agent-teams
Claude Code Agent Teams
Agent Teams is a Claude Code feature (v2.1.32+, experimental) where one session — the lead — spawns other Claude sessions — teammates — that share a task list, exchange messages via a mailbox, and can talk to each other directly. Unlike subagents (which only report back to their parent), teammates are first-class peers you can also address directly.
When to reach for this vs. nearby patterns
| Pattern | Shape | Best for |
|---|---|---|
| Subagents | Spawn → work → return summary | One-shot research / verification where only the result matters |
| Agent Teams (this skill) | Lead + N teammates, shared task list, mailbox, can argue | Coordinated parallel work where teammates need to talk |
/swarm |
Orchestrator + workers in git worktrees, role-specialized (Feature/CRAP/Mutation/Acceptance) | Hardening one feature with continuous quality pressure |
/nightshift |
Orchestrator spawns long-running independent agents in cmux workspaces | Overnight backlog churn, no human in the loop |
The deciding question for Agent Teams: "Do my parallel workers need to share state and message each other?" If yes, use this. If no, subagents are cheaper and /swarm//nightshift give better isolation.
Token reality: each teammate is a full Claude session with its own context window. A team of 5 burns ~5× the tokens of a single session. Worth it for review, hypothesis-testing, and exploration. Wasteful for routine implementation.