Subagent System

Delegating tasks to specialized AI agents involves creating subagents with custom system prompts, specific tools, and separate context windows. Use the /agents interface or manually create Markdown files in .claude/agents/ with YAML frontmatter defining agent name, description, and tools. Claude can then proactively delegate based on descriptions or you can explicitly invoke subagents by name.

A subagent system creates pre-configured AI personalities for task-specific workflows, solving problems of maintaining context and applying flexible permissions across complex tasks. Use subagents when you need specialized expertise handling different parts of a workflow—for example, delegating security review to one agent and test fixing to another.

Yes. Subagents support per-agent tool permissions and scope configuration. Each subagent is defined with its own set of tools in the Markdown frontmatter, allowing you to restrict capabilities to what each specialist needs and maintain security boundaries across delegated work.

Create subagents by defining Markdown files in .claude/agents/ with unique names, descriptions triggering automatic delegation, and system prompts reflecting each role's expertise. For code review and testing, define a code-reviewer subagent and a test-runner subagent, each with appropriate tools and instructions for their specialized tasks.

Automatic delegation occurs when Claude proactively invokes a subagent based on task descriptions and the agent's defined purpose. Explicit delegation happens when you directly invoke a subagent by name. Both approaches use the same underlying architecture and version-controlled best practices.

Yes. Best practices include version-controlled subagent definitions in .claude/agents/ directories, enabling consistent application across projects and user contexts. This ensures your specialized agent configurations remain auditable and reproducible.