Nodal-Agents
Concepts

Orchestrators

How orchestrators delegate work to child agents and compile results.

An orchestrator is an agent whose role is set to orchestrator. It doesn't do the work itself — it breaks a request down and hands pieces to specialist child agents, then compiles the result.

Two delegation styles

Every orchestrator gets both delegation toolsets at runtime. It picks the style that fits the request; the runtime ensures only one style runs per job.

Inline delegation — assign_<agent>

The orchestrator calls assign_<slug> to hand a task to one child agent and waits for its result before continuing. Use this for sequential work, or when the next step depends on what the previous one found.

  • Only one assignment per turn.
  • After the child returns, either call return_result to finish or assign the next step.

Parallel fan-out — create_task

The orchestrator calls create_task once (or several times in the same turn) to create independent tasks, each with an assigned_to set to a child's slug. Tasks run concurrently. After creating them the orchestrator ends its turn with a brief return_result acknowledgment — the task board runs everything in the background and compiles the answers automatically.

Use depends_on inside create_task to sequence tasks that must run in a specific order within the same board.

Team block

The orchestrator's system prompt includes a ## Your team section built entirely from the database at job time: child agent names, slugs, assigned skills, connectors, and MCP tool inventories. No metadata is hardcoded in the runner — if you change a child's skills in the dashboard the team block updates on the next job.

Soft mode lean

An orchestrator's orchestratorMode column (router or planner) is a soft default preference surfaced in the prompt, not a hard restriction on the toolset:

  • Router lean — the team includes sub-orchestrators; inline assign_* delegation is usually the right call.
  • Planner lean — the team is independent workers; create_task parallel fan-out is usually the right call for multi-part work.

The model chooses freely; the lean is guidance, not enforcement.

Delegation limits

Delegation depth is capped at 3 levels (parent → child → grandchild). A child that tries to delegate further fails loud. The overall chain-resume limit (15 per job) also applies — a long sequential workflow that needs more than 15 child completions will be killed before it finishes.

See Agents for the full guard table.

On this page