OpenAI’s New Dots Agent Keeps Working Between Conversations

OpenAI introduced dots at its September 29, 2026, DevDay event, pitching them as AI agents that can keep working after a conversation ends. A dot can follow a goal over time, use connected apps, work on its own cloud computer and return with progress or questions. The launch moves beyond the familiar pattern of opening a chatbot, asking for an answer and closing the window.

That does not mean a dot has unrestricted control over a user’s accounts or devices. What it can see and do depends on the connections a user enables, the task they give it and the rules that apply before it takes action.

What can a dot do?

Each dot is powered by GPT-6 Astra and has a cloud computer and browser for carrying out tasks. Users can give it a project, add instructions as the work develops and inspect its computer to see what it is doing. OpenAI says dots can connect through its plugin ecosystem to more than 4,000 apps, though an individual dot can use only the accounts and capabilities made available to it.

OpenAI’s examples include watching customer feedback, preparing and testing software fixes, updating a proposal as requirements change, and revising an analysis when new data arrives. These are examples of intended workflows, not a promise that every dot will complete every task correctly or without a person checking the result.

A dot can also run scheduled checks and send progress updates. Its more proactive background research can look through permitted connected information for useful developments even without a fresh request. OpenAI says that background research is restricted to read-only tools: it cannot, by itself, send messages, change material in connected apps or operate a browser or computer. Any action that follows must pass the normal permission and approval checks.

Users can talk to a dot in ChatGPT and, where enabled, through workplace messaging services such as Slack and Microsoft Teams. A dot may ask for a decision when it reaches a step that needs human judgment. Connecting it to a personal computer is optional; its cloud computer does not automatically give it access to local files.

Who can use it?

Dots began a gradual rollout on September 29 for ChatGPT Pro subscribers in eligible markets and Business Premium users in supported ChatGPT regions. For Pro users, the initial rollout excludes the European Economic Area, Switzerland and the UK. Enterprise, Edu and Healthcare workspaces can try a beta if an administrator enables it; Enterprise access is off by default. OpenAI has not announced general access for Free or Plus users.

An eligible user can start with one dot, included in the Pro or Business Premium plan at no additional charge. Initial setup takes place in the ChatGPT desktop app or a desktop browser. Once it is set up, the user can message it in the mobile app when that access is available; creating a dot on mobile is not supported at launch. Because the rollout is gradual, an eligible subscriber may not see it immediately.

OpenAI says conversations with a dot do not count toward ordinary ChatGPT usage limits. Deeper work has plan allowances, and tasks a dot starts in Codex or ChatGPT Work count toward those products’ usual limits. The company says it will share further usage terms after an initial launch period, so subscribers should not assume every kind of ongoing work is unlimited.

How much independence does it get?

The distinction from a regular ChatGPT conversation is continuity: a dot can retain context, pursue an ongoing responsibility and make progress between exchanges instead of waiting for every next step to be typed into a chat. It is still an assistant operating within permissions, not an independent account holder with blanket authority to act.

Users choose connected apps and can set rules for actions that should proceed, require approval or be handed back to them. OpenAI says a separate review system checks proposed actions that could affect accounts or share information. Purchases require approval, while steps such as changing a password remain with the user. People can inspect activity, redirect work or pause their dot.

Those controls matter because persistence raises the stakes of a misunderstanding. A mistaken answer in a chat can be corrected before anyone uses it; an agent working across files and accounts may encounter more opportunities to make a consequential error. OpenAI acknowledges that dots can make mistakes and advises users to review important work. For now, the practical question is not whether a dot can work around the clock, but which work a user is comfortable letting it pursue—and which decisions they want to keep for themselves.