LLM Wiki

Appearance

Built 26/04/17 09:08commit f8ff6f9

Codex Subagents ​

中文 | English

Summary ​

This source explains how Codex delegates bounded work across specialized child agents, which built-in roles exist, how custom agent files are shaped, and why thread depth and concurrency limits matter for cost and predictability.

Source ​

Key Contributions ​

  • Limits subagent workflows to explicit delegation rather than automatic fan-out.
  • Defines the built-in default, worker, and explorer roles and the file schema for custom agents.
  • Explains agents.max_threads, agents.max_depth, and job runtime controls as core coordination limits.
  • Shows both interactive review-style delegation and CSV batch fan-out as supported patterns.

Strongest Claims ​

  • Delegation only pays when the work is clearly decomposable; otherwise it mostly adds cost and coordination risk.
  • Depth and concurrency settings are guardrails, not tuning trivia.
  • Custom agents should be narrow, role-specific, and allowed to inherit only the settings they truly need to override.