OpenAI Launches Dots, Always-On AI Agents That Work Independently on User Tasks

OpenAI has unveiled Dots, a new category of autonomous AI assistants that operate continuously on behalf of users rather than waiting for individual instructions. Announced at DevDay 2026 in San Francisco, these agents represent the company's entry into the fast-growing personal AI market and a direct response to competitors like Meta's Muse.

How Dots Work#

Dots are powered by OpenAI's GPT-6 Astra model and operate through a fundamentally different paradigm than traditional chatbots. Rather than responding passively to queries, they actively monitor the web and connected apps, pulling in contextual information about user schedules, preferences, and past interactions to anticipate needs. The agents can be personalized and assigned complex, multi-step projects—and they handle them autonomously without constant prompting.

In demonstrations, OpenAI showed a Dot recognizing that a user's calendar indicated dinner conflicts and proactively messaging two GrubHub delivery options with pricing. Another example depicted an agent collaborating with a user to build and launch a website. These demonstrations underscore how Dots blur the line between assistant and collaborator.

Access and Integration#

Beginning today, Dots are available to subscribers of ChatGPT Pro ($100 per month) starting with a single agent per user. The platform allows interaction through ChatGPT, Slack, and Microsoft Teams, with context flowing seamlessly across these channels. Pro subscribers can also join a waitlist to communicate with Dots via iMessage or RCS messaging on Android.

OpenAI typically rolls new features first to its premium tier before expanding to broader user bases, suggesting Dots will eventually reach the wider user population.

Safety and User Control#

Recognizing the risks of autonomous agents, OpenAI has built in guardrails. Dots will seek explicit permission before executing sensitive operations such as installing software or changing passwords. Users can also define Custom Rules to establish boundaries around what tasks agents can perform independently versus those requiring human approval.