Run child agents
Admiral gives the agents it launches a set of tools named the admiral tools. With them, an agent can create child sessions, delegate work to them, and collect the results. You watch the whole family in the sidebar.
What agents can do
An agent with the admiral tools can:
- Create a child session, optionally in a new worktree under
admiral/in the repository. - Send the child a task and wait for it to finish.
- Wait for several children at once.
- Answer a child's questions.
- Receive the child's final report.
- Close a child session and remove its worktree when the work is merged or abandoned.
Ask for it in plain language
You do not touch the tools. You describe the delegation:
Split this migration into three independent pieces. Run each piece in its own child session in its own worktree, then report which pieces are done and which failed.
The agent creates the children, watches them, and reports back in the parent transcript. Each child appears as its own session row, so you can open any of them and read its transcript.
Boundaries
- Children are scoped to the parent. A child cannot reach outside its own subtree.
- The parent decides when a child's worktree is removed. A checked worktree removal refuses to delete unmerged work.
Turn the tools off
Some tasks should not delegate. Settings has a per-device switch for the injected admiral tools. Turn it off, and agents on that device run without them. See Settings pages.
Drive the same tools from a shell
The same capability exists as a command-line interface:
admiral chat --jsonIt supports a dry-run mode and typed exit codes, so scripts can drive child sessions too.
Next steps
- Start a session: the parent session.
- Settings pages: the injection switch.