OpenAI has launched Dots, always-on AI agents that run on their own cloud computers to autonomously fix bugs or send forgotten invoices, sometimes proactively. Users can access them through ChatGPT, Slack, and Microsoft Teams, and they use read-only access to find ways to help in the background. The move positions OpenAI against Meta's Muse.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
At DevDay 2026, OpenAI introduced Dots, always-on agents that each have their own cloud computer and can work autonomously around the clock. It ties "proactive research" to read-only permissions and approval rules, while shutting out European Pro users and the top-tier model GPT-6.1 Astra—capability expansion and risk control are placed on the same launch list.
At the DevDay 2026 developer conference, OpenAI introduced always-on AI agents called "Dots." Each Dot has its own cloud computer equipped with a browser, so it can work around the clock, conducting research, analyzing data, writing documents, or building software via Codex. This differs from the earlier Computer Use form in ChatGPT Work: the latter has users launch tasks one by one, with the agent operating the user's own screen, whereas a Dot by default works on its own computer, hands tasks to Codex when needed, and can only access the user's laptop when granted permission.
Users can reach their Dot through ChatGPT, Slack, and Microsoft Teams, and the Dot maintains context across these channels. OpenAI says the Dot learns from feedback how the user works and what the user considers a good result. Each Dot has a name and a cartoon avatar; in the demo, the desktop app shows chat on the left and what the agent is working on on the right, whether a Google Slides presentation or a cloud computer with a browser, terminal, and file storage. If the Dot is in control, the user can click "Take over" to step in.
When no one is actively using it, the Dot looks for ways to help in the background using read-only tools, which OpenAI calls "proactive research." This mode cannot send messages, change content, or control the browser or computer. This is a clear boundary between Dots and "autonomous execution": in the background it can look and find, but it cannot act and change.
For proactive tasks, a basic set of rules determines when the Dot acts on its own and when it confirms first. Custom Rules allow users to allow, require approval for, or prohibit individual actions, and operations involving accounts or sharing information go through an automatic check first. Sensitive tasks such as changing passwords are always done by humans. The Dot can use saved passwords without the model itself seeing them, safeguards are designed to block malicious instructions, and a monitoring system can pause or stop work if a safety issue arises.
OpenAI also acknowledges that Dots can make mistakes and recommends reviewing any result with real consequences. The source does not quantify the specific scope and frequency of error risk, so "making mistakes" is currently an officially acknowledged but unquantified risk, not an assessable probability.
Restricted release in Europe: currently Pro subscribers can only use Dots outside the European Economic Area, Switzerland, and the United Kingdom. This is a clear geographic constraint, meaning the same product creates an availability gap between European Pro users and Pro users in other regions.
Business Premium customers can use Dots in all supported regions, and Enterprise, Edu, and Healthcare customers can try the beta after an administrator enables it. Users first set up their Dot in the desktop app or browser, and can later use it on mobile as well.
A subscription comes with one Dot, and chatting with it does not count against usage limits; tasks it launches in Codex or ChatGPT Work are billed as usual, and the first-month limit is expanded. More Dots, higher speed, or greater workload will be charged extra later. OpenAI's long-term plan is for teams of Dots to work together, but at launch each user has only one.
Dots runs on GPT-6 Astra. At the same time, OpenAI released the cheaper model GPT-6.1 Sol, while the top-tier GPT-6.1 Astra was paused due to safety issues. The source does not explain the specific safety issues behind the pause, so the reason for the pause can currently only remain at the phrase "safety issues" and cannot be inferred further.
Putting these two things together: on one side is a product that lets agents work autonomously on a cloud computer around the clock, and on the other is a top-tier model not being released for safety reasons. This at least shows that in the same launch, OpenAI advanced both the capability boundary and the safety gate at the same time, rather than handling the two separately.
Dots is technically similar to Meta's Muse agent, launched three weeks earlier. According to Meta, Muse also runs on its own secure computer equipped with a browser, keeps working in the background, and requests approval before sensitive operations. This similarity judgment comes from reporting and Meta's own account, lacks independent verification, and can therefore only be treated as a source claim.
Plugins let one Dot connect to more than 4,000 apps and advance multiple projects at the same time. This figure comes from OpenAI and has not been verified by a third party. OpenAI also says that in one internal case, a Dot tracked a bug reported in Slack until it produced a usable new build; an early tester's agent found a forgotten invoice, prepared it, and sent it after the tester approved. These internal cases are all OpenAI's own accounts, with no independent evidence.
Everyday features such as ordering food via text message are only described as "coming soon," with no set timing.
This article is based on THE DECODER's reporting on OpenAI DevDay 2026. The launch of Dots, the cloud computer configuration, channel access, the read-only proactive research mode, approval rules, password safeguards, European release restrictions, enterprise customer arrangements, subscription and billing methods, the release of GPT-6.1 Sol, and the pause of GPT-6.1 Astra all come from OpenAI information cited in that report and are relayed at the level of confirmed facts. The claim that plugins connect to more than 4,000 apps, the technical similarity to Meta Muse, and the internal cases (tracking a Slack bug, finding a forgotten invoice) all come from OpenAI or Meta's own accounts and lack independent verification, so this article treats them as source claims. The specific safety issues behind the pause of GPT-6.1 Astra, the scope and frequency of Dots' error risk, and the launch timing of the text-message food-ordering feature are all not explained by the source.
Dots pushes the "always-on agent" from concept to a subscribable product: each Dot gets a cloud computer, maintains context across channels, conducts read-only proactive research in the background, and routes sensitive operations through approval. But the same launch also draws two boundaries—European Pro users cannot use it for now, and the top-tier GPT-6.1 Astra was paused due to safety issues. Capability expansion and risk control are placed on the table at the same time, while key claims such as 4,000 app connections, the similarity to Muse, and the internal cases remain at the level of OpenAI or Meta's own accounts.
Primary report
Primary source