Agent Architecture/Subagents & Agent Teams
Advanced10 min

Subagent Invocation Patterns

Claude Code provides four distinct ways to invoke subagents — from automatic selection to explicit CLI flags. Knowing which pattern to use determines whether your agents fire reliably or silently miss.

Quick Reference

  • Natural language: Claude picks the agent if the task matches the description
  • @agent-name: explicit typeahead invocation — bypasses Claude's judgment entirely
  • --agent agentName: CLI flag for scripted or programmatic session startup
  • Agent tool call: invoke via the Agent tool inside multi-agent workflows
  • Foreground blocks until done; background runs in parallel and returns a summary
  • Subagents do NOT inherit parent conversation history — only CLAUDE.md and MCPs
  • Results always return as a summary — design agents to produce structured output
  • @ typeahead was added in April 2026 — upgrade Claude Code if you do not see it

The Four Invocation Patterns

There is no single right way to invoke a subagent — each pattern exists because a different situation calls for different levels of control. Natural language is the most ergonomic. @-mention is the most reliable. The CLI flag is for automation. The Agent tool is for programmatic composition.

PatternSyntaxWhen to Use
Natural language"review this PR for security issues"When Claude should decide whether a subagent is appropriate
@-mention@security-reviewerWhen you want a specific agent and do not want Claude guessing
CLI flagclaude --agent code-reviewerWhen scripting or launching sessions programmatically
Agent toolAgent({ subagent_type: "code-reviewer" })Inside multi-agent orchestration workflows