On August 28, 2026, La Vanguardia published a story in Madrid that shifted the narrative surrounding code agents: OpenAI was designing a persistent and proactive version of Codex that would continue working until the user shut it down or put it to sleep. The original article was published by Wired (Maxwell Zeff). The full article is behind a paywall; we did not fabricate quotes from Wired. Gizmodo corroborated this on August 27: the code for Persistent mode appeared in the Codex CLI's public repository.
The news isn't that Codex "responds better." It's that OpenAI is testing a mode in which the agent doesn't stop after completing the initial request. This analysis separates what the press attributes to the code and a spokesperson from what a company can conclude today. Codex, as a product, already exists. The persistent mode doesn't: there's no public announcement or release date.
La Vanguardia, citing Wired, describes a change from current assistants, which typically stop after minutes or hours even if the task remains incomplete. By activating persistent mode, the assistant would continue active across multiple sessions until instructed to stop. The feature appears in the "reasoning effort" menu and, according to the same report, requires significant computing power and tokens for the model to "think" before responding.
An OpenAI spokesperson confirmed to Wired—as reported by La Vanguardia—that they are testing it, but will not launch it immediately. Thibault Sottiaux, head of core products, framed it this way:
"OpenAI has a very horizontal culture and many different things are explored in the open source repository, which is, in a way, our shared playground."
That statement confirms exploration. It doesn't confirm availability, plans, pricing, or who will have access. Gizmodo reached out to OpenAI and received no response before publishing: the only "yes, we're testing it" in the press on August 27 and 28 came via Wired, not a product announcement.
A discovery in a public repository is an early sign, not a release. Treat persistent mode as a confirmed prototype, not a feature you can already purchase.
What sets it apart from a chatbot isn't that the process takes longer. It's the lack of proactivity. According to the code analyzed by Wired, the agent could continue working even after responding: it would create follow-up tasks for itself, using previous interactions and its "user knowledge." It would also have a tool for sending spontaneous messages, with instructions to use them very sparingly.
That changes the contract. Today, a team invokes an agent, checks a diff, and closes the session. A persistent mode with proactivity maintains a goal beyond the prompt and can interrupt the human. Sam Altman, CEO of OpenAI, anticipated this in an interview with David Senra, quoted by La Vanguardia:
"There's a kind of single product, which is: I need to ask the AI something (...) over time, perhaps the AI should proactively offer me things." He added that this interface, which started as a chatbot and now includes programming agents, "will be perceived as a more consistent agent."
These are direction statements, not a technical specification. They align with what the code appears to be testing. They don't prove that the product already behaves this way in production.
Don't mix three layers. On its official page, Codex is the same agent in ChatGPT, the editor extension, and the CLI, all linked by the ChatGPT account. OpenAI states that it completes end-to-end tasks, supports multi-agent workflows with worktrees and cloud environments, and that Skills teach team norms. It also states, using product terminology, that it is "built for always-on background work": issue triage, alerts, or CI/CD.
That's marketing for a Codex that can be scheduled. It's not Persistent mode. A scheduled job starts when you or an event triggers it and finishes with the recipe. Persistent mode, according to Wired and La Vanguardia, stays on between sessions and proposes the next task until someone turns it off.
OpenAI has stated that over five million people use Codex every week, that over one million use it outside of development, and that Chat, Work, and Codex are available on the desktop app across all plans, including the free one. These are figures from the manufacturer. They help to understand that Codex is no longer an experiment. They don't prove that persistent mode is present in those areas.
Distinguish three things: the Codex that you can already open, the scheduled tasks that OpenAI is already announcing, and the persistent mode that Wired found and that the company says it is testing without a date.
La Vanguardia attributes the limitations to "the latest autonomous cyberattacks launched by OpenAI's advanced models." The article doesn't detail these incidents; we didn't invent victims, CVEs, or figures. It describes two restrictions of persistent mode. First: the assistant cannot unilaterally expand its capabilities on the system. Second: any modification outside the user's computer will require explicit approval.
This aligns with the documented Codex: OpenAI publishes approval pages, sandboxing procedures, and Codex Security, an agent that scans and suggests fixes. These controls don't make the prototype secure. They mean that, at least in the text found, OpenAI doesn't design a process with unlimited permissions.
Read them as design intent, not as a guarantee. An agent that stays running, invents tasks, and can send messages presents three risks that a brief chat does not: goal drift, uncontrolled computing power consumption, and actions taken when no one is looking. According to the press, the code attempts to mitigate the first and third risks. The second—the cost of a process that doesn't automatically shut down—remains as a warning within the report itself.
Security must exist outside the model: separate credentials, minimal scope tools, and a person authorizing what leaves the local computer.
La Vanguardia points out that Anthropic, Meta, and OpenAI are competing to automate everyday tasks, and that the users of these tools are still primarily programmers. That's a market analysis, not a buy order. On August 28, 2026, without waiting for a button that doesn't exist, an organization can:
In fourteen days, simply document what your agent does today when you get up from your chair and test a reversible workflow—a request, a triage, some tests—with three rules: it doesn't write outside the test repository, it doesn't communicate with external systems without approval, and someone reviews it the next day. If no one knows how to undo the action, the agent shouldn't execute it yet. Don't install a mode that isn't in the menu. Prepare the governance for when—if—it appears.
Persistent mode matters because it shifts Codex from a "tool I invoke" to a "process that stays running." Wired found it. A spokesperson confirmed the test and the lack of an immediate release. Altman already describes a more persistent agent. None of this justifies treating the prototype as if it were in production.
My recommendation is straightforward. Don't look for a Persistent mode switch or plan a project around a date OpenAI hasn't provided. Instead, review three things this week: what permissions Codex has on your machine and in the cloud, who approves changes outside the local computer, and what your team would do if the agent kept creating tasks at 3:00 AM. If you can't answer these questions, a mode that doesn't automatically go to sleep isn't an advantage: it's a bill and an incident waiting to be addressed.
The manufacturer's promise, when it arrives, will be to complete more work without you being there. The company's job is to demonstrate that the work was done correctly, that it can be stopped, and that it can be undone.
Editorial note: Quotes from Thibault Sottiaux and Sam Altman are from La Vanguardia, which attributes the research to Wired. We are not reproducing any additional quotes from Wired. Persistent mode was not announced or publicly available at the time of this note.
Si tu proyecto requiere una solución más enfocada, entra directo a la landing ideal para tu negocio y envíanos tu información en el formulario correspondiente.
0 Comentarios