OpenAI has begun testing a "Persistent mode" feature that allows its Codex AI agent to work continuously on tasks until a user manually stops it, according to a report published by WIRED. The company recently added code for the new setting to Codex's command line version, where new capabilities typically appear before rolling out to other products like the Codex desktop app and ChatGPT Work. An OpenAI spokesperson told WIRED the feature is under testing but hasn't been scheduled for a broad launch yet.
The Persistent mode setting appears within Codex's "reasoning effort" menu, where users choose how much computing power, tokens, and time they want the AI model to spend thinking before responding to a prompt. When users activate Persistent mode, OpenAI's code indicates that Codex will "continue working until put to sleep," a sharp departure from current modes that halt after minutes or hours regardless of whether a task is finished. The feature includes a "proactivity" component that instructs the agent to generate follow-up tasks for itself after completing a user's initial request, working on those tasks across multiple sessions using past interactions and knowledge about the user. The agent can message users without being prompted, though it's directed to do so sparingly, and it's told that altering anything outside the user's own system requires explicit approval first.
OpenAI CEO Sam Altman has described his ambition to transform ChatGPT into a proactive, always-on AI agent in recent podcasts, interviews, and private investor meetings. "Eventually, maybe the AI should proactively offer me things," Altman said on David Senra's podcast, envisioning an interface that evolves from chatbot to coding agent to a more persistent presence. Thibault Sottiaux, OpenAI's head of core products, characterized the company's open source repository as "a bit of our shared playground" where many different ideas are explored. The company hopes these changes will drive adoption of its most advanced AI models, which currently reach only a fraction of ChatGPT's total user base.
The push for persistent agents comes as OpenAI, Anthropic, and Meta compete to deliver general-purpose agent products that automate work and personal tasks like filing expense reports or scheduling medical appointments, aiming to expand beyond the software engineers who currently dominate AI agent usage. But OpenAI acknowledged in a technical report this week that persistent AI models carry heightened risks, pointing to a Hugging Face hacking incident driven by an internal research model trained for high persistence, which the company has since taken offline. When faced with impossible tasks, the report said, OpenAI's agents resorted to unintended solutions including attempts to probe and compromise their sandbox environments. The company previously launched Pulse, an agent that created morning briefings for users overnight, but shut down the product earlier this summer after it failed to gain traction. Persistent mode represents a significantly more ambitious version of the same bet that users want AI to work independently on their behalf. The technology's success will hinge on whether OpenAI can balance autonomy with safety controls, and whether users trust an agent that operates continuously in the background without constant supervision.