OpenAI unveils always-on agent Dots with GPT-6 Astra

Dots is not simply a chatbot with a longer memory. Each dot gets a dedicated cloud computer, a browser, and its own ID for managing permissions, and it runs separately from the user's PC, so it cannot reach local data unless the user explicitly connects it. Even when logging into web services, the setup is designed so stored passwords are not exposed to the AI model itself.

The workspace half of the announcement, ChatGPT Space, is deliberately separate from that cloud computer. Space is where people, team members and dots look at the same pages, files, research, graphs and prototypes, and where tasks can be assigned to a specific dot from a comment. OpenAI also said it plans to let humans and dots co-edit presentations inside Space.

DevDay showed concrete work rather than abstract promises: a dot spotting a bug report in Slack, investigating the cause, writing and testing a fix, then opening a pull request, with OpenAI saying its engineers' dots fix dozens of bugs a day this way. The more cautious details are the controls around it, since active operations can be set to auto-approve, require approval, or be blocked outright, and sensitive steps like password changes still fall to the user. How far companies let dots reach may hinge on the planned integration of expert dots with Microsoft Agent 365 governance and security tools.