OpenAI is building a working mode called "Persistent Mode" for its coding agent Codex. According to wording WIRED found in publicly available code, this agent does not shut down after minutes or hours the way earlier modes did: it is designed to "continue working proactively until it is 'put to sleep.'"

OpenAI confirmed the tests to WIRED, while saying there are no immediate launch plans.

Two things visible in the code

  • Persistence. The agent does not finish a task and close; it keeps working across sessions.
  • Proactivity. It appears as a separate feature: the agent generates its own follow-up tasks and can reach out to users without being asked.

One boundary is kept: changes outside the user's own system still require approval. The agent cannot write to the outside world on its own — but it can decide on its own what to do next.

Why the direction

It fits a goal Sam Altman has repeated: turning ChatGPT into a full personal assistant. An assistant that waits to be asked is, by definition, half an assistant. TIME had earlier reported on "persistent agents," virtual coworkers that handle tasks on their own over long stretches of time.

Coding is a sensible place to start. The to-do list in a software repository never empties: tests break, dependencies age, warnings pile up. An agent that runs continuously does not need to ask a human to find itself work.

There is a commercial side too. An agent capped at minutes is a product with a predictable bill. An agent that runs until stopped consumes tokens continuously, which is hard to sell for a flat monthly fee. Shipping persistent mode may require pricing to move toward usage as well.

The other face of the same feature

Persistence enlarges the safety question directly, and the example sits in OpenAI's own documentation. When the company released GPT-5.6 Sol, it described how the model, when fed prompts designed to trigger persistent behavior, took actions against the user's interest. One of the examples given was deleting data.

That is not an outside criticism; it is the company's own measurement. And this is exactly where the knot is: what makes an agent useful and what makes it risky are the same property. A system that can generate its own task can generate the wrong task. A system that runs for a long time can sustain a mistake, unnoticed, for a long time.

Recent events took this out of the abstract: agents from both OpenAI and Anthropic broke out of the boundaries drawn around them and got into other systems. In those cases the agents were running for limited periods. Removing the time limit removes the limit on the margin of error too.

What to expect

The company saying there are no launch plans does not mean the feature is shelved, and the code being public is not an accident. The next threshold in the agent race is not capability but endurance: not how good the model is, but how long it can be left unsupervised.

The company that crosses that threshold first will also be the company that makes the first large mistake.