“OpenAI recently shipped a feature for Codex that allows one thread to act as a high-level orchestrator controlling other, separate threads.”