Creating Subagents

Reference files: ${SKILL_DIR}/references/

IMPORTANT: Do NOT search the project directory for skill files. If you cannot find a file, use Glob: .claude/plugins/cache/**/creating-subagents/**/*.md

Subagents enable delegation of complex tasks to specialized agents that operate autonomously without user interaction, returning their final output to the main conversation.

1. Run /agents command
2. Select "Create New Agent"
3. Choose project-level (.claude/agents/) or user-level (~/.claude/agents/)
4. Define the subagent:
   • name: lowercase-with-hyphens
   • description: When should this subagent be used?
   • tools: Optional comma-separated list (inherits all if omitted)
   • model: Optional (opus, sonnet, haiku, or inherit)
5. Write the system prompt (the subagent's instructions)

• Code quality and maintainability
• Security vulnerabilities
• Performance issues
• Best practices adherence

</example>
</quick_start>

<file_structure>
| Type | Location | Scope | Priority |
|------|----------|-------|----------|
| **Project** | `.claude/agents/` | Current project only | Highest |
| **User** | `~/.claude/agents/` | All projects | Lower |
| **Plugin** | Plugin's `agents/` dir | All projects | Lowest |

Project-level subagents override user-level when names conflict.
</file_structure>

<configuration>
<field name="name">
- Lowercase letters and hyphens only
- Must be unique

</field>

<field name="description">
- Natural language description of purpose
- Include when Claude should invoke this subagent
- Used for automatic subagent selection

</field>

<field name="tools">
- Comma-separated list: `Read, Write, Edit, Bash, Grep`
- If omitted: inherits all tools from main thread
- Use `/agents` interface to see all available tools

</field>

<field name="model">
- `sonnet`, `opus`, `haiku`, or `inherit`
- `inherit`: uses same model as main conversation
- If omitted: defaults to configured subagent model (usually sonnet)

</field>
</configuration>

<execution_model>
<critical_constraint>
**Subagents are black boxes that cannot interact with users.**

Subagents run in isolated contexts and return their final output to the main conversation. They:
- ✅ Can use tools like Read, Write, Edit, Bash, Grep, Glob
- ✅ Can access MCP servers and other non-interactive tools
- ❌ **Cannot use AskUserQuestion** or any tool requiring user interaction
- ❌ **Cannot present options or wait for user input**
- ❌ **User never sees subagent's intermediate steps**

The main conversation sees only the subagent's final report/output.
</critical_constraint>

<workflow_design>
**Designing workflows with subagents:**

Use **main chat** for:
- Gathering requirements from user (AskUserQuestion)
- Presenting options or decisions to user
- Any task requiring user confirmation/input
- Work where user needs visibility into progress

Use **subagents** for:
- Research tasks (API documentation lookup, code analysis)
- Code generation based on pre-defined requirements
- Analysis and reporting (security review, test coverage)
- Context-heavy operations that don't need user interaction

**Example workflow pattern:**

Main Chat: Ask user for requirements (AskUserQuestion) ↓ Subagent: Research API and create documentation (no user interaction) ↓ Main Chat: Review research with user, confirm approach ↓ Subagent: Generate code based on confirmed plan ↓ Main Chat: Present results, handle testing/deployment

</workflow_design>
</execution_model>

<system_prompt_guidelines>
<principle name="be_specific">
Clearly define the subagent's role, capabilities, and constraints.
</principle>

<principle name="use_pure_xml_structure">
Structure the system prompt with pure XML tags. Remove ALL markdown headings from the body.

```markdown
---
name: security-reviewer
description: Reviews code for security vulnerabilities
tools: Read, Grep, Glob, Bash
model: sonnet
---

<role>
You are a senior code reviewer specializing in security.
</role>

<focus_areas>
- SQL injection vulnerabilities
- XSS attack vectors
- Authentication/authorization issues
- Sensitive data exposure
</focus_areas>

<workflow>
1. Read the modified files
2. Identify security risks
3. Provide specific remediation steps
4. Rate severity (Critical/High/Medium/Low)
</workflow>

❌ Bad: "You are a helpful assistant that helps with code" ✅ Good: "You are a React component refactoring specialist. Analyze components for hooks best practices, performance anti-patterns, and accessibility issues."

• <role> - Who the subagent is and what it does
• <constraints> - Hard rules (NEVER/MUST/ALWAYS)
• <focus_areas> - What to prioritize
• <workflow> - Step-by-step process
• <output_format> - How to structure deliverables
• <success_criteria> - Completion criteria
• <validation> - How to verify work

• Use role + constraints + workflow minimum
• Example: code-reviewer, test-runner

Medium subagents (multi-step process):

• Add workflow steps, output_format, success_criteria
• Example: api-researcher, documentation-generator

Complex subagents (research + generation + validation):

• Add all tags as appropriate including validation, examples
• Example: mcp-api-researcher, comprehensive-auditor

Keep markdown formatting WITHIN content (bold, italic, lists, code blocks, links).

For XML structure principles and token efficiency details, see the creating-skills skill's use-xml-tags.md (Glob: .claude/plugins/cache/**/creating-skills/references/use-xml-tags.md) — the same principles apply to subagents.

> Use the code-reviewer subagent to check my recent changes

> Have the test-writer subagent create tests for the new API endpoints

• Project: .claude/agents/subagent-name.md
• User: ~/.claude/agents/subagent-name.md

Subagent usage and configuration: ${SKILL_DIR}/references/subagents.md

• File format and configuration
• Model selection (Sonnet 4.5 + Haiku 4.5 orchestration)
• Tool security and least privilege
• Prompt caching optimization
• Complete examples

Writing effective prompts: ${SKILL_DIR}/references/writing-subagent-prompts.md

• Core principles and XML structure
• Description field optimization for routing
• Extended thinking for complex reasoning
• Security constraints and strong modal verbs
• Success criteria definition

Advanced topics:

Evaluation and testing: ${SKILL_DIR}/references/evaluation-and-testing.md

• Evaluation metrics (task completion, tool correctness, robustness)
• Testing strategies (offline, simulation, online monitoring)
• Evaluation-driven development
• G-Eval for custom criteria

Error handling and recovery: ${SKILL_DIR}/references/error-handling-and-recovery.md

• Common failure modes and causes
• Recovery strategies (graceful degradation, retry, circuit breakers)
• Structured communication and observability
• Anti-patterns to avoid

Context management: ${SKILL_DIR}/references/context-management.md

• Memory architecture (STM, LTM, working memory)
• Context strategies (summarization, sliding window, scratchpads)
• Managing long-running tasks
• Prompt caching interaction

Orchestration patterns: ${SKILL_DIR}/references/orchestration-patterns.md

• Sequential, parallel, hierarchical, coordinator patterns
• Sonnet + Haiku orchestration for cost/performance
• Multi-agent coordination
• Pattern selection guidance

Debugging and troubleshooting: ${SKILL_DIR}/references/debugging-agents.md

• Logging, tracing, and correlation IDs
• Common failure types (hallucinations, format errors, tool misuse)
• Diagnostic procedures
• Continuous monitoring

• Valid YAML frontmatter (name matches file, description includes triggers)
• Clear role definition in system prompt
• Appropriate tool restrictions (least privilege)
• XML-structured system prompt with role, approach, and constraints
• Description field optimized for automatic routing
• Successfully tested on representative tasks
• Model selection appropriate for task complexity (Sonnet for reasoning, Haiku for simple tasks)