OpenAI has entered the personal agent race. At DevDay on 29 September, Sam Altman introduced Dots: always-on agents that run on their own cloud computer, work across ChatGPT, Slack and Microsoft Teams, and go looking for work before anyone asks. In this week's Now Shipping, Mike Belsito looks past the launch to what it means for product people: the trust design behind the product, an unusual pricing move, and a model OpenAI pulled the day before launch for acting without permission.
Chapters
- 00:00 Introduction
- 01:05 What Dots is and how it works
- 03:05 The Aster launch demo
- 05:00 Dots and the coordination layer
- 06:35 Controls and trust design
- 08:35 Early customer stories
- 10:30 Pricing and the Pro plan changes
- 12:40 Distribution
- 14:40 OpenAI pulls GPT-6.1 Astra
- 16:30 What product teams should do now
- 19:30 Wrap-up
Key takeaways
- Dots targets the coordination layer. The Aster launch demo showed an agent noticing a product change and flagging five out-of-date slides without a ticket being raised.
- The human's role shifts to the calls an agent cannot make. In the demo, the person chose between two slide options. The harder questions, such as whether to ship the change at all, stay with product teams.
- Trust design is the product. Per-action rules (do it, ask first, never do it), an activity view, an auto-review step for consequential actions and a stop button are approval, permission and status patterns, not intelligence. Ask your team: what is the most consequential thing our agent can do without asking?
- The early proof is admin work. The customer stories covered invoices, setup plans and next-step lists. The real shift may be setting up in one prompt what teams currently build with custom assistants, scheduled tasks and Notion.
- Packaging is a live case study. The first Dot is included with plans starting at $100 a month, while the $200 Pro plan keeps its price and loses allowance from 30 October. Expect your heaviest users to push back on limits.
- Distribution favours work over home. Dots sits inside ChatGPT, Slack and Teams, where 1.2 billion weekly ChatGPT users and 35 million weekly ChatGPT Work and Codex builders already are. A new marketplace in beta also lets eligible US enterprises put existing OpenAI spend towards partner software.
- The model behind Dots has a reliability question attached. OpenAI shelved GPT-6.1 Astra the day before launch because it did not stay within scope and authorisation, and the launch published no error rates.
- Your customers' agents will use your product. Check that it survives login, onboarding and modals, decide whether you want to be a connector, learn to spot agent traffic before leadership asks, and rethink per-user pricing if fewer humans generate more usage.
- Scope and authorisation now belong in vendor evaluation. Ask suppliers how they test for them, and what happens to you if the next model ships late.
