Subagent Configuration

Subagent Configuration

Configure custom subagents in Claude Code with specialized tools, models, and permissions for task-specific workflows.

Purpose

Subagents are pre-configured AI personalities that delegate specialized tasks. Each subagent operates in its own context window with custom system prompts, specific tool access, and tailored permission modes. This skill provides guidance on configuring subagent YAML frontmatter and system prompts effectively.

When to Use This Skill

Use this skill when:

• Creating new subagents for specialized tasks
• Configuring tool access for security or focus
• Selecting appropriate models for performance/cost trade-offs
• Setting permission modes for automation workflows
• Loading skills automatically when subagents start
• Understanding configuration field options and values

Configuration Overview

Subagents are defined in Markdown files with YAML frontmatter:

---
name: agent-name
description: When this agent should be invoked
tools: tool1, tool2, tool3    # Optional
model: sonnet                 # Optional
permissionMode: default       # Optional
skills: skill1, skill2        # Optional
---

System prompt content goes here.

File Locations

• Project subagents: .claude/agents/ (highest priority)
• User subagents: ~/.claude/agents/ (lower priority)
• Plugin agents: Provided by installed plugins

Required Fields

name (Required)

Unique identifier using lowercase letters, numbers, and hyphens only.

Format: lowercase-with-hyphens

Examples:

• code-reviewer
• test-runner
• data-scientist
• security-auditor

Invalid:

• Code_Reviewer (uppercase, underscores)
• test runner (spaces)
• my.agent (periods)

description (Required)

Natural language description of when Claude should invoke this subagent. This field is critical for automatic delegation.

Best practices:

• Be specific about when to use the agent
• Include trigger phrases (e.g., "Use proactively when...")
• Mention task types clearly
• Use natural language Claude will understand

Examples:

Good:

description: Expert code reviewer. Use proactively after code changes to check quality, security, and best practices.

Good:

description: Debugging specialist for errors, test failures, and unexpected behavior. Use proactively when encountering any issues.

Weak:

description: Helps with code

Optional Fields

tools (Optional)

Comma-separated list of specific tools the subagent can use. If omitted, inherits all tools from main thread.

Common tools:

• Read, Write, Edit - File operations
• Bash - Shell commands
• Grep, Glob - Search operations
• WebFetch, WebSearch - Web operations
• Task - Delegate to other subagents
• NotebookEdit - Jupyter notebooks
• MCP tools: mcp__server__tool

Examples:

Read-only agent:

tools: Read, Grep, Glob

Full-stack developer:

tools: Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Task

Code reviewer (no modifications):

tools: Read, Grep, Glob, Bash

Security consideration: Limit tools to minimum needed for the subagent's purpose.

model (Optional)

Specify which AI model the subagent uses.

Values:

• sonnet - Balanced capability (default for subagents)
• opus - Most capable, higher cost
• haiku - Fast and efficient, lower cost
• inherit - Use same model as main conversation

Examples:

Fast searching with Haiku:

model: haiku

Complex reasoning with Opus:

model: opus

Match main conversation:

model: inherit

Defaults: If omitted, uses configured subagent model (typically sonnet).

permissionMode (Optional)

Controls how the subagent handles permission requests.

Values:

• default - Normal permission flow (asks for confirmation)
• acceptEdits - Automatically accepts edit operations
• bypassPermissions - Bypasses permission system entirely
• plan - Plan mode (read-only, no execution)
• ignore - Ignores permission requests

Examples:

Auto-formatter (accepts edits):

permissionMode: acceptEdits

Data analyst (bypasses for automation):

permissionMode: bypassPermissions

Architecture planner (read-only):

permissionMode: plan

Standard agent (asks permission):

permissionMode: default

Security note: Use bypassPermissions carefully, only for trusted workflows.

skills (Optional)

Comma-separated list of skill names to auto-load when the subagent starts.

Format: skill-name-1, skill-name-2

Example:

skills: test-driven-development, code-reviewer

Skills are loaded into the subagent's context automatically, providing specialized knowledge.

Configuration Patterns

Pattern 1: Minimal Configuration

For simple, general-purpose subagents:

---
name: simple-helper
description: A basic helper agent for simple tasks
---

Provide quick answers and basic guidance.

Pattern 2: Read-Only Agent

For exploration and analysis without modifications:

---
name: documentation-reader
description: Read-only agent for exploring documentation
tools: Read, Grep, Glob
model: haiku
---

Explore documentation and provide information about what exists.
Never suggest modifications.

Pattern 3: Automation Agent

For automated workflows that shouldn't ask permission:

---
name: auto-formatter
description: Automatically formats code files without asking permission
tools: Read, Edit, Bash
model: haiku
permissionMode: acceptEdits
---

Automatically format code files when invoked.
Run formatting tools without asking for confirmation.

Pattern 4: Expert Agent with Skills

For specialized domains requiring loaded knowledge:

---
name: pdf-expert
description: PDF processing expert with specialized skills
tools: Read, Write, Bash
model: sonnet
skills: pdf-processing, form-filling
---

Process PDF files, forms, and document extraction using specialized skills.

