Claude Code Subagents Parallel Work Guide 2026

Three research agents running in parallel. Codebase analysis dropped from 10 minutes to 3. I'm a solo developer, but it felt like I suddenly had a team.

A Claude Code subagent is a specialized AI assistant that runs in its own context window inside your main Claude session. Each one gets its own system prompt, tool access, and model configuration. You can run up to 10 simultaneously.

This is Part 3 of the "Claude Code Power User Guide" series. Part 1 covered MCP, Part 2 covered Hooks, and this final part is about using subagents and agent teams to parallelize your workflow.

Built-in Agents You Already Have

Claude Code ships with six built-in agents. The Explore agent runs on Haiku for fast, read-only codebase search. No risk of accidental file modifications. The Plan agent handles research in plan mode, and the General-purpose agent has full tool access for complex multi-step tasks.

For small projects, built-ins are enough. Explore fires automatically when you need code analysis, General-purpose handles complex refactoring. But as the codebase grew, I wanted more control. I wanted a code-reviewer agent that runs automatically after every code change. I wanted tests and docs updated in parallel, not sequentially.

Three Ways to Create Custom Agents

The easiest path is the /agents command. Claude walks you through an interactive creation flow and drops a markdown file in .claude/agents/.

You can also write the markdown file directly. The YAML frontmatter takes name, description, tools, model, and permissionMode. A code reviewer agent, for example, gets Read, Grep, Glob, and Bash only. Read-only by design. Model set to sonnet for speed.

---
name: code-reviewer
description: "Reviews code for bugs, performance, and security issues"
tools: [Read, Grep, Glob, Bash]
model: sonnet
permissionMode: plan
---

For one-off use, pass agents inline via CLI: claude --agents '{"reviewer": {"description": "...", "prompt": "..."}}'.

Invocation works four ways. Write a good description and Claude auto-delegates matching tasks. Ask in natural language: "use the code-reviewer subagent." Direct mention with @"code-reviewer (agent)". Or lock an entire session to one agent with claude --agent code-reviewer.

Foreground, Background, and Agent Teams

Subagents default to foreground execution. Blocking. Permission prompts pass through the parent. Switch to background (background: true or Ctrl+B) and multiple agents run concurrently with pre-approved permissions.

My actual pattern: when refactoring, I spin up three background research agents. One explores related files. One analyzes dependencies. One checks test coverage. Running in parallel cuts codebase analysis from 10 minutes to 3.

Agent Teams, introduced in February 2026 with Opus 4.6, go further. Subagents report only to their parent. Agent Teams send messages to each other. TeamCreate builds the team. SendMessage enables inter-agent communication. A shared TaskList tracks progress across all agents. Multiple independent sessions, coordinating.

The memory: project setting matters here. It lets agents remember coding patterns, architecture decisions, and recurring issues across sessions. I turned this on for my code-reviewer agent and review quality improved noticeably. Patterns flagged in one session get caught consistently in the next.

Practical Configuration Patterns

The isolation: worktree setting runs an agent in an isolated git worktree. Your main branch stays untouched while the agent experiments with changes.

Combine PreToolUse hooks for finer control. A DB query validator agent can automatically check that SQL is read-only before Bash execution. The effort parameter lets you tune reasoning depth: low for quick searches, max for complex debugging.

For a solo developer, this changes the math. Writing code, reviewing, testing, documenting: you do it all alone. Build an agent for each role and four of them work simultaneously. I keep writer, reviewer, tester, and architect agents configured per project, dispatching them in parallel before every work session.

You develop alone, but you don't have to work alone. Agents become the team.

---

References

• Claude Code Subagents Documentation - Anthropic
• Agent Teams Documentation - Anthropic
• Claude Code - Anthropic