Delegating Work Across Multiple AI Sessions

Start with a session that has context — a plan, some exploration, a partial build. Run /fork to copy it into a new session. The original is unchanged; the fork has the same history and can continue from there. Now you have two sessions, and you can send one a task while you work in the other.

send_message with a to (the session's name or id) and a message (the task) sends the task to that session. Add notify_when_done: true to get a notice when it finishes. The target session receives the message as a peer message and starts working on it.

send_message can also set the target's mode (mode: "build") and model (model: "claude-sonnet-5-5"). This is useful when you want the delegate to work differently from the leader — a cheaper model for tests, a faster model for exploration, a stronger model for the hard parts.

When the delegate finishes, the leader gets a notice (if it asked for one). It can also run list_sessions with result_of to see the delegate's latest turn: how it ended, its reply, and the files it changed. The leader can then review the work, merge it, or send follow-up tasks.

You are building a feature with an API and tests. Fork the session twice. Send one fork the API half with mode: "build". Send the other the tests with mode: "build" and a cheaper model. Work on the integration in the original. When both forks finish, read their results and review the changes. Approve the good ones, reject the bad ones, and continue.

Each session is independent. A delegate does not see what the leader does after the fork, and the leader does not see the delegate's intermediate steps. If the delegate needs context the leader has, send it in the message. If the leader needs to see the delegate's work, read the results or open the session.