OpenAI introduced dots on 2026-09-29, the headline product of DevDay 2026, describing them as “remarkably capable, always-on agents built to handle everything.” Each dot is powered by GPT-6 Astra, has “their own cloud computer,” its own browser and the apps a user connects (OpenAI says over 4,000 through its plugin ecosystem), learns from feedback and “can work towards your goals 24/7.” Users reach their dot in ChatGPT on desktop, web and mobile, by voice call, or in Slack and Teams, with texting “coming soon.” Rollout started across Pro, Business Premium and Enterprise plans in eligible markets. OpenAI also previewed “specialist dots” with their own identity for access management, IT-provisioned hardware and deep integrations with company systems of record, including through Microsoft Agent 365.
The design keeps the agent off the user’s machine by default: “Each dot works on its own cloud computer, while your computer and its contents stay separate unless you choose to connect it.” When not actively in use, a dot does what OpenAI calls “proactive research” through connected apps using tools restricted to read-only. Actions that could affect accounts or share information pass through an auto-review step that checks them against the user’s instructions, Custom Rules and safety requirements; some tasks, such as changing a password, always stay with the user, and a monitoring system can pause or stop a dot. Saved passwords can be used to sign in without being exposed to the model.
OpenAI gave in-house examples (dots investigating a bug as soon as it is reported in Slack, turning a new design into a working app) and one from outside: an early tester’s dot noticed he had forgotten to invoice a publication, prepared the invoice and sent it after his approval.
Why it matters: dots shift ChatGPT from a conversation the user starts to a standing agent that keeps working between conversations, with its own machine and a growing model of the user’s preferences. It is a mass-market bet on the persistent-agent pattern that coding tools had been testing. What it does not show: the announcement is a product page with no published measurement of reliability or error rates for unattended work, Pro users in the EEA, Switzerland and the UK are excluded at launch, and how well the read-only and auto-review safeguards hold up in real use is not yet documented.