Hierarchical Multi-Agent Systems

What variant of Codex are you using?

Desktop

What feature would you like to see?

TicketID:
019da542-c9b7-7873-a441-a9544c6dc739

[Feature Request - Hierarchical Multi-Agent Systems]
This would be a huge improvement for agent coding.

1. I imagining having a button on the sidebar called something like "Take Off" or "Orchestrate" or something similar.

1. When you click on that button, a dialog opens.

1. In the dialog, you would have 1 main text area. In that top "main agent" text area, you describe a high level task in either plan/agent mode what you want to do. This is where you use GPT-5.4 main big model to lay out the high level plan.

1. Below the "main agent" text area, you have a a group called "Operators".

Inside that group, there is a dropdown where you select a model. I am imagining selecting smaller models, like gpt-5.4-mini or gpt-5.4-nano.
You also have a description field for custom information for each "operator".
Maybe something more, like if you want to create a new git branch/worktree.

1. When you now submit from the dialog, it will kick off multiple "operators".

I'm thinking in a linear terms, meaning the "main agent" will always lay out the foundation and then kick off each "operator" with its tasks.

You can visually imagine this sitting in the control room at the airport controlling all planes.

----

Token Efficiency: This "Orchestrator" model is much more cost-effective. Instead of one massive chat history (where every message gets exponentially more expensive), only the "Main Agent" holds the long-term context. The "Operators" start with fresh, small contexts for their specific tasks. This saves a huge amount of tokens for both the user and OpenAI.

Scalability: By using smaller models (like Nano/Mini) for the heavy lifting of writing code, the system becomes faster and more responsive. It avoids "context saturation" where the AI starts to hallucinate or get sluggish in long sessions.

Win-Win: This approach reduces inference costs through better use of prompt caching and smaller models while giving me a more reliable and better tool in the long run.

Additional information

_No response_