Pattern 5: Security-Focused Agent

For security audits with restricted tools:

---
name: security-auditor
description: Security audit specialist. Use when reviewing code for vulnerabilities.
tools: Read, Grep, Glob, Bash
model: sonnet
permissionMode: default
---

Find vulnerabilities: SQL injection, XSS, CSRF, exposed secrets.
Provide severity ratings and remediation steps.

Pattern 6: Full-Featured Agent

For complex tasks requiring all capabilities:

---
name: full-stack-developer
description: Full-stack development expert. MUST BE USED for complex features requiring both frontend and backend changes.
tools: Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Task
model: opus
permissionMode: default
skills: code-reviewer, test-driven-development
---

Implement complex features end-to-end:
1. Analyze requirements
2. Plan implementation strategy
3. Implement backend then frontend
4. Write comprehensive tests
5. Document changes

Writing Effective System Prompts

The system prompt (Markdown body) defines the subagent's behavior and expertise.

Best practices:

• Be specific about role and expertise
• Include step-by-step workflows
• List key practices or checkpoints
• Define what NOT to do when relevant
• Provide examples or templates
• Keep focused on the subagent's purpose

Structure example:

You are a [role] specializing in [domain].

When invoked:
1. [First step]
2. [Second step]
3. [Third step]

Key practices:
- [Practice 1]
- [Practice 2]
- [Practice 3]

For each [task], provide:
- [Output 1]
- [Output 2]
- [Output 3]

Managing Subagents

Using /agents Command (Recommended)

Interactive management:

/agents

Provides:

• View all available subagents
• Create new subagents with guided setup
• Edit existing subagents
• Delete custom subagents
• See priority when duplicates exist
• Manage tool permissions easily

Direct File Management

Manual creation:

# Project subagent
mkdir -p .claude/agents
nano .claude/agents/my-agent.md

# User subagent
mkdir -p ~/.claude/agents
nano ~/.claude/agents/my-agent.md

Changes take effect on next Claude Code session start.

Best Practices

Design Focused Subagents

Create subagents with single, clear responsibilities:

✅ Good:

• code-reviewer - Reviews code quality
• test-runner - Runs and fixes tests
• security-auditor - Finds vulnerabilities

❌ Avoid:

• everything-agent - Does everything
• helper - Too vague

Write Detailed Prompts

Include specific instructions, examples, and constraints. More guidance = better performance.

Limit Tool Access

Only grant tools necessary for the subagent's purpose:

Security agent:

tools: Read, Grep, Glob, Bash  # No Write/Edit

Formatter:

tools: Read, Edit, Bash  # No WebFetch/Search

Version Control Project Agents

Check project subagents into git:

git add .claude/agents/
git commit -m "Add team subagents"

Team members automatically get the same subagents.

Test Subagent Behavior

After creating, test by:

1. Invoking explicitly: "Use the X subagent to Y"
2. Testing automatic delegation with relevant queries
3. Verifying tool access works as expected
4. Checking permission mode behaves correctly

Additional Resources

Reference Files

For detailed configuration examples:

• references/configuration-examples.md - 12 complete subagent configurations with explanations
• examples/ - Working subagent files ready to use

Built-in Subagents

Study Claude Code's built-in subagents:

• general-purpose - Multi-step tasks, uses Sonnet, all tools
• explore - Fast searching, uses Haiku, read-only
• plan - Architecture planning, uses Sonnet, plan mode

Related Documentation

• Subagents: /docs/en/sub-agents
• Plugins: /docs/en/plugins
• Settings: /docs/en/settings

Quick Reference

Field Summary

Common Tool Sets

Read-only: Read, Grep, Glob Basic editing: Read, Write, Edit With execution: Read, Write, Edit, Bash Full access: Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, Task

Permission Modes

• default - Standard (asks permission)
• acceptEdits - Auto-approve edits
• bypassPermissions - Skip all permissions
• plan - Read-only planning
• ignore - Ignore permission dialogs

Troubleshooting

Subagent not triggering automatically:

• Check description is specific with trigger phrases
• Include "Use proactively when..." in description
• Mention specific task types clearly

Tools not working:

• Verify tool names spelled correctly (case-sensitive)
• Check tools aren't blocked by permission settings
• Ensure MCP tools use mcp__server__tool format

Wrong model being used:

• Check model field is set correctly
• Verify model aliases: sonnet, opus, haiku, inherit
• Confirm model is available in your account

Permission issues:

• Review permissionMode setting
• Use default for standard behavior
• Use acceptEdits for auto-approval
• Avoid bypassPermissions unless necessary

Implementation Workflow

To create a subagent:

1. Identify purpose: What specific task will this subagent handle?
2. Choose tools: What tools does it need? (Minimum required)
3. Select model: Haiku (fast), Sonnet (balanced), Opus (capable), Inherit (match main)
4. Set permissions: How should it handle confirmations?
5. Write description: When should Claude invoke this? Include trigger phrases
6. Create system prompt: Define role, workflow, practices
7. Add skills (if needed): What knowledge should auto-load?
8. Test: Invoke explicitly and test automatic delegation
9. Iterate: Refine based on actual usage

Focus on creating focused subagents with clear purposes, minimal tool access, and specific descriptions for effective automatic delegation.