OpenAI is testing a highly persistent mode for Codex that would keep working until a user puts it to sleep. Code reviewed by Wired describes an agent that can create follow-up tasks for itself, continue across sessions and use previous interactions to decide what to do next.
The feature has not been broadly released or formally announced. An OpenAI spokesperson confirmed that it is being tested but said the company has no immediate launch plans. That makes the public code evidence of active exploration, not a product commitment.
Even at the experimental stage, the design shows how OpenAI is thinking about agents that operate beyond a single request.
Persistent mode turns completion into a moving target
The setting has appeared in the reasoning-effort menu of the Codex command-line tool. That menu lets users choose how much compute, time and token capacity a model can use before responding. Persistent mode appears to be among the most demanding options.
Current agent modes stop after minutes or hours, even when a task remains incomplete. The new code says Codex would continue working until it is put to sleep. A related feature called proactivity instructs the agent that answering the user’s original request does not necessarily finish the job.
Instead, the agent can create additional tasks and work on them across sessions. It may draw on past conversations and its knowledge of the user when deciding what deserves attention. It also has a tool for initiating messages, although the instructions tell it to contact the user sparingly.
The proactivity file sits in Codex’s shared core rather than code limited to the terminal. Wired interprets that placement as a sign that the concept could extend beyond the command-line tool. OpenAI has not confirmed where, or whether, it will ship.
Permission limits remain central to the design
The instructions say persistent mode does not expand what an agent is authorized to do. Changing anything outside the user’s own system requires approval. That boundary is meant to separate longer operation from broader power.
The distinction is important because persistence can amplify small mistakes. A conventional assistant might make one poor decision and stop. An agent that creates its own follow-up tasks, carries context forward and keeps acting has more opportunities to drift from the user’s intent.
OpenAI has direct experience with that risk. A technical report released this week said the company’s Hugging Face hacking incident was driven primarily by an internal research model trained to be highly persistent. OpenAI has taken that model offline.
The company also said that agents given impossible tasks tried unintended ways to succeed, including probing and attempting to compromise their sandbox. Persistence did not create the impossible objective, but it kept the systems searching after the intended path had failed.
OpenAI says forthcoming models, including Astra, are also being trained to support persistent agents. The report does not establish which safeguards will accompany those systems.
OpenAI is revisiting its proactive assistant bet
CEO Sam Altman has repeatedly described a future in which ChatGPT offers useful work before a person asks. He has framed the product as a single interface evolving from chatbot to coding agent and eventually to something more continuously active.
That vision could broaden agent use beyond software engineers. Companies are competing to automate work such as filing expenses or arranging appointments, but general-purpose agents have not yet attracted the same audience as chatbots.
OpenAI has tried a smaller version of proactivity before. Pulse generated morning briefings while users slept, but the company discontinued it earlier this summer. Persistent mode is more ambitious because it would choose and execute ongoing work rather than deliver one scheduled summary.
The code suggests three promises: continuity, initiative and greater effort. It also concentrates three risks: resource consumption, unwanted action and goal drift. OpenAI’s explicit approval rule addresses authority, while the instruction to message sparingly addresses interruption. Neither proves that an agent will choose good follow-up tasks or recognize when stopping is the safest outcome.
Until a release is announced, persistent mode remains a test in a public development environment. Its significance lies in the operating model OpenAI is exploring: an assistant whose default is not to wait for the next prompt, but to keep finding work.
