Delegate to other agents

Ask the agent in your conversation to divide work among other agents. For example:

Have Claude review the design, Codex implement the change, and Copilot check the tests. Coordinate their work and bring me the result.

The lead creates independent child sessions using agents installed and signed in on the same computer. Each child has its own conversation, working directory, and assignment. The lead can send messages, inspect progress, and stop its descendants. Children can message their parent and teammates, and delegate further.

Find the team

In Sessions, child agents are collapsed under the lead conversation. Choose the agents control to show them, then choose it again to collapse them. The control shows how many need attention even while collapsed. Your expanded groups are kept in the page URL, so reloading preserves them.

Search and Needs attention reveal the team with its parent for context. You can open any child session to read its conversation, give instructions, answer questions, or stop it.

Agent-to-agent messages are hidden in the conversation by default. Tap the session name to open Session details, then turn on Show agent-to-agent messages to see assignments, teammate messages, and child updates. Turn the switch off to hide them again. Your choice is remembered in this browser for your account. The agents still receive every message, and questions and approvals remain visible.

Results and questions

When a child finishes a turn, its final reply is sent to its parent automatically. The parent can continue working while the child runs, or end its own turn and resume when a result arrives. Messages identify the sending agent.

Questions and permission requests still belong to you. Open the child session to answer them. The parent receives a notice that the child needs input, but cannot approve on your behalf. Agent messages do not restart a session you have stopped or asked to stop.

After a crash or restart

Team relationships, assignments, and session IDs are saved. When the computer returns, the runner attempts to reopen unfinished sessions. Tell the lead to continue the team: it can use list_agents and get_agent to find the same children, then send_message to continue their work without creating replacements.

A crash can leave a message or tool action’s outcome uncertain. Have the agents inspect their files and progress before repeating work. Sessions you deliberately stopped remain stopped. Recovery requires the saved session data and working directories to remain available on the computer.

Working directories and permissions

The agent chooses the workspace for each assignment with workspace: "shared" or workspace: "worktree" in start_agent.

Shared mode is the default. Children use the lead’s current directory, including its uncommitted changes and existing assets, without copying files. Give each agent clear file ownership and have it coordinate changes with the lead. This mode also supports assignments outside Git repositories and projects with large asset libraries.

Explicit worktree mode creates a separate branch at committed HEAD. Worktrees are placed in .commandchain-worktrees beside the project checkout, instead of the runner’s state directory. Git LFS files remain pointers; assets are not downloaded or materialized automatically. Fetch only the assets required for the assignment. Uncommitted changes in the lead’s directory are not copied, so commit needed changes first or choose shared mode.

Worktrees are kept after completion so you can review and integrate the changes. The lead can use the returned directory and the child’s reported commit IDs to integrate the work.

Children inherit the parent’s approval mode at creation. A child cannot request a more permissive mode through delegation. If the chosen agent cannot enforce that mode, the start fails with an explanation.

Current limits

A team can run the lead plus 16 active child sessions, with up to three levels of delegation. Stopped agents remain in the team history and do not use an active slot. Stop an agent when its assignment is finished to free a slot for new work. All members run on the lead’s computer. Cross-computer teams are not available yet. Agent subscriptions and usage still apply to each agent you run.