Claude Code Agent Teams lets you spawn multiple independent Claude Code instances that coordinate through a shared task list and mailbox. Unlike subagents which run within a single session and only report back, these are fully separate sessions that talk to each other directly.
I gave it a rough product idea, asked for a PM, architect, three builders, and a reviewer — Claude spawned all of them. tmux split-pane mode made this work visually. Each agent gets its own pane, everyone’s visible at once, and you can click into any pane to interact. It turns multi-agent coordination into something you can actually watch.
The coordination was what impressed me most. The PM wrote a PRD, sent me questions, I answered, the architect picked up the finalized doc, builders waited for dependencies to resolve automatically, then started in parallel. No manual handoffs — the mailbox and task dependency system handled it.
With Claude’s 1M context window each agent holds enough project context to stay coherent across a full build. By the end — multiple phases, multiple commits — I had a working app on localhost.
Good to know: the feature is experimental — enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS in settings. Avoid two agents editing the same file. 3-5 teammates is the sweet spot since token usage scales with team size. Teammates auto-load your CLAUDE.md, MCP servers, and skills.
