The coding agents are arming up. OpenAI is bringing its most capable mode — “Ultra” — to Codex, its rival to Claude Code. The core idea: instead of a single agent working on a task, Ultra coordinates several sub-agents that tackle different parts in parallel.
What Ultra does
Ultra is the top tier of the GPT-5.6 family that OpenAI launched on July 9. The mode splits a complex task into pieces and runs a dedicated sub-agent for each one. A root agent hands out the work, the sub-agents run concurrently, and the results get synthesized into a single answer at the end. The point: if you can meaningfully divide a task, you finish faster.
In Codex, Ultra is available to users on the Plus plan and up; in ChatGPT Work, to Pro and Enterprise. The multi-agent feature is still in beta.
Why this is a jab at Anthropic
For readers of this blog, it matters for one reason: Claude Code has done this for a while. Anthropic built sub-agents in early, and recent releases even added nested sub-agents — agents that spawn their own agents, up to three levels deep by default. Orchestrating multiple models is business as usual in Claude Code.
OpenAI catching up and putting multi-agents front and center in Codex tells you one thing above all: the single, all-knowing agent is no longer the goal. The competition is shifting to who coordinates many agents best — fast, cheap, and without descending into chaos at the end.
My take
I’m cautiously optimistic about these multi-agent promises. Parallelism sounds great — and for cleanly divisible tasks (multiple files, independent tests, separate modules) it genuinely is. But it comes at a price: more tokens, more coordination overhead, more places where something can go wrong. Not every task gets better with five agents. Some just get more expensive.
For me, the real win isn’t the mode itself — it’s the competition behind it. When OpenAI and Anthropic push each other on orchestration, the people who benefit are us, the ones using these tools every day. I’ll still stick to my rule of thumb: first understand what the agent is supposed to do. Then decide whether you really need five of them.
Sources: