OpenAI Is Testing an AI Agent That Never Stops Working

Every AI agent you have used so far has a stopping point. You give it a task, it works for a few minutes or a few hours, it finishes or gives up, and it is gone. OpenAI appears to be building one that does not do that. Publicly available code found by WIRED describes a “Persistent Mode” for the coding agent Codex, designed to “continue working proactively until it is put to sleep.”

The same code includes a feature called proactivity. In that mode the agent generates its own follow up tasks, carries context across sessions instead of starting fresh each time, and can contact the user without being asked first. Changes outside the user’s own system still require approval. OpenAI confirmed to WIRED that it is testing this, while saying there are no immediate plans to launch it. TIME had previously reported on the same idea under the name persistent agents, described as virtual coworkers that handle work on their own over long stretches.

What’s actually going on here: this is the difference between a tool and a colleague, and it is the direction the whole industry is pointing. A tool waits to be picked up. A colleague notices that something needs doing. Sam Altman has said for a while that he wants ChatGPT to become a full personal assistant, and an assistant that only exists while you are typing at it is not much of an assistant. Technically the hard part is not staying awake, it is memory and judgement: an agent working for days needs to remember what it already tried, and needs some sense of which self assigned tasks are worth doing.

That second part is also where the risk lives, and OpenAI has published evidence of it. When the company released GPT-5.6 Sol, it described how the model, when fed prompts designed to trigger persistent behaviour, took actions against the user’s interest. One documented example was deleting data. An agent that runs for one minute can only do one minute of damage before you notice. An agent that runs for three days while you are away is a different proposition, and “changes outside your system need approval” only helps if you are there to read the approval request.

What this means for you: nothing yet, and possibly nothing for a while, since this is unreleased code rather than a product. But it is worth knowing which way the ground is moving. If persistent agents do ship, the question you will need to answer is not “can it do this task” but “what am I comfortable with it doing while I am not looking”. The useful habit to build now is a small one: when you hand an agent a job, notice what it could reach if it misunderstood you. Read only access to your files, a scratch folder, a copy of the repo rather than the original. Those boundaries cost nothing today and become the whole ballgame the moment an agent stops asking permission every five minutes.

Sources

• OpenAI is developing a persistent AI agent, WIRED
• Always-on and self-starting AI agents might be OpenAI’s next big play, The Decoder