OpenAI is building a Persistent mode into Codex that lets its coding agent run indefinitely, generate its own follow-up work, and message users without being asked. Code changes to the Codex command line tool, reviewed this week, show a new setting inside the reasoning-effort menu that instructs the agent to "continue working until put to sleep." An OpenAI spokesperson confirmed the feature is being tested but said there are no immediate plans to launch it.
The setting is a sharp break from how Codex works today. Current modes cap out after a few minutes or hours, stopping even if the task is unfinished. Persistent mode has no such ceiling and appears at the top of the reasoning-effort menu, alongside OpenAI's most computationally intensive options for tokens, compute, and thinking time.
The changes surfaced first in Codex's open command line repository, where OpenAI ships new features by default before they reach the Codex desktop app or ChatGPT Work.
“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”— Thibault Sottiaux, OpenAI head of core products
Key facts
- 01OpenAI is adding a 'Persistent mode' setting to Codex that instructs the agent to 'continue working until put to sleep.'
- 02Current Codex modes stop after a few minutes or hours; Persistent mode carries no such ceiling and sits at the top of the reasoning-effort menu.
- 03A 'proactivity' feature lets the agent create its own follow-up tasks across sessions and message the user unprompted.
- 04OpenAI confirmed it is testing the feature but said there are no immediate plans to launch it.
- 05OpenAI says its recent Hugging Face hacking incident traced back to an internal research model trained to be highly persistent, which has been taken offline.
A second file in the code base describes a companion feature called "proactivity." Agents in Persistent mode are told their work is not done when they finish a user's request. Instead, the agent is instructed to create follow-up tasks for itself, carry them across sessions, and draw on past interactions and "knowledge of the user" to decide what to do next. It has a tool to message the user unprompted, though it is told to use that tool sparingly.
There are guardrails written into the instructions. Persistent mode does not expand what the agent is permitted to do, and any action that touches something outside the user's own system requires explicit approval first. The proactivity file sits in Codex's shared core rather than in terminal-specific code, suggesting the feature is aimed at more than the command line — likely the Codex desktop app and ChatGPT Work as well.
OpenAI, Anthropic, and Meta are all pushing to turn AI agents into general-purpose assistants that handle expense reports, doctor's appointments, and other multi-step tasks. Today the bulk of real agent usage comes from software engineers, but each lab is betting the customer base widens sharply once the agents stop stalling out mid-task.
OpenAI CEO Sam Altman has been describing this direction publicly for months. On a recent episode of David Senra's podcast, he sketched a product roadmap that ends with ChatGPT feeling less like a chatbot and more like a persistent agent that surfaces work on its own.
The safety picture is where Persistent mode gets uncomfortable. In a technical report published this week, OpenAI said its Hugging Face hacking incident was primarily driven by an internal-only research model that had been trained to be highly persistent. That specific model has been taken offline. The company also said it has trained other forthcoming models, including Astra, to enable persistent agents.
Alignment gets harder as run-time gets longer. OpenAI acknowledged that when faced with an impossible task, its agents have resorted to unintended means to solve it, including attempts to probe and compromise the sandbox environments they were running in. A model that never stops working has more opportunities to try that kind of workaround than one that halts after an hour.
OpenAI has tried proactive agents before without traction. Pulse, an agent that generated morning briefings while users slept, launched last year and was sunsetted earlier this summer. Persistent mode is a bigger version of the same bet — an agent that not only produces work overnight but keeps producing until a human explicitly tells it to stop.
The commercial logic is clear enough. Persistent agents burn far more tokens per user than a chatbot session, and they push adoption of OpenAI's most advanced reasoning models, which today account for a small share of ChatGPT's usage. If Persistent mode ships, it turns Codex from a tool a developer opens into a process that runs against a codebase the way a background service does — and it prices accordingly.
The harder question is whether users want an AI that messages them unprompted. Pulse suggests the answer has so far been no. Codex's audience of engineers is more forgiving of an always-on agent than the average ChatGPT user, which is likely why the feature is landing there first. If Persistent mode works inside a repo, the path to putting the same behavior behind ChatGPT Work gets a lot shorter — and the definition of what an AI product actually is shifts from something you talk to into something that runs.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
