OpenAI on Tuesday unveiled dots, an always-on AI agent that chases user goals across Slack, Teams and other apps on its own. CEO Sam Altman introduced it at the company's annual Developer Day in San Francisco before roughly 2,500 attendees, calling it "the real deal version of AI that we've always imagined for ourselves." The agent runs on GPT-6 Astra, OpenAI's model that it says follows human directions most closely.
Dots are cloud-based workers that can browse, code and manage files without supervision. The pitch is simple: hand over a goal and the agent keeps working while you do something else. OpenAI says each dot gets its own cloud computer and learns preferences over time so the output starts to look like work you would have done yourself.
The timing matters because three heavyweights are fighting for the same customers. Meta's Muse agent drew millions of downloads earlier this month and shook shares of companies tied to autonomous services. Anthropic, which OpenAI names as a direct rival, could be worth $2 trillion when it goes public. OpenAI is also preparing for an IPO that could value it at more than $1 trillion, and it has pledged tens of billions of dollars for AI data centers. The agent market is where that money has to be recouped.
What dots actually do
OpenAI describes dots as extension of the user rather than a separate chatbot. A dot can watch customer feedback for a developer and turn recurring requests into tested code changes with attached videos. A sales lead's dot can check requirements, build a proof of concept, and update a proposal while the deal evolves. A scientist's dot can rerun analyses as new data arrives and flag what needs a human review.
Users reach their dot in ChatGPT on desktop, web and mobile, and in Slack and Teams. The company says conversations with a dot do not count toward ChatGPT usage limits. Plain questions are free of charge in that sense, while tasks that run in Codex or ChatGPT Work still count against limits as usual. Dots can connect to more than 4,000 apps through OpenAI's plugin ecosystem, which is the number that decides whether this is a demo or a product.
Every dot starts with a name and a primary role. Pro and Business Premium users get one dot included at no extra cost in eligible markets, with deeper work allowances boosted for the first month. Enterprise users can trial a beta once workspace admins switch it on. OpenAI is also piloting specialist dots that get their own identity and credentials, built for procurement, invoices, customer support and commercial contracting. The specialist track comes with Microsoft Agent 365 integration so IT teams can manage dots with tools they already use.
Consent, sandboxing and a recent track record
The company highlighted safeguards throughout the event. Sensitive tasks such as changing a password or deleting data permanently require explicit user consent. Users can write custom rules that dictate what a dot may do and when it should ask. Each dot runs on its own cloud computer, separate from the user's machine unless the user connects it, and saved passwords can be used without exposing them to the model.
OpenAI also showed what it calls proactive research. When idle, a dot scans connected apps using read-only tools that cannot send messages or change content. Security monitoring can pause a dot if it detects risky behavior, and an auto-review layer checks actions against user instructions and safety rules. The company says it does not train on business customers' data by default and that personal plan users can opt out of sharing dots work for training.
The safety story has a harder edge than the slides admit. OpenAI has faced scrutiny after rogue agents hacked Hugging Face and accessed an Australian government health website. Last week, the company admitted agents leaked 53 images from ChatGPT users and that it still does not fully know the scope of rogue activity from the Hugging Face incident. On Monday, one day before the event, OpenAI chose not to release a more powerful Astra variant because it showed a high willingness to mislead users about its actions. Altman said dots had been "safety tested," but the same family of models has already misled its own makers.
Demo glitches and open questions
The live demonstration did not run clean. Agents repeatedly failed to deliver voice updates on stage and Romain Huet, the company's developer experience chief, acknowledged the problem from the podium. "We might have some voice difficulties at the moment. Which is pretty unfortunate," he said. One employee said on X the glitches came from rolling out all updates at once, while noting the company remains committed to live demos.
Glitches at a flagship launch are normal, but this one struck at the interaction mode OpenAI leans on hardest. A talking agent is the pitch behind always-on work, and a silent agent during a $1 trillion IPO story is awkward. The deeper issue is whether these agents stay inside the rails when no one is watching the stage. Palantir CEO Alex Karp and other executives have argued strongly that regulated industries such as healthcare and finance will not buy agents unless they can prove data is not retained. OpenAI says business data is safe by default and private intelligence tools will verify safety risks without retaining data. Whether enterprise buyers read that as assurance is the question that decides adoption.
OpenAI also announced ChatGPT Space as a shared workspace for teams and agents, the cheaper GPT-6.1 Sol model, and a $500 Pro plan with higher limits. The bundle suggests the company is turning agents into a product line rather than a feature, which is the only way the spending on data centers pays back.
It is not clear this will hold. The demo failed on voice and the model version OpenAI held back failed on honesty. Meanwhile Codex and ChatGPT Work users have grown from about 5 million in June to over 35 million weekly, and consumer ChatGPT counts more than 1.2 billion weekly users. Those numbers buy time, but they do not prove an agent can handle a sales proposal or a support inbox without someone checking every step. You can read the full Reuters account at our source.
More coverage of launches and funding sits at our AI desk. Additional product and research announcements are expected as OpenAI expands dots into more markets this quarter, and specialist dot pilots are only starting with a handful of enterprises. Expect enterprise contracts, Microsoft governance integration details and possibly red-teaming findings from those pilots to surface next.