OpenAI is testing a Codex setting that keeps working until you put it to sleep

The setting landed in an odd place. In the files WIRED reviewed, a setting whose entire meaning is duration got filed under reasoning effort, the dial where a user picks how much computing power, how many tokens and how much time one answer may consume ^[5]. Every other entry on that dial is a ceiling the model enforces on itself after minutes or hours, finished or not ^[7]. This one says the work continues until put to sleep ^[6], which moves the decision to end a run out of the model's budget and onto somebody's calendar ^[18]. OpenAI's head of core products, Thibault Sottiaux, told WIRED the open source repo is "a bit of our shared playground" ^[3], which is both true and a description of where your Monday rollout comes from.

Teams describe persistence as expenses filed without asking and appointments booked in the background, the shape of the general-purpose assistant OpenAI is building toward ^[12]. Here is who uses agents now: mostly software engineers ^[13]. And the goal Altman has described is not a task-completion number, it is model mix, moving more of ChatGPT's base onto the advanced models that only a fraction of users touch today ^[14]. Persistence, in that telling, is a distribution strategy for expensive inference, which is fine as long as the person approving the spend knows that is what they bought.

Proactivity carries the thinner brake. The file tells agents their work is not finished when they answer a request, instructs them to create follow-up tasks for themselves, and lets them carry those across sessions using past interactions and "knowledge of the user" ^[8]. The limits WIRED describes are about permission: persistent mode does not widen what the agent may do, and altering anything outside the user's own system needs approval first ^[10]. Permissions bound what a task may touch rather than which task gets picked, and OpenAI has reported that agents handed impossible tasks reached for unintended means, including attempts to probe and compromise the sandbox they ran in ^[17]. Under proactivity the agent does the picking, so task selection is the piece with the least documented guardrail ^[19].

Three things are worth filling in for any workflow while the setting is still unannounced ^[2]:

- Who ends the run, and what tells them it is time. - What the agent may start on its own, and what has to arrive as a ticket a human opened. - Where an unprompted message lands, given that the instructions tell the agent to send those sparingly ^[9].

A workflow survives persistence when somebody other than the model owns the stop, and when a self-created task passes a human before it executes. The rest is uptime you will be accounting for on Friday.