I made a quick video on a Codex setting most people never open, because flipping it on changed how I run bigger tasks. Multi-agent mode lets Codex run several agents at once, each with its own context window, reporting back to the main agent. In the video I turn it on and push it to a live run of ten.
Why this setting exists
I recorded this after a tweet from Peter Steinberger, creator of open claw, the most popular AI coding tool right now. He commits at a scale most people do not believe, over 1,800 contributions on a single day, and one way he handles the onslaught of PRs is by spinning up 50 Codex agents in parallel because nothing off the shelf was built for that scale. That is the ceiling. The setting is what gets you started.
Where the toggle lives
Open Codex, go to the Experimental settings, and make sure multi-agents is checked. When I looked the day before, mine was off by default. Out of the box Codex sends one agent through your task start to finish. With the box checked it can spawn several that split the work and run at the same time.
The live demo
Rather than leave it theoretical, I enable it on screen and ask Codex to spin up 10 agents to audit my downloads folder and suggest a cleanup. Codex confirms it is planning parallel agents and spawns 10, each auditing from a different angle, large files and installers among them. You watch the whole set run at once instead of one agent grinding through in sequence.
Why concurrency helps
Each agent gets its own context window, so the main agent is not bombarded with all the noise and tokens from every subtask. That keeps the primary thread clear and cuts how often you have to compact the context window. When one agent gathers context while another edits, they run at the same time instead of one after the other, and the main thread only sees the summarised results.
The downloads cleanup is a toy example on purpose, so the mechanics are easy to see. The real use is the Steinberger case: a pile of independent work that one agent would grind through in sequence, split across many agents that each own a slice. Bug sweeps across a repo, auditing a folder from several angles, or triaging a stack of PRs all fit the same shape.
Watch the walkthrough above, and subscribe for more practical AI coding tips.
I specialize in setting up parallel-agent workflows so teams ship faster without drowning the main context. Book a call at cloudyeti.io/meet.