defineHarnessSubagent

Purpose and behavior

Define a tool that runs a built-in child agent, with a fresh history, in the parent’s borrowed sandbox. The parent model sends { prompt }; the tool returns JSON text with the child’s text and conversation identifier, and a child failure is handled like any tool error. Child tokens count toward the dispatch usage and every ancestor budget, and ancestor permissions also apply to child tools.

Complete example and detailed rules.

Parameters and properties

• optionsRequired
  HarnessSubagentOptions
  Tool name, description shown to the parent model and built-in child agent. Unknown keys are rejected.
• options.nameRequired
  string
  Unique tool name exposed to the parent model, using 1–64 letters, digits, underscores or hyphens.
• options.descriptionRequired
  string
  Tells the parent model when to delegate to this child; sent with the tool schema.
• options.agentRequired
  CustomAgent
  Built-in agent from createAgent({ harness: createHarness(…), model }) that supplies the child’s instructions, tools, limits and permissions. A CLI agent is rejected with code configuration.

Returns

HarnessSubagent

Signature

export declare function defineHarnessSubagent(
  options: HarnessSubagentOptions,
): HarnessSubagent;

Related contracts

• HarnessSubagent
• HarnessSubagentOptions