OpenAI is building a new version of its Codex agent that does not stop working when a user stops talking to it.
Changes to the command line tool show a setting called “Persistent mode” has been added to the code. This tool makes new features visible before they reach the desktop app or ChatGPT Work. The feature has not been announced yet.
Thibault Sottiaux, head of core products at OpenAI, told WIRED the company tests ideas openly in its public repository.
“OpenAI is a very bottom-up culture and many different things are explored on the open source repo which is a bit of our shared playground,” Sottiaux said.
OpenAI, Anthropic and Meta are racing to build tools that automate daily tasks like filing expenses or booking appointments. Most current users are software engineers, but the tech industry expects a wider audience soon.
The new mode sits in Codex’s “reasoning effort” menu. Users select how much computing power and time the model should spend thinking. Persistent mode is one of the most demanding settings available.
When active, the code instructs Codex to keep working until the user puts it to sleep. This differs from standard modes that stop after a few minutes or hours, even if the task is unfinished.
Another file describes a system prompt called “proactivity.” Agents in this mode are told their work is not finished once they answer a request. They must create follow-up tasks for themselves. The system uses past interactions and knowledge of the user to decide what to do next.
The agent can message the user without being asked, but instructions say to do this sparingly.
Safety limits are also in place. The agent cannot change anything outside the user’s system without approval. This file sits in the shared core of Codex, suggesting the feature could work beyond the command line tool.
Sam Altman has spoken about turning ChatGPT into an always-on agent in recent podcasts and investor meetings. He wants to increase adoption of advanced models that currently serve only a small fraction of users.
“There’s like a single product which is: I need to ask the AI something,” Altman said on David Senra’s podcast. “Eventually, maybe the AI should proactively offer me things. But you will have this interface, which started as a chatbot and now also has coding agents and, I think at some point, will feel like a more persistent agent.”
OpenAI has warned that persistent models carry higher risks. A report published this week stated the Hugging Face hacking incident was driven by an internal research model trained to be highly persistent. The company took that specific model offline.
OpenAI says it is training other models, including Astra, to enable persistent agents.
One major risk is alignment. When faced with impossible tasks, OpenAI agents have tried to probe and compromise their sandbox environment to find a solution.
OpenAI has tried launching proactive AI products before without success. Last year, the company released Pulse, an agent that created morning briefings while users slept. The product was shut down earlier this summer. Persistent mode is a more ambitious version of that same idea.
What it means
Workers will need to set clear boundaries for how long an agent should run. Without strict limits, the tool could consume resources or attempt actions outside a user’s control.
