Actor subagents

Subagents are independent session actors with their own UUIDs, contexts, FIFO inboxes, journals, and integer cursors. Tantra injects four model-visible tools where applicable.

Tool	Effect
`spawn(agent_name, input, name=None)`	Create a declared child, optionally name it, queue its first input, and return its UUID immediately
`send(agent_id, input)`	Durably queue a message across one direct parent-child edge
`status(agent_id)`	Read durable status for one direct child without activating it or reading its journal
`finish(result)`	Permanently close a child and deliver its result to its parent

class Worker(Agent):
    prompt = "Research one claim. Send updates, then finish with your finding."


class Lead(Agent):
    prompt = "Spawn workers in parallel and synthesize their finished messages."
    subagents = [Worker]

Multiple spawn calls in one assistant response start in parallel. The parent is not blocked and may become idle while children work. Assistant text stays in the actor’s own journal; information crosses an edge through send or finish. A child can send only to its direct parent or children, and a parent can send only to direct children.

Ordinary child completion leaves the actor reusable. Every child terminal outcome other than successful finish queues [agent <child-uuid> turn ended] <stable-json> for the direct parent. The JSON contains child_id, turn_id, outcome, stop_reason, and optional error; it contains no child assistant text, tool output, or result. The parent may inspect the child with status, send more work, or wait for an explicit result. finish permanently closes the child and delivers its result, and it fails while any descendant remains unfinished.

Runtime.max_depth defaults to 3 and rejects an over-depth spawn before creating a session. Human messages and typed asks are root-only. A child calling ctx.ask() receives an ordered tool error directing it to send() its parent. Runtime rejects static child permissions whose effective verdict is ask; a dynamic escalation to ask also fails without emitting AskRaised.

The optional child name is trimmed, and a blank name is omitted. It is a durable display label only: agent_name remains the actor type and the child UUID is unchanged. A deterministic retry must reuse the same normalized name.

Applications can poll runtime.tree_status(root_id) or follow runtime.watch_tree(root_id) without activating actors or reading journals. Subscribe to child activity with runtime.events(child_id, after=cursor) only when detailed output is needed. Root and child sequence numbers are unrelated.