OpenAI launches always-on Dots agents across connected apps
OpenAI launched Dots, what it calls ‘always-on agents’ that can pursue a user's work across connected applications, at its DevDay conference on 29 September. Powered by GPT-6 Astra, a personal Dot can draw on more than 4,000 app connections, work in ChatGPT Space and communicate through Slack or Microsoft Teams. The first personal Dot is available on eligible Pro and Business Premium plans, while Enterprise administrators can enable a beta. The agent can conduct background research and return with findings without a fresh prompt, moving OpenAI from answering individual requests toward handling continuing assignments. The product puts the company's large ChatGPT audience and its business integrations behind a more ambitious claim on daily work.
The launch begins with one personal Dot per user, while specialist agents for roles and teams are in pilot. OpenAI describes a Dot as learning a person's preferences and project context, then deciding when to use tools, ask a question or present work for review. It runs on a cloud computer rather than depending on an open browser tab or a laptop that stays awake. A user can give it a continuing objective, see its progress and set rules about which tasks require approval. This design matters because it changes the unit of engagement from a chat session to a relationship that persists between sessions, with more chances to consume inference and more opportunities to demonstrate utility.
OpenAI has placed permission boundaries at the center of the release. The initial proactive work is read-only research; sensitive actions require a user's approval, and administrators control enterprise access. Microsoft said Dots will integrate with Agent 365, while the launch description shows work continuing through established collaboration channels rather than requiring every interaction to happen in ChatGPT. Reuters reported that a live voice demonstration encountered a glitch, a reminder that the polish of an agent's handoff matters alongside the underlying model. The company's published safeguards describe the intended operating limits; customers will judge the product through the ordinary mistakes, interruptions and ambiguous requests that arise in sustained use.
The commercial reach is considerable even before specialist Dots arrive. Reuters reported OpenAI's figures of 1.2 billion weekly ChatGPT users and more than 35 million weekly users of Codex and ChatGPT Work. Those are audience measures across different products, not a Dot adoption count. OpenAI has included a first personal Dot in certain higher-priced plans, making the immediate aim deeper engagement and retention as well as a reason to move heavy users into premium tiers. A Dot that brings together existing files, conversations and connected services could become the interface through which users encounter other paid tools in the company's growing ecosystem.
The launch also lands in a fraught week for autonomous agents. OpenAI had described unexpected behavior in internal model work and paused some of its most capable tool-using research systems, while it separately shelved a planned GPT-6.1 Astra release. Dots use the already released GPT-6 Astra, a different model. The distinction is important to the product story: OpenAI is expanding consumer and business access to a persistent agent while tightening the controls around more capable development systems. Chief executive Sam Altman said at DevDay that AI should give people more agency and creativity. The practical test for Dots is whether persistent assistance saves enough attention to justify both compute consumption and the trust demanded by broad app access.
Analysis
Dots shift OpenAI's opportunity from selling a model response toward owning the sequence of work that generates many responses and tool calls. Including one agent in premium plans can raise usage costs, but a reliably useful agent also increases switching costs as it accumulates context and becomes the route into connected services. The 35 million weekly Work and Codex users are a meaningful distribution base, yet the larger 1.2 billion ChatGPT audience is not equivalent to users willing to delegate actions. The economic prize is an enduring workflow relationship; approval friction, execution quality and the cost of persistent compute determine whether that relationship yields attractive revenue per user.