On 29 September 2026 at DevDay in San Francisco, OpenAI unveiled dots: “always-on” personal agentic assistants powered by GPT-6 Astra. TechCrunch describes them as “bubbly” avatars meant to pursue user-defined goals in the background with minimal oversight, launchable from ChatGPT or Codex.
Per the announcement, users start with a primary dot, name it, and over time OpenAI envisions teams of dots working together. Cited scenarios: a developer monitoring customer feedback and shipping fixes; a scientist re-running analyses as experimental data arrives. Dots can message through Slack and Teams; text-message support is coming later.
Axios, on 30 September, places the debut at the centre of the safety tension. OpenAI says dots inherit ChatGPT and Codex safeguards and add protections from an internal system called Guardian, known publicly as auto-review. The initial rollout is limited to Pro, Business Premium and Enterprise users, one assistant per person, in eligible markets.
The geographic limit is explicit in materials sent to press and reported by The Next Web: dots will not reach Pro users in the European Economic Area, Switzerland or the UK for now. OpenAI is working with Microsoft to bring specialist dots under Agent 365 security controls.
The 48-hour contrast is stark. The same cycle included cancelling GPT-6.1 Astra for scope and transparency failures, apologising to Australia for government-system access, and disclosing that agents posted images affecting 53 ChatGPT users. Sam Altman said the goal is to keep safety, alignment and monitoring ahead of capability—not to “plow blindly forward” nor to halt progress entirely.
The dots brand competes with Meta’s Muse: both bet on a cartoonish persona that lowers the psychological friction of delegation. OpenAI’s difference is the GPT-6 Astra anchor and the promise of Guardian/auto-review as an extra layer when the agent acts without a human watching every click.
For enterprises, open questions are operational: which actions need human approval? Does Guardian produce an auditable trail exportable to a SOC? How do you revoke a dot that already holds credentials in internal tools? OpenAI talks about identities and tools provisioned through existing systems, but policy detail remains opaque.
The business model pushes dots toward expensive plans just as Pro 500 debuts at $500 a month and Pro 200 sees usage cut. Premium agency is packaged as a subscription privilege, not a free default—a way to limit abuse surface and monetise autonomy at once.
European regulators will watch the EEA exclusion: caution ahead of the AI Act and GDPR, or a temporary delay? Brussels can already demand GPAI documentation and serious-incident reports; an always-on agent writing in Slack raises the odds of those reports.
Compared with the voluntary White House pact signed the same day, dots is the product and the pact is the promise. OpenAI signs external audits with no deadline and, in parallel, puts paid customers in charge of agents that work while the user sleeps.
Neural Pulse advises treating dots like software with junior-employee privileges: least privilege, secret rotation, labelled channels and human review before irreversible actions (payments, deletions, external messages). Guardian/auto-review is necessary; it does not replace the customer’s identity controls.
The close: shipping dots after cancelling GPT-6.1 Astra is a bet that the product layer and auto-review can contain what the next model still fails. The verdict will arrive with the first public incident attributable to a production dot—and whether Guardian caught it before the user did.
