Protected vs left behind

This week Apple announced HomeView, its first smart display for the home. It costs $299 and it will not be sold anywhere in the European Union. Apple's reason fits in one line of the press release: Siri AI is not available on iOS in the EU, so the device built around it stays out too.

I live in Spain, so I read that line twice.

The same launch brought a new HomePod mini, which will be sold here, for 169 euros. Its main reason to upgrade is the new Siri, which we cannot use. On Xataka's Loop Infinito podcast, Javier Lacort compared it to a car whose battery is in place but which the law only lets you drive 300 kilometers. The hardware is there. The part that justifies the price is not.

That leaves a question I keep turning over. Are we being protected, or are we being left behind?

The dispute is about the Digital Markets Act, the EU law that already forced Apple to open the App Store and the default browser to competitors. The European Commission wants the same for AI assistants. If Siri gets deep access to your iPhone, rival assistants should be able to get it too.

Apple says it proposed several ways to do that and the regulators rejected all of them. Its June statement spells out what it thinks the DMA requires: giving any AI system "nearly unlimited access to a user's device," along with the ability to act on it. "That includes the ability to read and send messages, make purchases, access files, and execute actions across any app." Apple argues that is unsafe, because AI systems can be hijacked. So it is holding Siri AI back from iPhone and iPad in the EU, with no timeline. The Mac and Vision Pro get it here, because the DMA does not treat those as gatekeeper platforms.

Lacort's view is that both sides have a point, and I agree with him. The Commission is right that if the only assistant allowed into your home, your calendar and your messages is Apple's, that is exactly the bottleneck the DMA exists to prevent. Apple is right that an assistant with access to your whole personal context is the place where opening doors needs the most care.

What strikes me is that Europe is pulling in two directions at once. Competition law wants the door opened to any assistant. Privacy law wants to know exactly who walks through it and what they take. Both instincts are reasonable. They point at the same capability.

Europe is not just debating that capability in theory. In the US, it shipped this summer.

Between August and the end of September, four companies released something they call a personal agent:

Each one runs on a computer in the vendor's cloud, keeps working when you are not looking, and does the things on Apple's list. It reads and sends messages, buys things, opens files and acts across apps.

None of them is fully available here. Grok Bot's tagging feature on X is blocked for accounts in the EU and the UK, and its finance connection is US-only. OpenAI's personal Pro plans exclude the European Economic Area, Switzerland and the UK for dots at launch. Muse is in the US and Canada with no date for Europe. Instinct publishes no country list, and its privacy policy has no GDPR section. None of the vendors has said the rules are the reason. The pattern is still hard to miss.

So I should be clear. I have not used these agents. I have read their documentation, their terms and their safety posts, and the reviews of people who could use them. What follows is an argument built from that record, not a hands-on review.

Read side by side, the four products split on one question: when the agent does something, who is doing it?

Some agents act as you. Grok Bot gives each user a persistent cloud computer where the Bots sign into your real accounts and keep working around the clock. Its documentation says plainly that separate Bots are not a security boundary, because every Bot shares every login. A dot lives inside your ChatGPT account with its own cloud computer and browser. OpenAI built real limits into it. Background research is read-only, a separate review step checks actions before they run, and password changes and money transfers always come back to you.

Some agents act for you. Instinct has no app. You text it, call it or email it, and it acts through its own phone number, email address and cloud computer. Muse sits closer to this side. It works in your accounts, but a second agent Meta calls Sentinel must approve every outbound request, and the main agent never sees your real passwords.

Acting as you spreads fastest, because nothing else in your life has to change. Its cost is that nobody can tell the agent apart from you. On October 1, a Grok Bot user's "CFO" Bot, which had read-only access to his bank, posted his personal balances and spending into his company's Slack. To the Bot, a channel called "Exec-team" looked like his own agent group chat. Nothing was hacked. The login worked and the message went out under his name.

Acting for you gives separation, but a new party now holds part of your life under its own terms. Instinct's terms say it may keep using data it already indexed after you disconnect an account, unless you ask for deletion. They also say agreements the agent makes are binding on you. And separation only helps if people understand the permissions behind it. A YouTuber let Muse handle his Marketplace buyers and chose "Allow Always" on a permission prompt. Muse shared his pickup address and told a waiting buyer "Yep I'm here!" when he wasn't home.

Look at that incident log again. It was written by early users in the US, with their own bank details and their own addresses. Europeans were spared the beta.

OpenAI's own system card for the model behind dots adds a number. When the number of chained tasks doubled from five to ten, the rate of moderate scope violations rose from 8.6% to 19.7%. Longer leashes produce more wandering. Apple's warning that agents can be hijacked is self-serving, but it is not wrong.

And the questions European law forces are exactly the ones some of these products have not answered. What is the legal basis for processing my inbox? How long is the data kept? Where does it live? A privacy policy without a section for those questions is a fair reason to wait.

You cannot build judgment about agents from other people's reviews. Apple's own statement says EU developers will not be able to test or use the new Siri AI features for their apps on iPhone. If you build software in Europe, you will design for agents you cannot touch and learn their failure modes secondhand. That frustrates me more than any missing gadget.

The cost also lands on consumers. We pay the same price for a HomePod with fewer features, or we get no product at all. Lacort made a sharper point. Apple uses the absence as pressure, so every device that reaches Europe carries an unwritten footnote: this is what you're missing.

And the delay does not stay small. While Europeans wait, the people building the habits, the products and the norms around personal agents are somewhere else.

The delay is not protection. It is only time.

Nothing in the current gap makes these agents safer when they arrive. They will come with the same designs, the same shared logins and the same "Allow Always" button. Regulation can decide who is allowed through the door. It cannot decide what the agent does once it is inside your accounts. That is settled by permissions, one click at a time, by the person using it.

Line up the incidents and they split into two kinds. Some are agents misreading the job. Instinct cancelled a WIRED reporter's delayed DoorDash order against her instruction to cancel only with a refund, and it cost her $64. Better models will make those rarer. Others are permissions doing exactly what they said, like the shared connections behind the CFO Bot or the seller's "Allow Always." Better models will not fix those, because the agent did what it was allowed to do.

So the useful thing a European reader can do with the delay is arrive with a policy instead of an impulse.

There is an irony I should admit. The personal agent I actually use from Spain is the one nobody had to approve for my market, because I run it myself.

I use Hermes Agent, and Hermes helped research this issue, so read this part with that conflict in mind. OpenClaw is the larger project of the same kind. Both run where you put them, use whatever model you choose and work in any country, because no vendor decides where they launch.

That freedom moves the security work onto the owner. In both projects, tools run directly on the host machine by default, and sandboxing is something you switch on. OpenClaw's year included a high-severity vulnerability, hundreds of malicious skills in its open marketplace, and tens of thousands of instances exposed to the internet, many through a default network setting their owners never changed. No regulator stopped any of that. Self-hosting does not escape the question of what the agent may do. It hands that question to you.

When these products reach Europe, and some will, I would decide two things before installing anything:

If a product does not let me express those two decisions, it is not ready for my life, whatever country it ships in.

Are we being protected or left behind? For now, a bit of both. I still want to try all four. But I would rather arrive late with a policy than early with "Allow Always."

Back to the regular Tuesday schedule with Parallelism vs Coherence: when do more agents reduce elapsed time, and when do they merely multiply the cost of making the answers agree?