Sub Agent Kit

Agente especializado que pode ser conectado a um OrchestratorKit como sub-agente.

Sub Agent Kit

AI-generated

Summary

Configure and run a specialized sub-agent powered by an LLM, allowing customization of agent identity, behavior, skills, output actions, guardrails, and interaction state management.

Inputs

  • agentName (required) — Identifier for the agent used by an orchestrator to invoke it (no spaces, e.g., 'specialist').
  • agentDescription (required) — Description shown to orchestrator to decide when to delegate tasks to this agent.
  • systemPrompt (required) — Initial system prompt defining the agent's behavior and role.
  • modelOverride — Optional override for the model used, replacing credentials' default (e.g., 'anthropic/claude-sonnet-4-5').
  • maxIterations (required) — Maximum number of LLM iterations per task interaction.
  • stateless (required) — If true, passes orchestrator-provided conversation history each call instead of maintaining internal session memory; recommended for agents not using tools.
  • outputContentKey (required) — JSON key under which the message to the user is returned in the agent's output.
  • outputInstructionsKey (required) — JSON key under which routing or control instructions are returned in the agent's output.
  • outputActions — Declare possible string actions the agent can return; these are auto-injected into the system prompt for the LLM to structure output accordingly.
  • inlineSkills — Define inline skills with name, description, and content that can be used by the agent as callable tools or guideline fragments.
  • guardrails — Rules to validate or block input/output content pre- or post-processing, supporting keyword detection, PII, secrets, regex, jailbreak, NSFW, and custom model checks, with fallback responses.

Output shape

a single object representing the configured sub-agent instance exposing a call method to process tasks

The output object has a call method accepting a context with task, history, and optional state, returning the LLM-generated response and usage stats. The agent auto-applies configured guardrails pre- and post-response and manages conversation state unless stateless mode is enabled. Output JSON keys for message and instructions are customizable.

Links

Discussion