Child agents

An agent can declare other agents it may delegate to. Every binding names an agent file and chooses a delegation mode.

Native delegation

The bundled planner binds its Socrates reviewer this way:

subagents:
  socrates:
    agent: socrates
    mode: native

The parent delegates through its native harness. The child’s definition lives in agents/socrates.md.

Native child declarations support one level. Claude native children inherit the parent’s connections. Harness capabilities determine the native delegation behavior.

Process delegation

Use process mode to launch a separately generated configuration, including for cross-harness children:

subagents:
  worker:
    agent: worker
    mode: process

The bundled worker is a Codex agent. Its definition illustrates a worker with its own model and skill selection:

---
harness: codex
model:
  name: gpt-5.6-luna
  reasoning: max
description: Implement assigned work and return verification evidence.
skills:
  - implementer
---
 
Implement the assigned task, run relevant checks, and report evidence.
Return remaining questions to the parent.

Choose a model available to your harness and account. Process children receive their own generated configuration.

Keep instructions and identities distinct

A document under a skill’s references/ directory is supporting guidance. A child agent is an identity under agents/, selected through an explicit binding.

Inspect the parent graph before launching:

agent-farm inspect planner --workspace my-project

Use separate worktrees when parent and child work should make independent code changes. Generated bundles share access to the repository, credentials, and native global configuration.

Sources: Agent definitions , child behavior , and bundled worker .