Claude Code sessions can now message each other and split work
ListAgents and SendMessage shipped on macOS/Linux in 2.1.224; opt-in Agent Teams add a lead, a shared task list, and a mailbox that won't relay approvals.
Copy markdown
Two new tools: ListAgents and SendMessage
Claude Code 2.1.224 (Aug 7) lets one session discover others with ListAgents and hand them work with SendMessage — macOS and Linux only, with no Windows-native support. A message is text one Claude writes to another: never conversation history or files.
Agent Teams: a lead, teammates, a shared task list
Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and your main session becomes a lead that spawns teammates, each with its own context window. They claim work from a file-locked shared task list and talk through a mailbox. Still experimental and off by default.
Built for parallel review and adversarial debugging
The canonical plays: spawn three teammates to review a PR through security, performance, and test-coverage lenses, or five to argue competing bug hypotheses and disprove each other. Docs suggest 3-5 teammates; token cost scales linearly per teammate.
A teammate can't approve permissions for you
An incoming agent message is treated as untrusted input, so a relayed 'it's approved' never counts as your consent. In auto mode a classifier screens every inter-agent message — including shutdown and plan-approval protocol messages — before it is delivered.
In-process by default; split panes need tmux or iTerm2
The default runs every teammate in one terminal (arrow keys, then Enter to view and message one). Split-pane mode gives each its own pane but needs tmux or iTerm2's it2 CLI — and isn't supported in VS Code's terminal, Windows Terminal, or Ghostty.
Know the limits before you lean on it
/resume and /rewind won't restore in-process teammates, task status can lag and block dependents, and it's one team per session with no nesting. Start with research and review, not parallel edits to shared files.