ProBackend
AI Agents
5 days ago6 min read

OpenAI wants its Dots to keep working after you close the chat

OpenAI wants its Dots to keep working after you close the chat

OpenAI’s Dots are designed to carry out ongoing tasks across connected apps. Here’s what the company says they can do, and where user control still matters.

OpenAI’s Dots make a bolder promise than a chatbot that answers quickly: they’re meant to keep working when the conversation pauses. That shift could make an assistant genuinely useful for projects that span days. It also raises the harder question: what should an always-available agent be allowed to do without asking first?

OpenAI announced Dots on September 29, 2026, presenting them as proactive assistants for complex projects and everyday tasks. The concept is not just “ask a question, get a response.” A Dot can work in the background, use connected apps, and come back with progress, a question, or a decision for the user. That’s a meaningful change in the relationship between person and software: the assistant is expected to notice and follow through, not merely wait for the next prompt. OpenAI’s announcement

From chat replies to ongoing work

The appeal is easy to understand. Most personal and office work is made of small, scattered obligations: check a calendar, compare options, gather details, then decide what to do. A conversational assistant can help with each step, but the user usually has to keep the whole sequence moving. Dots are pitched as a way to hand over more of that coordination.

OpenAI says each Dot can use its own cloud computer, browser, and connected apps, while allowing users to inspect its work. It also describes the assistants as carrying context across channels. That makes the product sound less like a standalone chat window and more like a layer sitting across the tools where work already happens. Those are company-described capabilities, not proof that every task will be seamless or that every integration behaves the same way. OpenAI

A launch demonstration reported by WIRED gives the pitch a concrete shape: a Dot noticed a calendar conflict with dinner and offered two Grubhub options with prices. It’s a modest example, which is exactly why it lands. The value isn’t a dazzling piece of reasoning; it’s catching a small collision in the day and doing some of the comparison work before the user asks. Whether people want an assistant to make that kind of inference routinely is another matter. WIRED

Proactive help needs boundaries

“Always on” sounds convenient until the assistant’s idea of helpfulness diverges from yours. A system that can read connected information, notice patterns, and take steps needs more than a good answer. It needs a clear distinction between gathering information and changing something on a user’s behalf.

OpenAI says Dots have built-in rules for independent action, and that users can add custom restrictions or permission requirements. The company also describes an auto-review process that checks whether actions follow those rules and safety requirements, and can return tasks that need user involvement. These are useful design commitments. They’re not a guarantee that mistakes disappear, and the distinction matters: a review mechanism is a control to evaluate, not independent evidence that the product is safe in every situation. The Verge

OpenAI’s product material adds that the Dot’s cloud computer is separate from the user’s own computer unless the user chooses to connect it. For supported websites, saved passwords can be used without exposing them to the model. The company says background “proactive research” uses connected apps with read-only tools, which cannot send messages, change app content, or control the user’s browser or computer. It also says safeguards are intended to defend against malicious instructions and monitor for potentially harmful behavior. That separation between reading and acting is a sensible starting point; users will still want to know exactly when a task crosses from one to the other. OpenAI

The practical test won’t be whether the assistant can describe its rules. It’ll be whether those rules are understandable before a task begins, visible while work is underway, and reliable when a task goes wrong. OpenAI says users can set boundaries, follow progress, and remain involved in decisions that need them. Those controls are central to the product, not decorative settings. An assistant that saves a few minutes but leaves users unsure what it has read or changed won’t feel like a trustworthy coworker.

A product launch, not a settled category

The launch also lands in a crowded race to make assistants act rather than just talk. The New York Times framed Dots as a competitor to Meta’s Muse and other AI assistants. The strategic bet is clear: the next contest may be over who can maintain useful context and complete multi-step work, not who produces the most fluent single answer. The New York Times

OpenAI said the initial rollout would begin for Pro, Business Premium, and Enterprise plans in eligible markets. That wording is narrower than a general release: availability depends on plan and market, and should not be read as access for every ChatGPT user. The company has also previewed specialist Dots for organizational roles, alongside access management and connections to company systems. The Verge reported that these specialist versions were being tested for company roles. OpenAI The Verge

For businesses, the promise is appealing and the requirements are stricter. A general assistant can be a personal convenience; one connected to company systems has to fit permissions, oversight, and existing processes. Specialist agents could help with repetitive work, but an organization will need to decide which information they can access and which actions require approval. OpenAI’s mention of access management points toward that problem, without by itself answering how every company should handle it.

The Times also reported that Dots are powered by GPT-6 Astra and that OpenAI delayed GPT-6.1 Astra amid safety concerns. It placed the launch against broader scrutiny of agent safety, including reported incidents involving OpenAI agents and external websites. That context doesn’t establish that Dots will repeat any particular incident. It does explain why claims about safeguards deserve careful attention, especially when software is being asked to operate across services and act with less continuous prompting. The New York Times

The real question is what gets delegated

Dots are an ambitious answer to a familiar frustration: assistants often understand a task but leave the user to do the follow-through. OpenAI is trying to close that gap with persistent context, connected tools, and background work. If it works well, the improvement may feel less like a new chatbot feature and more like having one fewer loose end to manage.

But persistence changes the bargain. Users need to understand what the Dot can see, what it may do on its own, and where it must stop and ask. OpenAI’s descriptions of read-only background research, custom restrictions, review, and user oversight are promising pieces of that bargain. They remain claims about a product whose value will depend on how those protections perform in ordinary use.

For now, the most useful way to judge Dots is not by the “always-on” label. It’s by the handoff: can a person delegate a bounded task, see what happened, and retain control of consequential choices? Until that feels dependable, the assistant’s best work may be helping with the small things—and asking before it tries to do the big ones.

from chat replies to ongoing workfrom chat replies to ongoing work