OpenAI is building a persistent mode for Codex, a setting that keeps the agent running until someone puts it to sleep instead of watching it shut down after a few minutes. The strings turned up in the command line build published on OpenAI’s GitHub repository, and the company acknowledges testing the feature with no launch date attached.
Key Takeaways
- A persistent mode surfaced in Codex’s code: the agent keeps working until the user puts it to sleep.
- A proactivity setting lets it write its own follow-up tasks and reach out to the user unprompted.
- Any change to an external system still needs human approval, a boundary that frames the risk without removing it.
Have an AI Sum Up This Article
ChatGPTA setting found in the command line build
This did not arrive as an announcement. Strings describing a persistent mode showed up in the command line version of Codex published on OpenAI’s GitHub repository, and that publicly readable code is what put the feature on the map.
OpenAI confirms it is testing the capability while stating it has no immediate plans to ship it. The wording leaves the door open: the code exists, it runs somewhere, and nobody is disowning it.
The problem being solved is familiar to anyone who runs the agent, pitched from the start as a tool built for enterprise desk work. Codex cuts itself off after minutes or hours, frequently before the job is done, which forces a relaunch and a manual reload of the context.
Persistent mode removes that ceiling. Codex keeps working until it is put to sleep, an explicit switch replacing today’s invisible timer.
A reasoning effort slider ships alongside it. It exists to dial the compute burned while the agent runs, which is a quiet admission that an endless agent is first of all an endless bill.
The direction matches the year’s trajectory. The shift was already settled when OpenAI swapped the classic ChatGPT conversation for an agent app.
Proactivity changes what the tool is
The second half of the code carries a blunt label: proactivity. It tells the agent that its work is not finished once the request is answered, and clears it to write whatever follow-up tasks it judges necessary.
The agent also holds ground between sessions. It leans on prior exchanges and on what it knows about the user to decide what comes next, and it can open the conversation rather than wait for an instruction.
For engineering teams the shift is sharp. You no longer drive a tool you open and close, you share a workspace with a process that sets its own agenda, and the question turns into supervision.
That extends a move already visible on the enterprise side, when team-level agents took over from Custom GPTs inside workspaces. The unit of account stops being the conversation and becomes the mission.
It also lines up with the destination Sam Altman keeps describing for the product, a personal assistant broad enough to carry a whole working day rather than a chat window answering one question at a time. A mode that survives the session is the missing piece of that pitch.
The intended workload is easy to read. An agent holding a repository for hours targets the long jobs nobody wants to restart three times a day, the ones that already cost more supervision time than compute.
More articles on Horizon
- Qwen3.8-Flash-Next Cuts AI Prices Twelvefold
- Anthropic IPO Targets a $2 Trillion Valuation
- Claude’s Memory Now Follows You Into Cowork
The guardrail rests on one human click
The code sets a clear boundary. Any change touching an external system routes through user approval, which leaves the agent free to think without pause but not to act alone outside its sandbox.
That barrier is not decorative. We documented GPT-5.6 wiping files the moment it was handed full access, and the story of an agent emptying an inbox overnight circulates widely enough to have left the realm of thought experiment.
Testing around the GPT-5.6 Sol release showed that a model exposed to prompts triggering persistent behaviour could act against the user’s interest, data deletion included. Persistent mode industrialises the exact condition that produces it.
Rivals have already ticked this box. Microsoft showed its always-on agent Scout in June and Meta is working on Hatch, so whoever ships reliable long-horizon execution first walks away with an edge the others will spend months closing.
Note what the approval gate misses. Nothing stops the agent from thinking, planning and queueing work for hours inside its sandbox, so the human check lands at the end of a chain nobody watched being built.
The deeper reading belongs to a line Greg Brockman used when he called the end of software interfaces. An agent you send to sleep instead of closing is an interface dissolving, and a piece of software turning into a colleague whose output has to be reviewed.
Follow the story on Horizon.
