Agent Teams is the experimental multi-agent mode where several Claude Code instances work together with a shared task list and shared state.

What it does

A team has a lead (coordinator) and teammates (workers). The lead decomposes the job, assigns tasks, and synthesizes results. Teammates claim tasks, work in parallel, and communicate directly when needed. You see their work in panes if you like. It's designed for jobs that genuinely benefit from parallelism - audits, broad refactors, large research.

When to use it

Tasks with many independent subtasks that can run in parallel.
Codebase-wide audits or refactors.
Research spanning many files or sources.
Any time sequential delegation would waste wall-clock time.

Gotchas

Agent Teams is experimental. Expect rough edges and breaking changes.
More agents means more cost. Measure the speedup versus the bill.
Coordination overhead can swamp the benefit on small jobs. Use for genuinely parallel work.

Official docs: https://code.claude.com/docs/en/agent-teams.md