Paragraph 1: The Shift from Asking to Delegating (and the Leak)
For years, the most frustrating part of using an AI assistant has been its invisible, arbitrary stopwatch. You ask it to research a topic, draft a report, or fix a bug, and it gives you an answer, then goes blank—idle, awaiting your next command)Skip.
Paragraph 2: The “Reasoning Effort” Menu and the Definition of Persistent (Detailed Explanation of the feature)
Paragraph 3: The “Proactivity” Layer (How it creates its own tasks, works across sessions, and messages you).
Paragraph 4: Safety Limits and Platform Expansion (The guardrails, the shared core, implications for ChatGPT).
Paragraph 5: Sam Altman’s Vision and the Business Race (The competitive landscape, driving adoption of premium models).
Paragraph 6: The Existential Shift and Industry Context (AI as a coworker vs. a tool; what “put to sleep” means for human oversight; the future of digital companionship/interns).
Let’s calculate word counts as I draft. I’ll aim for around 333 words per paragraph.It’s a familiar frustration for anyone who has ever used a modern AI chatbot: you ask it to do something complex—read through a dense legal document, organize a chaotic folder of files, or generate a comprehensive report—and it stops halfway. The model hits an invisible wall, an arbitrary token limit, or a computing threshold, and politely tells you it’s done, even though the task is clearly unfinished. You then have to copy-paste, re-prompt, and baby-step the AI toward completion. This reactive, stop-and-go model has defined our experience with artificial intelligence since its public debut. But that is all about to change. Earlier this week, developers combing through the open-source codebase of OpenAI’s advanced coding agent, Codex, discovered the blueprints for a massive philosophical shift: a feature internally dubbed “Persistent mode.” This isn’t just a higher processing setting; it’s a fundamentally new way of working. The code explicitly instructs the model to “continue working until put to sleep,” a stark departure from the current paradigm where an AI finishes its batch of thinking and politely hands the digital reins back to the human. It marks the moment Silicon Valley stops building a reactive tool and starts building an autonomous employee. The implications are huge: an AI that doesn’t stop when it’s answered, but keeps going, keeps checking, keeps iterating—like a digital ghost in the machine, working through the night so you don’t have to.
The feature has been spotted in the codebase of Codex, OpenAI’s specialized engineering agent, located in the “reasoning effort” menu. This menu is essentially how users dial up or down how much “thinking” a model does before it responds. Currently, users can choose between low, medium, or high effort, which dictates the number of tokens (the AI’s computational “brain fuel”) it burns through. Persistent mode is emerging as a new, far more intensive option. Reading the code, when a user selects this, the system instructs Codex to “continue working until put to sleep.” This is a colossal departure from the status quo. In today’s architecture, an AI agent burns through its context window—its working memory—and hits a wall, often stopping mid-task because it has run out of compute or conversational bandwidth. Even if a task is 90% complete, the agent stops, waiting for the human to tell it to continue. Persistent mode essentially removes that stop-gap. It is programmed to work autonomously, potentially for hours or days, looping back on itself to check for errors, retry failed API calls, and allocate more compute as needed. It doesn’t just complete the initial request; it is told that its work isn’t finished until the underlying goal is achieved. For software engineers, this means handing over an architecture overhaul and coming back the next morning to a completed pull request, rather than an unfinished half-step.
Beyond just not stopping, the code describes a secondary layer called “proactivity.” This is arguably the more unsettling and revolutionary aspect of the update. Within Persistent mode, there is a system prompt that fundamentally changes the agent’s operating script. Instead of a standard “finish the task and output a result” directive, the agent is told that once it answers a user’s request, its job is not done. It is explicitly instructed to scan the user’s context, chat history, and connected files to generate its own “follow-up tasks.” It actively seeks out problems to solve without being prompted. For example, if you ask it to draft an email to a client, it might draft it, and then, using its proactivity, decide to look at your calendar to see when you have a meeting with that client, prompting it to update the meeting notes or draft a prep document for that future meeting. It works across sessions, meaning it remembers the context from a previous conversation and synthesizes that knowledge to continue working. The code even grants it a tool to message the user unprompted, though it is explicitly instructed to use this sparingly, begging the question of how a machine decides what is “sparingly” versus “annoying.”
Humanizing this, imagine you have a personal assistant who sees the next logical step in a project without you having to ask. If you use Persistent mode to organize your expense report, the agent doesn’t just sort the receipts and stop. It notices you have a trip planned for next month, so it pre-emptively creates a new expense folder related to that trip. It might even message you to ask for the hotel confirmation number. This level of atomic autonomy is the holy grail of “AI as a coworker.” However, this immense power necessitates strict guardrails, which OpenAI is clearly aware of. Within the code base, the instructions for Persistent mode contain strict safety overrides. The agent is told, bluntly, that while it can work indefinitely, it does not have expanded permissions. It cannot access external databases or alter files outside the user’s designated workspace. More critically, the code explicitly states that if the agent wants to take an action that affects the user’s system or environment outside of its sandbox—like sending an email without confirmation or modifying an external account—it must prompt the user for approval. This suggests OpenAI is trying to build a “bull in a china shop” that has been carefully tamed to ask permission before knocking anything over JavaScript.
The placement of this code is highly telling. Persistent mode and proactivity are not buried within the terminal-only version of Codex; they are located in the shared core of the system. This means OpenAI is designing this not just for command-line savvy programmers, but as a foundational feature for its broader consumer products, like ChatGPT. This aligns perfectly with Sam Altman’s recent public talks, where he has described his desire to turn ChatGPT from a browser tab you open into an ambient, always-on co-pilot for your entire digital life. The strategic imperative is clear. Currently, the most advanced proprietary models are used by a fraction of a percentage of the global population—primarily developers. To justify the billions in compute spending, OpenAI needs everyday consumers to use these features. A persistent, proactive agent that can autonomously handle tedious chores—drafting emails, organizing files, booking bookings, occasionally pinging you with a suggestion for a venue you were just talking about—is the perfect use case to drive adoption of their premium tier, OpenAI o1 and o3. In the race against Meta and Anthropic, who are also building similar autonomous agents, this persistent work ethic could be the killer feature that shifts the consumer market.
Ultimately, this is a philosophical shift as much as a technical one. The current paradigm treats AI as an oracle—you ask, it answers. Persistent mode signals a shift to treating AI as an employee. It is designed to have its own internal “to-do list” and to be self-motivated. While the safety guardrails prevent it from causing chaos, the fact that it can message the user unprompted and create its own work tasks is a massive leap toward true agency. In the short term, this feature will likely be rolled out to super-heavy users and enterprise clients, those willing to pay a premium for the always-on compute. But the longer-term implication is a world where the software doesn’t just help you use a computer; it uses the computer for you methodically, overnight, in the background. The conversation about AI taking jobs becomes less about “it wrote my code in 10 minutes” and more about “it works for 10 hours independently while I sleep.” The race between Meta, Anthropic, and OpenAI is not just about who can answer questions best anymore; it is about who can build the first truly autonomous, always-on digital employee that you trust enough to let work through the night. The code suggests Altman’s vision is closer to reality than we thought, and our relationship with software is about to shift from asking a question to managing a workforce.
