Skip to content

Root Agent & Delegation

The agent you talk to runs as the general role with its own tools. It can answer directly or delegate bounded tasks to specialist agents. The old classifier-controlled, tool-less orchestrator is retired; this page keeps its existing URL for links and bookmarks.

  1. A channel or API sends a message to the backend.
  2. The root agent receives the request with session context, tools, and applicable skills.
  3. It does the work itself or calls spawn_child for a specialist task.
  4. It collects results and returns a response. Saved artifacts and execution records provide evidence to review.

The keyword classifier scopes context and supplies small-model hints. It does not choose whether the root runs or deterministically select the specialist.

The swarm has three levels: root → Agent → Subagent. The root and depth-one agents can spawn; depth-two subagents cannot. spawn_worker and spawn_team are retired LLM-facing tools. Parallel work uses spawn_child and results can be gathered with collect_children. Root-only create_pipeline starts a staged workflow with configured checks and approval gates.

Roles describe tool access and responsibility. The general root is one of the 16 roles, not an additional routing-only role. See Routing and Pipelines.

Planned children use a configured executor when available; otherwise expert model preferences and topic bindings determine the model. Unresolved specialist bindings fail; the root can use the configured default. Children do not simply inherit the parent’s model. Fallback behavior depends on the execution path and configuration; it is not guaranteed to stay local or free.

swarm.levelDefaults configures token, wall-clock, fan-out, and detached-child limits by level. Tokens share the parent’s pool; parent time spent awaiting children is excluded from its active clock. Limits constrain execution but do not guarantee exact billing or immediate cancellation of external work.

ASK actions require approval on attended sessions; descendants inherit that approval surface. Unattended ASK is blocked. Stored DENY takes precedence over broad allow rules. A reviewed scoped grant can authorize bounded automation. See Security.

Receipts summarize observed tool activity. Scorers check selected conditions and can trigger bounded retries. Neither certifies correctness or security. Durable swarm events support reconciliation after restart, not arbitrary resumption of interrupted model computation.