Claude Code Subagents: Your AI Team, Explained Simply

Claude Code subagents are specialized AI assistants that your main Claude spawns to handle side tasks — each one works in its own separate context window and brings back only the summary. You create one in five minutes with the /agents command, no code required. I'm Evgeny, and this is the feature that turned my single AI assistant into a small department.

What are agents in Claude Code, anyway?

An agent is simply Claude with a job description. Your main conversation is the boss; a subagent is an employee the boss hires for one task: 'go research this', 'go review that code'. The employee gets its own instructions (called a system prompt), its own set of allowed tools, and — this is the magic — its own working memory. It does the messy digging in its own room and returns with a one-page report instead of dumping three hundred files on the boss's desk.

What are subagents in AI, in plain words?

The 'sub' just means subordinate: an agent spawned by another agent. Here's why it matters even if you never write code. Claude's working memory — the context window — is like a desk. Every file it reads piles onto the desk, and a cluttered desk makes a sloppy worker. A subagent rents a second desk: all the search results, logs, and dead ends pile up there, and only the conclusion comes back to yours. One sensible rule from the docs: subagents can't spawn other subagents, so your office never turns into infinite bureaucracy.

Why is everyone so obsessed with this?

Because delegation is built in, not bolted on. Claude Code ships with built-in subagents it uses automatically: Explore (a read-only researcher that searches your project without changing anything), Plan (gathers context when you're in plan mode), and general-purpose (for tasks that mix exploration with changes). You've probably benefited from them without noticing. The obsession kicks in when people discover the next level: custom subagents — your own specialists with your own rules, reusable across every project. That's also why 'Claude Code orchestration' became such a buzzword: one boss, many workers, your instructions.

How to create your first subagent in 5 minutes

Under the hood, a subagent is just a Markdown file with a small header — honestly the least scary 'configuration' in the business. Project-specific ones live in .claude/agents/, personal ones in ~/.claude/agents/. One caveat from the docs: files you edit directly on disk load at session start, so restart the session — agents created through /agents work immediately.

---
name: code-reviewer
description: Reviews code for quality and best practices
tools: Read, Glob, Grep
model: sonnet
---

You are a friendly reviewer. Check clarity, naming,
and bugs. Return a short prioritized list.

How to use subagents — and when Claude delegates on its own

The 'description' field is the whole hiring policy: Claude reads it to decide when to hand a task over. Write it like a when-to-call-me note — 'Use proactively after code changes' works better than 'a code reviewer'. You can always force the issue by naming the agent in your request: say to Claude, 'use the code-reviewer subagent on yesterday's changes'. And in the /agents panel, the Running tab shows live subagents, so you can watch your team actually working.

Start with one employee, not ten. A single read-only reviewer or explainer that you actually use beats a GitHub list of 'awesome subagents' you never touch. When you catch yourself giving the same instructions for the third time — that's not a prompt anymore, that's a job opening.