agent-teams

Agent Teams orchestration

Claude Code's experimental Agent Teams coordinates multiple Claude Code sessions. The session that creates the team is the lead (you, here); teammates are separate sessions with their own full context windows that message each other directly and pull work off a shared task list.

This differs from subagents (the Agent tool): subagents run inside one session and only report back to their parent. Teammates are peers — they coordinate through a shared mailbox and task list, and self-claim work.

The one rule this skill exists to enforce

Delegate and wait. Do not implement the parallel work yourself in the lead session. When a task decomposes into independent streams, the lead's job is to spawn teammates, assign/seed tasks, monitor, and synthesize — not to write module A then module B then the tests sequentially. If you catch yourself doing the work, stop and delegate it.

• ❌ Anti-pattern: "Write the auth module, then the API, then the tests" → lead does all three in order.
• ✅ Pattern: spawn 3 teammates (auth / API / tests), assign clear tasks, then wait for all to finish before synthesizing.

If asked to proceed and teammates aren't done: do not start doing their work — wait, or steer them.

Preflight — is it enabled?

Agent Teams is off by default and needs Claude Code v2.1.32+. It requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 (env or settings.json env). If a request clearly wants a team but the feature may be off, check first and tell the user how to enable it rather than silently falling back to doing everything solo:

// ~/.claude/settings.json
{
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" },
  "teammateMode": "auto"   // "auto" | "in-process" | "tmux"
}

teammateMode: auto (split panes if in tmux, else in-process), in-process (cycle with Shift+Down), tmux (needs tmux/iTerm2; one pane per teammate). Per-session override: claude --teammate-mode in-process. Teammate model is set via /config → "Default teammate model" (does not inherit /model).