How to enable parallel agents in OpenAI Codex
Open Codex, go to experimental, and check multi-agents. Then ask Codex to spin up several agents, each with its own context window, and report back to the main agent.
The main idea
Parallel agents means you ask a group of agents to do a task at the same time. You do not depend on one agent or one context window. Each agent has its own context window and reports back to the main agent, so you get the work done faster.
I made the video after I saw a tweet from Peter Steinberger, the creator of OpenClaw. He had a flood of pull requests and spun up 50 Codex agents in parallel. I wanted to try it myself.
To turn it on, open Codex, go to experimental, and make sure multi-agents is checked. It was not checked on my account when I tried it. Then you can ask in plain words. I typed a request to spin up 10 agents to look at my downloads folder and suggest a cleanup. Codex answered that it was spinning up 10 parallel agents, each auditing the folder from a different angle, such as large files and installers. Because each agent has its own context window, the main agent is not flooded with tokens and does not have to compact its context.
Key points
- In Codex, open experimental and check multi-agents.
- Ask for parallel agents in plain words, for example spin up 10 agents to look at my downloads folder and suggest cleanup.
- Codex spawned 10 agents, each looking from a different angle, such as large files and installers.
- Each agent has its own context window and reports back to the main agent.
- Separate context windows keep noise and tokens out of the main agent's context.
Save my free seat →