Today’s Codex will feel “primitive” by fall — and its own team’s roadmap backs it up

Thibault Sottiaux, who runs core products at OpenAI, dropped a blunt prediction on X late Monday: the Codex you're using right now will look primitive by fall. "It's pretty clear Codex is a good harness," he wrote, before pivoting hard. "But it will seem primitive in 2-3 months and we're about to go through another major evolution in how we use AI at the frontier." His kicker line is the one worth sitting with: "The next generation of models need more than your laptop."

He didn't spell out what's coming, but the timing lines up with moves OpenAI has already made. Codex got a GPT-5 update in early July and crossed 8 million users shortly after, so the product itself isn't struggling. The bottleneck, based on Sottiaux's comments, is architectural. Codex still leans on cloud infrastructure but often needs the developer's laptop to reach files and run tools. Close the laptop, and the agent can lose access to what it needs mid-task.

That's presumably why OpenAI announced in June it plans to buy Ona, a company (formerly known as Gitpod) that builds secure cloud development environments already used by 2 million developers. OpenAI framed the deal as the "next phase of Codex," one where an agent keeps grinding away in a customer's cloud long after the laptop that kicked things off is shut. The acquisition isn't finalized, and it's unclear whether Sottiaux's remarks are directly tied to it, but the direction is obvious: give agents a permanent workspace instead of a session that dies when you walk away from your desk.

OpenAI has already shown what sustained autonomy looks like. In a February experiment, Codex ran for about 25 hours straight, burned through roughly 13 million tokens, and produced around 30,000 lines of code building a design tool from nothing. For comparison, the piece notes Alibaba's Qwen3.8-Max agent recently worked autonomously for 16 days and shipped 265 commits without a human touching it. That gap between hours and days is exactly the kind of leap Sottiaux seems to be gesturing at.

But moving execution into the cloud doesn't erase the hard part, it just relocates it. Giving an agent standing access to a company's network or credentials is a real risk, which is why OpenAI says Ona's setup keeps things customer-controlled, with the organization deciding what the agent can touch while OpenAI handles the model and orchestration. Codex is already leaning into this agent-management identity, folded into the ChatGPT desktop app, running parallel tasks, and letting developers monitor jobs happening on devboxes or remote machines instead of their own laptop. Anthropic is chasing a similar problem from another angle, having acqui-hired Mendral to automate CI/CD chores like flaky tests. Whatever comes next, agents are going to need their own identities and audit trails, the same accountability humans already get, just layered on top of software that doesn't sleep.