The GenAI Field Guide · Trending

OpenAI dots

Always-on ChatGPT agents, each with its own cloud computer and browser, that keep working on a goal between conversations.

A dot is a persistent agent inside ChatGPT, running on GPT-6 Astra, that holds an ongoing goal, works on it in the background on a dedicated cloud computer, and comes back to you for approval when a rule says it must. It moves the product from answering a prompt to owning a responsibility. Anyone who runs agents, or who will be asked whether staff may use one against company apps, should understand its permission model before it lands on their desk.

What it is

OpenAI announced dots (lowercase is the official name) at DevDay on 29 September 2026. Reporting describes a dot as a personal agent that you name, give goals to, and then leave running: it keeps making progress between conversations, follows up on work, and can split work across sub-agents. OpenAI's own line, as quoted by the press, is that dots are always-on agents built to handle everything. Under the marketing, it is a long-running agent loop with a persistent sandbox, memory, and a connector layer, packaged as a consumer and workplace product rather than an API.

The interesting part is not any single capability. ChatGPT already had agents that could browse and use tools for a session. What changes is the time horizon and the initiative: a dot is expected to notice things in your connected apps and propose or do work without being asked each time. That is why most of the launch coverage, and OpenAI's own separate safety write-up, is about controls: what a dot may do alone, what needs approval, and what it can never do without you.

Coverage also places dots against Meta's Muse and Google's Gemini Spark as the same category of always-on assistant. One analyst quoted by CBS News read it as aimed at professional and enterprise users more than everyday consumers, which matches the plans it ships on.

How it works

Each dot gets its own cloud computer with a browser. Secondary reports describe it as an isolated Linux environment running Chrome, which you can watch or take over at any time. The dot reaches your work through plugins, with OpenAI citing more than 4,000 connectable apps. Access to your own laptop is a separate, optional grant that starts turned off and needs the ChatGPT desktop app and your confirmation.

When it has nothing assigned, a dot does what OpenAI calls proactive research: it looks through connected apps for ways to help, but through read-only access, so in that mode it cannot send messages, change content or drive a browser. The worked example in coverage is a dot that spotted an unpaid freelance invoice in email, pulled the details, drafted a PDF invoice, and waited for approval before sending it. That pattern, read freely and act only behind a gate, is the core design.

Action is governed by built-in rules plus Custom Rules you write. Reports describe four outcomes for an action: the dot does it alone, it is pre-approved, the dot asks first, or it hands off to you. An auto-review step checks planned actions that touch accounts or share information before they run, some actions such as password changes and money transfers always need you, and a monitoring system can pause or stop a dot when it detects a safety problem. OpenAI also describes Private Safety Processing, meant to let automated safety review run without giving OpenAI staff a new way to read protected customer content.

A dot learns your preferences from feedback and keeps memory across its work. According to OpenAI's privacy FAQ as quoted by secondary coverage, you cannot currently view, edit or delete individual dot memories; only deleting the dot clears them. Dots are a ChatGPT product, not developer infrastructure. The developer counterpart announced the same day is the Agents API, which offers durable sessions and a hosted browser for agents you build yourself.

How to use it

You need a ChatGPT Pro or Business Premium plan in an eligible market, or an Enterprise workspace whose admin has turned dots on (it starts off by default). Pro users in the European Economic Area, Switzerland and the UK are excluded at launch. Create the dot on ChatGPT desktop or web; after that you can also reach it from the mobile app.

  1. Open ChatGPT on desktop or web and follow OpenAI's Getting started with your dot help article to create your first dot, give it a name and describe the ongoing responsibility it owns.
  2. Connect only the apps the goal needs through plugins. Start narrow; every connection widens what proactive research can read.
  3. Write Custom Rules before giving it real work: list which actions it may take alone, which it must ask about, and which are blocked, and keep anything that sends, shares, pays or deletes behind approval.
  4. Leave laptop access off unless a task truly needs local files. If you enable it, it requires the desktop app and your explicit confirmation.
  5. Reach the dot where you work: ChatGPT, Slack or Microsoft Teams. Texting is described as coming soon or as a limited US Pro beta depending on the source.
  6. Watch its cloud computer for the first few tasks and correct it with feedback; that is how it learns your preferences.

Use cases

Sources

  1. Introducing dots, OpenAI, 2026-09-29
  2. Getting started with your dot, OpenAI Help Center, 2026-09-29
  3. OpenAI launches dots, always-on AI agents with their own cloud computers, The Next Web, 2026-09-29
  4. OpenAI launches Dots, its bubbly agentic avatar, TechCrunch, 2026-09-29
  5. OpenAI Dots are always-on agents that can do tasks for you, 9to5Google, 2026-09-29
  6. Sam Altman unveils dots, OpenAI's new AI personal agent, CBS News, 2026-09-29
  7. OpenAI launches dots, personal AI assistant built to handle everything, Al Jazeera, 2026-09-30
  8. OpenAI DevDay 2026: Dots and ChatGPT's Agent OS, The Neuron, 2026-09-30
  9. OpenAI Dots explained: always-on agents, pricing, and limits, eesel AI, 2026-09-30
  10. What Is OpenAI dots? Always-On Agents Explained, Spinnable, 2026-09-29
  11. ChatGPT Pricing October 2026, GEO Toolbox, 2026-09-29