Orchestration of multiple agents

[enriquemorenotent]

What feature would you like to see?

Today, to work on a frontend project, I opened a Codex session (let's call it Codex FE). While working on a problem, Codex told me that the best way to proceed depended on some details about the contract with the backend. So waht I did it:

1. Went to my backend codebase and opened a new Codex session (lets call it Codex BE).
2. Asked Codex FE to give me a prompt for Codex BE to give him all the information it needed.
3. Pasted the given prompt into Codex BE
4. Pasted the given reply into Codex FE, so that it knew the best way to proceed.

This brought me to a realization: We should be able to orchestrate coding agents. Codex instances should be able to talk to each other, the same way that human coders communicate with each other inside a company, since each one has a dedicated domain.

My example is a use case, but I am sure there a lot of situations where it could be useful.

Are you interested in implementing this feature?

This is no small task, by any means, but I would be willing to give it a go, as long as there is consensus on the usefulness of this feature, and the way to implement it. Even though the Codex team is probably better equipped to achieve this

[freelanceagent1]

Use an MCP like desktop commander that spawns terminals with codex agents envoked but how do you want to avoid context clutter while orchestrating 2-3 agents via single agent?

FIY. There are bmad-method and vibe-kanban github repos you'd want to look into if your main issue is interaction with the agent. BMAD method in particular is very useful by providing pre-defined YAML config files for agent entities and makes it easy to setup an agile dev flow, create epics and stories, test story execution etc. all tho it might require some finetuning for your use case.

[enriquemorenotent]

Use an MCP like desktop commander that spawns terminals with codex agents envoked

Will this work on an application like Codex, where the process does not end, until the user manually exits?

but how do you want to avoid context clutter while orchestrating 2-3 agents via single agent?

The idea is that the agents communicate to the main orchestrating agent only the information that is relevant to them (the same way that project members communicate with each other, without needing to share the whole context of their operation)

[grahama1970]

Use a shared db and store messages b/w tasks? id, agent_id, message, timestamp Some Claude-code-centric github projects do this, I think. Still, have fun with it :)

[wbdb]

I have already worked with various orchestrations for agents (Roo Code etc), and none of them were really satisfactory. As long as 5-10 % of the code still needs revision iterations or there were misunderstandings despite very specific instructions, things would simply multiply here. High costs, little sustainable output.

In most cases, I would say it is better to work with familiar structures, such as MVC frameworks (e.g. CodeIgniter), etc.

Then you can build migrations, models, controllers and views one after the other. And eliminate errors in good time before they become costly.

[pakrym-oai]

We are looking into similar use cases under subagent support https://github.com/openai/codex/issues/2604