OpenAI's Codex ships a 'Persistent' mode that runs until you stop it
The always-on agent race got a real code drop today. OpenAI is wiring a new "Persistent" setting into Codex's reasoning-effort menu — the closest a frontier lab has come to shipping a product around the idea of an agent that doesn't politely end its turn after a few minutes.
Codex in Persistent mode is told to "continue working until put to sleep," according to a file in the Codex core that WIRED spotted in a recent pull request. That phrasing is doing the real work: today's reasoning modes stop after a task is "done" (or after a few hours), even if the job isn't finished. Persistent mode flips that — the agent is the one that decides when to stop, not the timer. The same file tells the agent that Persistent mode does not expand what it's allowed to do and that altering anything outside the user's own system requires explicit approval first, a guardrail that reads like OpenAI is trying to make the always-on behavior land without the obvious safety scares.
The setting sits in Codex's shared core, not the terminal-specific code, which is the tell that this is intended for more than the command-line tool. Thibault Sottiaux, OpenAI's head of core products, told WIRED the company explores a lot internally in its open repo, but the strategic intent is no secret: in a TIME interview two days ago, Sottiaux said OpenAI is "close to showing a product built around persistence and always-on execution," accessible from "almost anywhere," that "keeps doing useful work based on new information and feedback." He called it "definitely going to feel like a new thing to people."
That phrasing tracks the broader product line OpenAI has been quietly assembling. ChatGPT Work — the assistant that does tasks, not just answers — launched to coincide with what insiders call "the Merge," the decision to pull Codex's agentic capabilities into the chatbot and combine the compute and product teams behind both. Business revenue passed consumer revenue at OpenAI in July for the first time; the company is at a roughly $40 billion annualized run rate, and CFO Sarah Friar has been running public-company drills with the implication that an IPO could land by 2027. The corporate plumbing is being rearranged around the assumption that the next big product surface is the always-on agent.
The strategic backdrop is Sam Altman's long-stated goal of turning ChatGPT into a proactive personal assistant that "acts before being asked, recommends things on its own, and performs mundane tasks such as buying concert tickets autonomously." The Decoder frames the new Persistent mode as OpenAI's concrete answer to that vision — and as a direct response to xAI's Grok Bot, the always-on agent app that ships with its own cloud computer, bundled into SuperGrok Heavy at $300 a month. We covered SpaceXAI's launch of Grok Bot back in August, when the framing was that Grok had drawn first blood on a category every frontier lab now wants to claim. Today's WIRED scoop is the first real evidence that OpenAI is more than watching from the bench.
What the file does not tell you is how OpenAI plans to price or distribute a Persistent Codex when it ships. There's a real tension in the company's own playbook here: the same week Astra was held back over the Hugging Face agent-internet incident, OpenAI is preparing to ship an agent whose entire product premise is that it does not stop on its own. The "continue working until put to sleep" instruction is paired with the explicit boundary that it cannot expand its own permissions, which is the narrowest possible version of an always-on agent — one that has to ask before it does anything new. That is going to be the architectural shape of the product category for a while: agents that can outlast a session, but only inside a fence.
What to watch: whether OpenAI rolls Persistent mode out to ChatGPT Work users first or holds it for an enterprise SKU, and whether the explicit "ask before doing anything new" guardrail shows up in any competitor's agent docs in the next 30 days — that would be the early signal the fence is becoming table stakes.
OpenAI just put an always-on agent in its own reasoning-effort menu. Would you trust a model that's allowed to keep working while you're not watching? Tell us in the comments.
