Orchestrator Kit

Agente supervisor que delega tarefas aos nós SubAgentKit conectados.

Orchestrator Kit

AI-generated

Summary

Route tasks through multiple AI agents using either an LLM supervisor (orchestrator mode) or deterministic handoff chaining (handoff mode) to delegate work and return a unified response.

Inputs

  • Modo de Execução (required) — Choose between 'orchestrator' mode where an LLM routes tasks to agents dynamically, or 'handoff' mode where agents route to each other deterministically via action instructions.
  • Campo da Mensagem de Entrada (required) — JSON field path in input items containing the user message or task to process (default 'message').
  • Campo do ID de Sessão (required) — JSON field path for session identification to maintain context (default 'sessionId').
  • Campo de Saída (required) — JSON field path where the orchestrated response will be stored (default 'response').
  • Modelo Alternativo — Optionally override the language model used by the orchestrator.
  • System Prompt — Prompt text for the orchestrator LLM describing its supervisor role (only in orchestrator mode).
  • Máximo de Iterações — Maximum iterations the orchestrator LLM will make to decide routing and delegate tasks (default 20).
  • Agente de Entrada — Name of the first sub-agent to call, required in handoff mode. Must match a connected SubAgentKit agent.
  • Máximo de Saltos — Maximum number of handoffs between agents in handoff mode before terminating the chain (default 5).
  • Mapa de Roteamento — User-defined routing table mapping agent 'action' values to next agent names in the handoff chain; missing actions end the chain.
  • Ações Terminais — Comma-separated list of actions that terminate the handoff chain. The chain stops when one of these actions is encountered.
  • Campo de Skills — Field path to input JSON containing skills loaded from 'Skill Loader' node, used in orchestrator mode.
  • Skills Inline — Inline definition of skills with name, description, and content available to the orchestrator LLM.
  • Guardrails — Pre- and post-processing content checks to block or fallback on problematic inputs or outputs, such as NSFW or PII.

Output shape

A single item per input with JSON containing the original input augmented with the orchestration result under outputField; includes executionTrace of agent calls and detailed token usage statistics.

Supports two operation modes: 'orchestrator' uses an LLM supervisor to dynamically route among connected SubAgentKit agents with skill and guardrail integration; 'handoff' mode runs a deterministic routing chain among agents based on user-defined routingMap and terminalActions. Input fields for message, session ID, and output can be customized. Memory (AiMemory input) is integrated for conversational context if connected. Guardrails run on input (pre) and output (post) to filter content. Errors occur if required parameters or agents are missing.

Examples

Example 1: Orchestrator mode with LLM routing

Set Modo de Execução to 'orchestrador'; provide a system prompt; define inline or loaded skills; connect SubAgentKit agents on AiAgent input; specify inputField 'message' and sessionIdField 'sessionId'. The node routes the task dynamically and returns the orchestrator’s response.

Example 2: Handoff mode deterministic routing

Set Modo de Execução to 'handoff'; provide entryAgent as the starting sub-agent name; define routingMap with actions mapped to agents; specify terminalActions to end the chain; connect agents accordingly. The node iteratively calls agents following the routing map until a terminal action is reached, returning the final response.

Links

Discussion