The most underrated new Codex feature.
How I let OpenAI Codex run a team of AI agents from a single Codex coordinator thread.
Hey 👋,
OpenAI’s coding agent, Codex, recently shipped one of its most powerful features, and almost nobody is talking about it. Codex can now manage its own threads.
It sounds trivial, but this is the most interesting AI coding feature I’ve seen in ages.
If you haven’t used Codex, it’s an AI coding assistant, similar to Claude Code. You give it a task in plain English and it writes the code for you. A thread is just a single Codex conversation, one chat working on one task.
Until now, if you wanted Codex to work on five things at once, you had to open five conversations and steer each one yourself. That works, but you are the bottleneck. Now you can create one coordinator thread that acts like a manager, spinning up and managing other threads for you.
I got the idea from the Codex team. They posted something along the lines of: if you ever get tired of managing your Codex threads, just let Codex manage them itself. That is exactly what this does.
The coordinator gets a small set of tools. It can create threads, list them, read them, send messages to other threads, pin them, archive them, and set their titles. The one that matters most is being able to send messages to other threads. That is what lets a single coordinator thread talk to other conversations and direct their work.
Here is what I did with it. I have been building a coding agent of my own, called Neo, and I had a backlog of tickets I wanted to get through. Rather than working on them one at a time, I gave four of them to a single coordinator thread and asked it to coordinate four independent worker threads, one per task.
Each worker had to use its own branch and its own git worktree. This part is really important. The separate branch and worktree is what lets every thread run at the same time without getting in each other’s way. Without it they would be editing the same files and tripping over each other.
I also gave each worker the same simple workflow to follow. Take the ticket. Implement the changes. Write tests where appropriate. Use a sub-agent to review the work. Address any valid feedback. Then open a separate draft pull request.
So I ended up with one thread spinning up several others and managing all of the work, with four tickets moving at the same time. I set the work up, then watched the coordinator run it.
A couple of things stood out to me.
The separate branch and worktree per worker is the part that makes this work. It is easy to skip, and it is the whole reason the parallel work is safe.
Letting the agent clean up after itself helps more than I expected. Codex can pin the threads you care about and archive the rest, which keeps your workspace usable once you have a lot of them running.
This might sound like a small change, but it is one of the most useful features I have seen in a long time. It means I can work on a lot more at once than I used to, because I am setting up the work and reviewing it rather than driving every thread by hand.
I made a video covering this in more depth, including the full demo. You can watch it here:
Thanks for reading. Have a great week.
P.S. If you want to go deeper on building real software with AI agents, that’s what I’m building inside AI Engineer.