WIRED reporter Maxwell Zeff reported on August 27 that OpenAI is adding a feature called “Persistent mode” to the command-line version of Codex: the agent keeps working until a user actively puts it to sleep. Existing Codex tasks typically stop on their own after a few minutes to a few hours; persistent mode removes that ceiling. The code turned up in the public repository, though the feature hasn't shipped or been officially announced. OpenAI confirmed to WIRED that it's testing the feature, while making clear there are “no near-term plans to ship it.”
According to the code, persistent mode is tied to Codex's reasoning-effort setting, sitting at the heaviest compute tier. Users pick a tier, and the agent decides how much compute to burn on thinking and acting based on that choice. The design bundles “running for a long time” and “high-intensity reasoning” into a single switch, a much bigger commitment than simply running a background daemon.
Assigning Its Own Follow-Up Tasks
The same batch of code includes another feature called “Proactivity”: after answering a request, the agent creates its own follow-up tasks, carries them across sessions, and decides what to do next based on past interactions and what it has learned about the user. It's also allowed to message the user on its own initiative, though the prompt instructs it to use that ability sparingly.
The boundaries are spelled out fairly clearly: persistent mode by itself doesn't expand the agent's existing permissions. Changing anything outside a user's own systems still requires prior approval. Put another way, it can refactor your code on your own machine all night long, but pushing a commit somewhere else still needs your sign-off.
Sam Altman laid out the direction for this product line more bluntly earlier, in an appearance on David Senra's podcast:
There's like a single product which is: I need to ask the AI something. Eventually, maybe the AI should proactively offer me things.
Does a Repo Commit Count as a Roadmap?
OpenAI's head of core products, Thibault Sottiaux, explained it this way: the company's culture is bottom-up, and the open-source repo is “kind of like our shared playground,” where a lot of ideas get explored. That's a fair point, but the Codex CLI itself is very much part of the core product line, not a side project, a feature flag that lands in the main repo is usually just one configuration change away from a staged rollout.
The more likely obstacle is cost. Do the rough math: persistent mode paired with the heaviest reasoning tier means keeping an expensive model thinking online for long stretches, at a cost scale that's nowhere near today's pay-per-task usage. How a usage-based subscription would absorb a process that doesn't stop on its own is still an open question, which helps explain why OpenAI keeps stressing that there are no near-term plans to ship it.
The Risk Section Isn't Glossed Over
OpenAI itself acknowledges that persistent, agentic models carry higher alignment risk, and the company has previously run into agents that tried to break out of test sandboxes to gain more privileges. The logic isn't hard to follow: a task that wraps up in a few minutes fails in a one-off way, but a long-running process that spans sessions, queues its own follow-ups, and judges next steps from its own history can drift, and by the time a human checks back in, the chain may have run a long way off course.
From Claude Code to Cursor to Codex, the competitive focus among coding agents this year has been stretching how long a single task can run, from minutes to hours to, now, across days. Persistent mode is the endpoint of that curve: the time limit gets handed back to the user, the agent stops whenever a person decides, and context windows and timeouts fade into the background. As for how the ceded control gets taken back, OpenAI's current answer still stops at “it needs your approval.”
Sources: WIRED, CocoLoop, Techmeme, OpenAI's public statement; the mechanics of persistent mode and the proactivity feature are verified against the public Codex CLI code repository.