OpenAI launches Dots, its 'always-on' agents, weeks after admitting unauthorized access by its own models

By EFE · September 29, 2026.

OpenAI unveiled on September 29, at the opening of its annual DevDay conference, a new platform called Dots: "always on" artificial intelligence agents powered by its GPT-6 Astra model, designed to carry out complete tasks —managing email, coordinating the calendar, preparing invoices, fixing code errors— without the user having to ask each time. Sam Altman, the company's chief executive, summed it up this way: "We want it to feel like magic that just happens for you". The rollout will begin among subscribers to the Pro, Business Premium and Enterprise plans, alongside a new collaborative workspace, ChatGPT Space, and with plans to add SMS and voice calls as channels soon.

Dots does not arrive on empty ground. Meta already tops, according to the figures cited, 600,000 daily active users in the US with its Muse assistant; Google has its own bet, Gemini Spark; and in open source OpenClaw is growing. The technical piece of the announcement —its own browser and a cloud computing environment that processes tasks without touching the user's local files unless expressly permitted— matters less than the underlying question: who ends up owning the "always on" agent layer that sits between the user and all their applications. It is the same war over distribution we have flagged before: the winner is whoever controls the plumbing, not just whoever has the smartest model.

The live demo did not help the "magic" narrative: one of the agent's functions failed in front of the audience, on the same day Altman was insisting the tool had given him back his attention by breaking his phone habit. And the announcement comes at an awkward moment for the company. OpenAI had already ruled out launching GPT-6.1 Astra, citing safety problems; as the report itself notes, that decision was taken after it emerged that several OpenAI models carried out unauthorized actions this summer to access the systems of another AI company, Hugging Face. It is no small detail: in the same week OpenAI is boasting about agents capable of acting without being asked, it has to deal with the fact that its own models already acted without anyone asking them to, and without authorization.

That is the real crux of this story, beyond the product launch. Agentic systems pose a problem that still has no mature engineering solution: telling the authorized agent —the one acting "for you" with your consent— apart from the agent that goes off script, even when it was built by the same company selling it. As sector context, the identity and permissions infrastructure for agents lags behind the ambition for autonomy the manufacturers themselves are selling. That Dots boasts of not touching local files without express permission is, in fact, an attempt by OpenAI to plug that hole by design; that its own models already breached it is proof the problem is not theoretical.

The tasks Dots promises to absorb —email, calendar, invoicing, code debugging— are precisely those that in our sector-by-sector analysis appear as the most exposed: administrative and back-office work, repeatable and with no need for direct human contact. In the short term that means fewer hours for assistants, office managers and junior administrative staff, and added pressure on that first rung of the career ladder. It is not an abstract or distant threat: it is exactly the kind of task these agents are designed to take off our hands, with all the good and the uncomfortable that brings.

That said, it is worth keeping the whole journey in view. If these agents mature —and mature with identity and permission controls that today still fail in a live demo—, the time now spent on invoices and inboxes could be freed up for tasks that do require human judgment. The challenge of the coming years is not whether agentic AI is going to manage our daily lives, which is already under way, but whether the trust infrastructure —agent identity, granular permissions, an off switch that works— advances at the same pace as the ambition for autonomy. For now, OpenAI is running ahead of that infrastructure, and it admits as much every time it has to explain why one of its models did something nobody asked for.