Claude Code agent teams start with a blank slate

Enabling CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS in settings.json lets a Claude Code session spawn teammates that work in parallel instead of one after another, per the docs. One orchestrator hands out pieces of a job and several agents run at once. The catch: each teammate starts with none of the context from the chat that spawned it.

No shared memory. Each teammate reads CLAUDE.md and whatever MCP tools are configured, and nothing else, according to the docs. It doesn't know what you and the orchestrator already worked out. Any file path, constraint, or decision you already settled has to go straight into the spawn prompt.

Teammates coordinate with each other through the file system, not shared chat. That's a different way of working than one long conversation, where everything said earlier stays in view. Writing a spawn prompt that carries the full context a teammate needs is its own skill, separate from describing what to build.

Tokens scale per teammate, so it's not something to leave running by default. The docs reserve it for research, review, and building new modules, jobs with a clear brief and a clear finish line.