dots

dots are OpenAI's agents, launched on 29 September 2026. Each one runs on GPT-6 Astra with a cloud computer and a browser of its own, connects to more than 4,000 apps through ChatGPT's plugins, and keeps working while you are not talking to it.

It is in this directory because of where it reaches you. You message or call it inside ChatGPT, and you can also message it in Slack and Microsoft Teams, carrying the same context across all of them. OpenAI's own page says texting is coming next and does not say over what. Until it does, the channel row here records what is stated, and iMessage, SMS and RCS are not on it.

It also messages you first, with progress, questions and decisions that need you. Between conversations it does what OpenAI calls proactive research, reading the apps you have connected through tools restricted to read-only. That work cannot send messages, change anything in those apps or drive a browser.

The work OpenAI describes is long-running. A developer's dot watches customer feedback and brings back tested fixes. A scientist's reruns an analysis whenever new data arrives. Its one example from outside the company is a tester's dot that noticed an unsent invoice, prepared it and sent it after he approved.

On control, OpenAI published a separate safety write-up on launch day. Your computer stays separate unless you choose to connect it, you can open the dot's own computer at any time to see what it did, and Activity View shows background work as it happens. Saved passwords come from an encrypted credential service and sign in without the model seeing them.

It names prompt injection, a webpage or email carrying instructions meant to redirect the agent, and says a monitoring system can pause a dot mid-task if it flags a concern.

The write-up is also specific about what a dot does without you. It can pay with a card already saved on a merchant's site, and every such purchase waits for your approval. Health data goes only to a recipient you name. Changing a password or moving money between accounts is handed back to you, and Custom Rules can tighten any of this but cannot remove those steps.

On training, the answer splits by plan. Business, Enterprise and Edu workspace content is not used to improve models by default, and personal plans follow the "Improve the model for everyone" setting. Background research is not trained on directly, though a note from it can be once it is pulled into a conversation. Disconnecting an app stops new information flowing in, and what a dot has already learned stays in its context until you reset it.

Specialist dots, set up by a company with their own identity and credentials for one job, are in enterprise pilots, and OpenAI says it is working with Microsoft to manage them through Agent 365.