Spawned subagents

create_subagent (alias spawn_subagent) is a system tool: the parent LLM creates a specialist during a turn. It starts one pre-authored Agent manifest for a single task and returns that child's final response as the tool result. It differs from delegate_to_<id>: delegation calls a peer in the MAS topology, while creating a subagent materializes a temporary child from a named template.

Enablement

Declaring the system tool is the whole contract: there is no separate capability flag. An agent without this entry cannot spawn. Use either name:

spec:
  tools:
    - kind: system
      name: create_subagent
      params:
        templates:
          - id: reviewer
            ref: ./agents/reviewer.yaml
            description: Review the proposed change
        max_spawns: 8
        max_depth: 3

The runtime advertises both create_subagent and spawn_subagent to the model so it can "create a subagent" without a hidden host API. Template IDs are unique within the manifest. References are containment-checked and must resolve to valid Agent manifests. The model selects a declared template and supplies its task; it cannot create a manifest or grant new tools at runtime.

Bounds and lifecycle

The default session budget allows eight child runs at a nesting depth of three. SpawnLedger enforces both limits inside the ctl spawner. Each call materializes one child, registers it on the local bus, runs one turn, returns the result, and removes the instance and endpoint in a finally block. Parent-call identity is preserved for trace attribution.

This release can run a child synchronously (one shot) or concurrently when the async driver is enabled (ainvoke, dispatch: parallel on a workflow). A child is torn down after its turn unless the session keeps it. Bounds still apply.

See Tutorial 9 for an incident coordinator that creates researcher and reviewer specialists, and the runnable manifest pair.