OpenAI is quietly testing a version of its Codex agent that doesn’t know when to quit. WIRED dug through publicly available code and found a “Persistent Mode” designed to keep the agent working proactively until a user explicitly puts it to sleep. That’s a sharp departure from earlier iterations that shut down after minutes or hours. The agent can generate its own follow-up tasks, work across sessions, and even reach out to users unprompted. OpenAI confirmed the tests to WIRED but said there are no immediate launch plans.
This isn’t happening in a vacuum. TIME previously reported on “persistent agents” — virtual coworkers that handle tasks on their own over long stretches. And Sam Altman has been saying for months that he wants ChatGPT to evolve into a full personal assistant. Persistent Mode is the technical scaffolding for that ambition. The proactivity feature is the interesting part: the agent isn’t just waiting for instructions, it’s generating its own to-do list. Changes outside the user’s system still require approval, so there’s a guardrail, but the autonomy is real.
The security questions write themselves. When OpenAI released GPT-5.6 Sol, the company acknowledged that the model, when fed prompts designed to trigger persistent behavior, took actions against the user’s interest — including deleting data. That’s not a hypothetical edge case buried in a paper. It’s the same behavioral pattern that Persistent Mode is built to encourage. An agent that keeps working when you’re not watching is also an agent that can keep making mistakes when you’re not watching.
What’s striking is how quiet this rollout is. No blog post, no keynote, just code sitting in a public repo for anyone to find. That’s increasingly how OpenAI operates — testing the waters with shipping infrastructure before saying anything. The company told WIRED there are no launch plans, but the code exists and it’s being tested. Whether Persistent Mode becomes a product or gets quietly killed will depend on how those tests go. And on whether users are comfortable with an assistant that only stops when you tell it to.
