subagent-teams-skill_skill

Overview

This skill orchestrates in-session Task tool subagent teams for lightweight parallel work inside a single Claude Code session. It lets you fan-out research, implementation, review, and documentation across multiple subagents while sharing a single codebase and session context. Use it when you need fast, concurrent tasks without the overhead of full isolated worktrees.

How this skill works

You create a Team and spawn Task-tool subagents with parameters like subagent_type, model, prompt, and run_in_background to execute in parallel. Each subagent runs against the shared session context and reports progress via TaskCreate/TaskUpdate so you get live spinners and status updates. When agents finish, you manually fan-in results into a synthesized summary or next-phase tasks.

When to use it

• Parallel read-only or non-overlapping edits across files
• Fan-out research strategies to explore different parts of a codebase
• Run multiple reviewers in parallel for pre-PR checks
• Implement independent components concurrently (architect → builders → reviewer)
• Quick documentation updates across many files

Best practices

• Limit parallel agents to 5–7 to avoid context overflow
• Pick models by task: haiku for searches/reviews, sonnet for code, opus for architecture
• Avoid spawning agents that will edit the same files; use agent-teams for conflicting edits
• Use TaskCreate/TaskUpdate activeForm to show live progress and TaskUpdate addBlockedBy to sequence phases
• Design prompts so background agents are independent—background agents cannot see each other

Example use cases

• Research team: launch three Explore agents (auth, DB, API) and synthesize findings
• Implement team: Plan with an Opus agent, spawn 2–3 Sonnet builders, then a Haiku code-reviewer
• Review team: three haiku code-reviewers parallelize security, consistency, and performance checks
• Doc team: multiple general-purpose agents update README, CHANGELOG, and API docs independently