OpenAI just gave Codex's multi-agent mode a practical new capability: Sol can now directly assign tasks to Luna. For example, you can have the strongest agent, Sol, break down a project and do the final review, while a bunch of faster, cheaper Luna agents work through the pieces in parallel.
Codex already supported one agent spinning up multiple sub-agents and splitting work across different models. But Luna had a special limitation: it could only execute tasks on its own, and couldn't be pulled into a team by a higher-level agent like Sol. That restriction is now gone.
OpenAI is also explicitly recommending this division of labor. Sol is built for complex, open-ended tasks; Luna is better suited for well-defined, repetitive jobs like searching, organizing, extracting, and classifying. One agent handles the thinking and the acceptance checks; a squad handles the batch execution.
https://twitter.com/thsottiaux/status/2088725163923984638?s=46