Claude Sub-Agents
View SourceClaude includes a powerful sub-agent system that lets you create specialized AI assistants for specific tasks in your project.
Documentation
For complete documentation on Claude Code's sub-agent system:
- Official Sub-Agents Guide - Complete guide with examples
What Claude Includes
When you install Claude, you automatically get:
- Meta Agent - A built-in agent that helps you create new sub-agents following best practices
- Usage Rules Integration - Automatic injection of library best practices into your sub-agents
- Project-specific configuration - Sub-agents configured in
.claude.exs
Configuration
Sub-agents are configured in .claude.exs:
%{
subagents: [
%{
name: "test-expert",
role: "ExUnit testing specialist",
instructions: "You excel at writing comprehensive test suites...",
usage_rules: ["usage_rules:elixir", "usage_rules:otp"]
}
]
}Creating Sub-Agents
The easiest way to create a new sub-agent is to ask Claude:
Create a sub-agent for handling GraphQL queriesThe Meta Agent will automatically:
- Generate a complete sub-agent configuration
- Choose appropriate tools and permissions
- Include relevant usage rules from your dependencies
- Add it to your
.claude.exsfile
Manual Configuration
You can also manually add sub-agents to .claude.exs:
%{
subagents: [
%{
name: "database-specialist",
role: "Database and Ecto expert",
instructions: """
You are an expert in Ecto and database operations.
When invoked, you should:
1. Analyze the database schema
2. Write efficient queries
3. Handle migrations properly
""",
tools: [:read, :write, :edit, :bash, :grep],
usage_rules: ["usage_rules:ecto", "usage_rules:elixir"]
}
]
}Usage Rules Integration
Sub-agents can automatically include best practices from your dependencies through usage rules:
usage_rules:elixir- Elixir language best practicesusage_rules:otp- OTP patterns and practicesusage_rules:phoenix- Phoenix framework patternsusage_rules:ecto- Database and query best practices- And any other dependencies that provide usage rules
Important Notes
- Clean Slate: Sub-agents start fresh each time - they have no memory of previous interactions
- Tool Inheritance: When tools are omitted, sub-agents inherit all available tools
- Performance: Keep instructions focused and context minimal for best performance
- Delegation: The main Claude agent automatically delegates to sub-agents based on their role descriptions
For more details on sub-agent architecture, delegation patterns, and advanced usage, see the official documentation.