4. Agents that start agents
A Claude Code session can run fleet spawn too. Then it is a
lead agent: it divides a task, starts one worker for each part,
and reads their reports. The page draws this as a tree: who started whom.
- A lead agent outside fleet uses the
fleetCLI or the control MCP at/mcp:agents,spawn,send,stop,resume,transcriptand more. - A worker inside a turn uses the worker MCP or the CLI:
report,done,ask,inbox,send,spawn,context. - The server records the parent of each agent. A child keeps the user of its parent.
- The spawn policy limits the tree: depth 3, and 32 live
children for each agent. A person who spawns has no limit
(
fleet/internal/spawnpolicy/spawnpolicy.go). A worker at depth 3 may still spawn one-shot probes withfleet spawn <name> --probe(haiku, at most 0.50 USD and 900 s, its own project, at most 25 in its life). - When a worker ends a turn with no
done, its parent gets a note in its inbox. - The swarm plugin is one lead agent. It starts workers through fleet. See a ticket, from issue to deploy.
- An agent can have a kind:
worker(the default),front-door,chieforkubejob. The kind tells fleet where the agent can run.
The agent graph draws the tree. The Builds panel joins the workers to the issues.
Next: 5. More machines: nodes. One machine has room for some turns only.