a complete user settings page with: profile editing, notification preferences, billing management, and theme customization. Use the Plan agent to create a plan, then spin up separate worktrees for each section. When all are done, merge them and run a review.

Try It Yourself