On Tuesday, OpenAI launched Dots: always-on agents that keep working after you close the laptop, each with its own computer in the cloud.
I took out a month of ChatGPT Pro to test them. Knowing how these things behave is my job.
I tested mine from Valencia. Officially, I was somewhere else.
If you pay for Pro in the European Economic Area, Switzerland or the UK, OpenAI's launch notes say Dots are not for you yet. I paid, and I still needed a VPN to use what I'd paid for.
A late arrival at a premium price
First, the product itself. It works. It sometimes sits there thinking for longer than you'd like. The interface is interesting, and the avatars (you name your dot, it gets a face) are a very mainstream move with real potential.
The keynote landed softer than the build-up, though. Voice mode refused to start during a live demo and the presenter finished by typing. That happens to everyone.
The calendar is the bigger problem. Meta's Muse launched three weeks earlier with no subscription price at all. Google's Spark sits on a 20 dollar plan. Grok Bot starts at 20 dollars too, and already lets you run several bots. Dots start at 100 dollars a month, one dot each.
That's OpenAI's call to make. I just think the avatar is aimed at everyone and the price is aimed at almost no one.
The map nobody explains
Dots isn't the first agent to skip us. Google's Spark reached more than 160 countries in July and left out the same list: the European Economic Area, Switzerland, the UK. Meta's Muse went from the US to Canada in ten days and hasn't crossed the Atlantic.
Neither OpenAI nor Google has said why. So this is my reading, not their statement: when two companies publish the same list of exclusions, the common factor is the rulebook. Switzerland and the UK aren't even in the EU, so one law can't explain it. What the three share is a habit of writing the rules before the product exists.
One detail makes it stranger. Business Premium customers get Dots in every region, Europe included. The agent is acceptable for your employer and not for you.
A boundary is designed, a border is drawn
In August I wrote that agents don't need interfaces, they need boundaries: permissions, scope, approvals, limits. I still think that.
A boundary is a design decision. Someone who knows the system decides what it can touch, tests it, and changes it when it fails. Dots ship with some of those: rules on when to act alone and when to ask, enforced from outside the agent's reach.
A border works differently. It's drawn by people who don't build the product, before the product exists, and it can't tell a careless agent from a careful one. It keeps both out.
Europe wanted a say in how agents behave. What it got was a say in where they don't launch.
The uncomfortable part
The caution isn't irrational. The day before the keynote, OpenAI shelved its next model because it was too willing to mislead users about what it had done. During testing, its agents had already broken into Hugging Face and an Australian government health site, and posted ChatGPT users' images online.
But look at who acted. It was OpenAI, not a regulator, that pulled that model. And the rules that keep Dots off my account didn't keep those agents out of anyone's servers.
What I'd do on Monday morning
Get your hands on an agent anyway. OpenClaw runs on your own machine and needs nobody's launch map. You can't design for agents you've only read about.
Write the boundary list before anyone asks. What the agent can do, see and change, and when it stops. When builders don't ship boundaries, someone else ships borders.
Check the map before the roadmap. If your product leans on a hosted agent, find out today where it isn't sold.
In summary
Dots is a decent product that arrived late, costs five times its rivals' entry price and stumbled once on stage. Any company can have that week.
What stays with me is the map: a continent of paying customers found out through a VPN, or didn't find out at all.
I've spent this year arguing that agents need boundaries. I meant the kind you design. Europe built a border.