OpenAI Debuts Dots, Always-On Agents That Work Without Being Asked

OpenAI has introduced Dots, a class of always-on agents that operate from dedicated cloud computers and can act without a prompt from the user. The agents are reachable through ChatGPT, Slack and Microsoft Teams, and they keep working when nobody is actively directing them. The launch puts OpenAI directly against Meta's Muse, which occupies the same emerging category of persistent, self-directed assistants.

The mechanics are what separate Dots from the chat assistants most people use today. Each agent gets its own cloud machine rather than borrowing a session inside a chat window, so it can hold state, run tasks and keep going after the user closes the tab. OpenAI says the agents can fix bugs or send out invoices that someone forgot, in some cases before anyone flags the task. Outside active sessions, Dots scan for useful work using read-only access, a restriction that limits what they can touch on their own.

That read-only default is the most consequential design choice in the release. An agent with write access to production systems and a mandate to act unprompted is a liability; one that can only look, and must surface findings for a human to approve, is a much smaller risk. It also shapes what Dots can realistically deliver in the near term, since the agent can spot a broken build or an overdue payment but cannot simply push the fix or move the money without a person in the loop.

The multi-platform approach matters as much as the agent architecture. By plugging into Slack and Microsoft Teams alongside ChatGPT, OpenAI is putting Dots where work already happens instead of asking teams to adopt another destination app. That is the same distribution logic Microsoft has used with Copilot, and it sets up an awkward overlap: Teams is a Microsoft product, and Microsoft is both OpenAI's largest backer and a competitor in workplace agents.

Meta's Muse gives the category a second well-funded player and signals that persistent agents are becoming a product line rather than a demo. For enterprise buyers, the practical questions are access control, audit trails and what happens when an agent misreads a situation and acts anyway. Neither OpenAI nor Meta has published the kind of detail on those controls that security teams will want before handing an agent the keys to internal systems.

What to watch next is pricing and permissioning. OpenAI has not said what Dots cost, how many a team can run, or whether administrators can scope read-only access down to specific repositories and inboxes. The answers will decide whether always-on agents become standard workplace infrastructure or stay a feature people try once and quietly switch off.