Custom Subagents

Subagents are specialized instances of Claude with:

• Specific expertise or focus area
• Custom system prompts and instructions
• Defined tool access and permissions
• Ability to handle complex, multi-step workflows

Based on early documentation, custom subagents will likely be configured in:

• Location: .claude/agents/ (project) or ~/.claude/agents/ (user)
• Format: Configuration files (format TBD)
• Purpose: Specialized workflows, domain expertise, complex tasks

A specialized agent focused entirely on code review with:

• Deep knowledge of code review best practices
• Security vulnerability detection expertise
• Performance optimization experience
• Access to static analysis tools
• Checklist-driven review process

Potential workflow:

1. Analyze changed files
2. Run security scanners
3. Check performance implications
4. Review test coverage
5. Generate comprehensive review report

An agent specialized in creating and maintaining documentation:

• Technical writing expertise
• Documentation structure knowledge
• API reference generation
• Tutorial and guide creation
• Keeping docs in sync with code

Potential workflow:

1. Analyze code structure
2. Identify undocumented APIs
3. Generate draft documentation
4. Create usage examples
5. Suggest improvements to existing docs

Focused on comprehensive test creation:

• Testing framework expertise
• Coverage analysis
• Edge case identification
• Mock and fixture creation
• Test refactoring

Potential workflow:

1. Analyze code to be tested
2. Identify test scenarios
3. Generate unit tests
4. Generate integration tests
5. Create test fixtures
6. Verify coverage

Specialized in code quality improvement:

• Design pattern knowledge
• Code smell detection
• Refactoring techniques
• Safe transformation steps
• Impact analysis

Potential workflow:

1. Analyze code structure
2. Identify improvement opportunities
3. Propose refactoring plan
4. Execute transformations incrementally
5. Verify tests still pass

Expert in database schema evolution:

• SQL knowledge
• Migration strategies
• Data preservation
• Rollback planning
• Performance considerations

Potential workflow:

1. Analyze schema changes needed
2. Generate migration scripts
3. Create rollback scripts
4. Test on sample data
5. Document breaking changes

name: Code Review Agent
description: Specialized agent for comprehensive code reviews
expertise:
  - security analysis
  - performance optimization
  - code quality
tools:
  - Read
  - Grep
  - Bash(security-scanner:*)
system_prompt: |
  You are a code review expert focused on security,
  performance, and code quality. Always check for:
  - Security vulnerabilities
  - Performance bottlenecks
  - Best practice violations
  - Test coverage

workflow:
  - step: analyze_changes
    description: Review all changed files
  - step: security_scan
    description: Run security analysis
  - step: performance_check
    description: Check for performance issues
  - step: generate_report
    description: Create comprehensive review

allowed_tools:
  - Read
  - Grep
  - Bash(git:*)
  - Bash(security-scanner:*)
denied_tools:
  - Write  # Review agent shouldn't modify code
  - Bash(rm:*)

Follow Claude Code development:

• Claude Code GitHub
• Official Documentation
• Anthropic Blog
• Release notes and changelogs

Claude Code has built-in subagents accessible via the Task tool:

• general-purpose - Research and multi-step tasks
• Explore - Codebase exploration and search
• Plan - Implementation planning

Usage:

Claude, use the Explore agent to find all authentication code

Create skills for automatic activation:

• Agent Skills - Configure in .claude/skills/
• Context-based activation
• Single-purpose, focused tasks

Create commands for explicit invocation:

• Slash Commands - Configure in .claude/commands/
• User-invoked workflows
• Parameterized prompts

Extend capabilities with MCP:

• MCP Servers - Configure in .mcp.json
• External service integration
• Language-specific tooling

• Narrow focus: Expert in one specific domain
• Complementary: Works alongside other agents
• Non-overlapping: Doesn't duplicate built-in functionality

• Clear entry point: When should this agent be used?
• Defined steps: What's the agent's process?
• Clear exit: How does the agent finish?

• Minimum necessary: Only tools needed for the task
• Safety: Can't perform dangerous operations
• Audit trail: Actions are logged and reviewable

• Composable: Can work with other agents
• Stateful: Maintains context across steps
• Resumable: Can pause and resume workflows

Using the Task tool to create multi-step workflows:

Task(
  subagent_type="general-purpose",
  prompt="First, analyze the codebase. Then, identify issues. Finally, propose fixes."
)

Embedding complex workflows in slash commands:

---
description: Multi-step code analysis
---

# Analysis Workflow

1. **Understand**: Read and comprehend the code
2. **Identify**: Find issues and opportunities
3. **Analyze**: Deep dive into each issue
4. **Propose**: Suggest specific improvements
5. **Prioritize**: Rank by impact

[Detailed instructions for each step...]

Creating highly focused skills:

• One skill per specific task
• Clear activation triggers
• Defined tool access
• Single responsibility

• Claude Code Documentation
• Claude Code GitHub
• Anthropic Blog

• GitHub Discussions
• GitHub Issues

• Anthropic Agent Research
• Anthropic SDK

—

Note: This file serves as a placeholder and will be significantly expanded when custom subagent functionality becomes available in Claude Code. Check back for updates.