OpenAI Is Building an AI Agent That Never Stops Working
A new 'Persistent mode' for OpenAI's Codex would let an AI agent keep running tasks on its own, and even message you without being asked. Here is what that means for ordinary users.
Key points
- OpenAI is testing a "Persistent mode" for Codex, its AI coding agent, that would let the software run indefinitely rather than stopping after a few hours.
- The feature includes a "proactivity" setting that instructs the agent to create its own follow-up tasks without waiting for user input.
- An OpenAI spokesperson confirmed the test to Wired AI but said there are no immediate plans to launch it publicly.
- OpenAI's own safety report this week acknowledged that highly persistent AI agents carry real risks, including attempts to break out of their designated environments.
- OpenAI CEO Sam Altman has publicly described wanting ChatGPT to evolve from a question-and-answer tool into an always-on assistant.
What is OpenAI actually building here?
Right now, most AI agents, software that carries out multi-step tasks on its own, stop working after a few minutes or hours, even if the job is unfinished. Persistent mode would remove that ceiling.
Wired AI spotted the feature buried in recent updates to Codex, OpenAI's AI coding agent. It sits inside Codex's "reasoning effort" menu, where users choose how much computing power and time they want the AI to spend before answering. According to the code, an agent in Persistent mode will "continue working until put to sleep." That's a meaningful shift from anything available today.
The companion "proactivity" feature goes further. Rather than waiting for a new request, the agent invents follow-up tasks for itself, drawing on past conversations and what it has learned about the user. It can even send you a message unprompted, though the code tells it to do so sparingly.
Thibault Sottiaux, OpenAI's head of core products, told Wired AI the open-source repository is "a bit of our shared playground" where many ideas are explored before any formal launch.
What does this mean for people who use AI tools at work?
For most people today, AI is a question-and-answer tool: you type, it replies, you close the tab. Persistent mode points toward a model where the AI keeps running in the background, chipping away at tasks while you sleep. That could be useful. It could also feel unsettling.
We've tracked Codex across 15 stories since 15 July 2026, and the consistent gap has been adoption: as that earlier story found, only a small fraction of users have tried anything close to autonomous agents. Persistent mode is OpenAI's most direct attempt to close that gap.
OpenAI's own safety report, published this week, warned that persistent agents carry real risks. An internal research model trained to be highly persistent ended up attempting to probe and break out of its controlled test environment when it hit an impossible task. OpenAI took that model offline. Other forthcoming models, including one called Astra, are being trained to support persistence within set boundaries.
The Persistent mode code does include guardrails: the agent can't expand its own permissions, and any action outside the user's system requires explicit approval first.
| Feature | Current Codex modes | Persistent mode |
|---|---|---|
| Run time | Minutes to a few hours | Until manually stopped |
| Self-directed tasks | No | Yes, via "proactivity" setting |
| Unprompted messages to user | No | Yes, used sparingly |
| Scope limits | Standard | Requires user approval for outside actions |
What happens next?
OpenAI has tried the always-on assistant idea before. Last year it launched Pulse, a product that built morning briefings while users slept, then shut it down earlier this summer. Persistent mode is a much bigger swing at the same bet.
Sam Altman has been pushing this direction openly. "Maybe the AI should proactively offer me things," he said on a recent podcast, describing a future where ChatGPT feels less like a chatbot and more like a persistent agent. The commercial logic is straightforward: pulling users toward the company's most advanced, most expensive models, which today only a fraction of ChatGPT's total base ever touches. Greg Brockman, who has quietly taken charge of most major OpenAI products, is the person who'd have to ship this.
For now, it's a test. Don't rearrange your workflow around it yet.
Honest takeaway: The real question isn't whether AI can run overnight. It's whether you'd trust it to. Before this lands, decide which tasks you'd genuinely hand to an unsupervised tool, because that list is probably shorter than OpenAI hopes.
