Agent Teams Composition Skill

You have expertise in composing and orchestrating Claude Code Agent Teams. Use this knowledge when the user's task would benefit from parallel multi-session work.

Agent teams are off by default. Use them only when the user explicitly started the session with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; otherwise use ordinary subagents, and don't propose teams unprompted.

When Agent Teams Fit

With teams enabled, they fit when:

• The user asks for comprehensive code review (suggest 3-reviewer team)
• The task involves work across multiple layers (frontend + backend + database)
• The user describes a bug with unclear root cause (suggest competing hypothesis investigation)
• The user wants to evaluate multiple technologies or approaches (suggest research panel)
• The task can be clearly split into independent, file-separated workstreams

Do NOT suggest agent teams when:

• The task is sequential or involves same-file edits
• A single session or subagent would suffice
• The coordination overhead would exceed the benefit
• The user is on a tight token budget

Team Templates

Full Review (3 teammates)

Security reviewer + Performance reviewer + Correctness reviewer (edge cases, error paths, callers).
Each reviews independently, then lead synthesizes findings.

Feature Dev (3 teammates)

Architect (plan approval required) + Implementer + Reviewer.
Architect plans first, implementer follows, reviewer validates the diff by review.

Debug Squad (3-5 teammates)

Each teammate investigates a different hypothesis.
Adversarial debate structure -- teammates challenge each other's theories.

Cross-Platform (3 teammates)

iOS developer + Android developer + Shared architect.
Each platform teammate owns their own directory. Architect coordinates API contracts.

Full-Stack (3 teammates)

Frontend + Backend + Database.
Clear file ownership per layer. Coordinate on API contracts first.

Research Panel (3 teammates)

Simplicity advocate + Performance advocate + Devil's advocate.
Each evaluates from a different perspective, then synthesize.

Spawn Prompt Formula

Every spawn prompt should include:

1. Role: "You are the [role] teammate responsible for [domain]"
2. Files: "You own files in [directories]. Do not modify files outside your scope."
3. Focus: "Specifically look for / implement [details]"
4. Context: "[Project-specific information the teammate needs]"
5. Deliverable: "Report your findings as [format] / Implement [specific output]"

Cross-Platform Awareness

• macOS and Linux: full support (in-process + tmux split panes)
• Windows: in-process mode only (automatic with auto setting)
• Always use --teammate-mode auto (default) for automatic detection

Related Documentation

• ~/.claude/docs/AGENT-TEAMS.md - Full guide with examples
• ~/.claude/docs/reference/workflows/agent-teams.md - Decision framework
• ~/.claude/commands/assemble-team.md - Quick team assembly command