OpenAI introduced a new class of AI assistants called Dots during its DevDay 2026 conference in San Francisco. The service presents users with customizable, blob-shaped agents that remain active continuously, handling tasks on the user’s behalf. Unlike traditional single-turn chatbots, Dots are designed to operate in the background, monitoring information sources and acting when prompted.
The agents run on OpenAI’s GPT-6 Astra model, which powers their persistent capability. By staying online, a Dot can continuously crawl the web, retrieve data, and incorporate context from linked applications, gradually learning a user’s preferences. This “always-on” architecture differentiates the product from episodic conversational tools that require a new prompt for each interaction.
OpenAI demonstrated the system with two scenarios. In the first, a Dot examined a user’s calendar, recognized a free evening, and suggested two meal-delivery options from GrubHub, complete with price details. The user selected a choice and later instructed the agent to place the order. A second demo showed collaborative website creation, where the Dot assisted the user throughout the launch process.
Users may converse with their Dots through several platforms, including the ChatGPT interface, Slack, and Microsoft Teams, with shared context across these channels. OpenAI also opened a waitlist for Pro subscribers to interact via iMessage on Apple devices or RCS messaging on Android, expanding the reach of the agents into mobile messaging environments.
The launch aligns with OpenAI’s push toward enterprise adoption. At the same event, CEO Sam Altman announced “specialist Dots” tailored for corporate use, capable of handling functions such as accounting, email marketing, and legal analysis. These domain-specific agents are intended to embed within professional workflows, offering businesses a customizable AI coworker that can be constrained by explicit user permissions.
OpenAI emphasizes built-in safety mechanisms, yet the continuous nature of Dots raises privacy and security considerations. Agents request explicit consent before performing sensitive actions like software installation or password changes, and users can configure a Custom Rules feature to forbid certain operations. Nevertheless, connecting personal data and enabling model training on that data could expose information if the agents are misled or compromised, so caution is advised.