OpenAI Dots vs Claude Duet vs Meta Muse vs Grok Bot vs many more: The AI Agent Race Has a “Europe Problem”

Every time a new AI agent launches, I have a little ritual.

I watch the beautiful demo. I imagine all the things it could take off my plate. The emails answered, the research done, the calendar magically reorganised while I’m having my morning coffee.

Then I skip the benchmarks, scroll past the pricing, and look for one very specific sentence.

Where is this available?

Lately, that sentence has been ruining an otherwise exciting autumn.

OpenAI launched dots on September 29. For personal Pro users, the European Economic Area, Switzerland and the UK are excluded.

Meta launched Muse in September. First the US and Canada, then Japan on October 23. Japan was its first market outside North America.

Anthropic opened the Claude Duet research preview on October 21. US only, Pro and Max subscribers, waitlist required.

Three impressive launches. Three products I’d genuinely like to spend time with. Three dead ends before I even get to the sign-up screen.

And I live in Milan, not on Mars.

I lead design work at Accenture Song and run multiple R&D experiments on the side. I spend a lot of time working with AI, thinking about agentic experiences, and trying to understand what happens when we stop using models as chatbots and start giving them actual responsibilities.

I’m exactly the sort of person these products are supposed to appeal to.

Yet the most important feature of an AI agent in 2026 might be something nobody puts in a keynote.

Whether you can actually use it where you live.

This isn’t going to be a complaint about Europe being left out. It’s something more practical: a look at six agent platforms, what I find interesting about each of them, and why the one I currently rely on isn’t necessarily the one with the most exciting launch video.

A quick disclosure: I pay for Grok Bot and have written extensively about Hermes. I can speak from experience about those tools, but for dots, Muse and Duet, I’m evaluating the published product design, not pretending I’ve tested products I cannot access on a personal plan.

And that’s a distinction worth making.

First, what are we actually comparing?

There’s something interesting happening across these six products.

We’ve spent the last few years talking to AI. Asking questions. Generating copy. Writing code. Getting suggestions.

Now the products are shifting toward something else: giving an AI its own computer, access to our tools, a memory of our work and permission to do things without us constantly supervising it.

That’s a much bigger change than another model getting better at reasoning.

Because once an AI can send an email, update a CRM, book something, publish a post or touch a client’s files, the conversation stops being purely about intelligence.

It becomes about trust.

Who has access to what? When does the agent need to ask? What happens when it makes a mistake? Can I separate two clients’ information? Will it keep running when my laptop is closed?

And, naturally, how much is all of this going to cost me?

Those are the questions I used to look at the six contenders.

The first big distinction is between hosted agents—dots, Muse, Duet and Grok Bot—which give you a managed computer in the cloud, and self-hosted agents—Hermes and OpenClaw—which give you much more control but also make you responsible for running the thing.

Neither approach is automatically better. They just come with very different bills, and not all of those bills are financial.

Three agents I’d love to try, from the other side of the Atlantic

OpenAI dots: the one that understands permission

The thing that interests me most about dots isn’t the model behind it. It’s the way OpenAI is thinking about approvals.

Dots has four possible behaviours for an action: proceed automatically, proceed if explicitly pre-approved, ask first, or hand the task back to you.

That may sound like a minor detail. To me, it’s one of the most important design decisions in the whole category.

Think about the difference between asking an assistant to draft an email and asking it to send one.

A human assistant understands that distinction. An AI agent needs to understand it too, consistently, across hundreds of actions.

OpenAI’s documentation makes the boundary explicit: “Asking for a draft doesn’t authorize sending it.”

I love that sentence.

It’s exactly the kind of principle I’d want written on the wall of any studio designing agentic experiences.

Dots also gets its own cloud computer, with the option to hand control back to you. Access to your personal machine is separate and opt-in.

On paper, there’s a thoughtful division between autonomy and human responsibility.

Unfortunately, my personal Pro account in Italy isn’t eligible.

There is a nuance here: OpenAI offers dots to eligible Business Premium workspaces in supported regions, and some Enterprise workspaces can access the beta with administrative approval. So Europe isn’t universally blocked across every plan.

But that creates another distinction.

An agent operating inside my employer’s environment is not the same thing as a personal assistant I can use for my own projects or freelance clients. Those are different worlds, with different access, rules and responsibilities.

I don’t want to blur them just to get access to a new product.

OpenAI dots documentation

Meta Muse: the one I’d probably use through WhatsApp

Muse might be the most interesting of the group from a consumer-experience perspective.

Not because it promises to do something radically different, but because Meta seems to understand where people already spend their time.

You can talk to Muse through its own app or directly in WhatsApp.

Anyone who has worked in Italy knows why that matters.

Clients reply on WhatsApp. Suppliers reply on WhatsApp. Your architect, your accountant, the person delivering your furniture. Sometimes even the people who really should be answering emails.

An agent that lives inside that behaviour has a meaningful advantage over one that asks you to adopt yet another interface.

But what really caught my attention was the architecture.

Meta gives each user a dedicated cloud environment called Muse Secure VM. There’s also a separate agent, Sentinel, responsible for controlling internet access.

Muse doesn’t get to read your passwords or payment details, and sensitive actions such as sending emails or making purchases require confirmation.

I haven’t tested any of this, so I’m not going to declare it the safest system on the market.

What I like is the underlying design idea: security isn’t just a set of warnings layered over an autonomous agent. It’s part of how the environment is structured.

That’s the right direction.

And yet there’s something almost comical about an AI agent designed to work inside WhatsApp being unavailable in Italy.

The country where I’ve seen entire businesses operate through WhatsApp voice notes can’t officially subscribe to it.

You have to appreciate the irony.

Meta’s Muse announcement

Claude Duet: probably the closest to how I actually want to work

Of the three, Claude Duet is the one whose positioning feels closest to my daily life.

It’s built around professionals and small business owners. Not just organising your dinner reservations, but managing work, following up on things, preparing documents and keeping track of projects.

It has channels for different areas of your life, shared memory, connected applications and its own cloud computer.

In other words, it feels less like a chatbot with extra permissions and more like the beginning of a proper digital colleague.

What I particularly appreciate is that Anthropic is fairly open about what doesn’t work yet.

Duet checks accounts at intervals, meaning urgent things can arrive between checks. You can’t watch its browser work or take over that browser. And all that background activity consumes your existing Claude plan limits.

These aren’t small limitations.

But I’d rather see them clearly explained than hidden behind another video of an AI apparently running an entire company while its founder goes surfing.

Duet is designed to ask before it sends messages, makes payments, signs documents or deletes things. Again, that’s the kind of boundary I want to see.

The problem is access.

US account. Pro or Max subscription. Waitlist invitation. Research preview.

By the time I’ve read the eligibility requirements, I’m already out.

Claude Duet documentation

The alternative: run your own agent

At this point, someone will inevitably say: why not just self-host?

And they’re not wrong.

This is where Hermes and OpenClaw become particularly interesting.

Both let you build an assistant that runs on hardware or infrastructure you control. You choose the model, configure the integrations and decide how the system operates.

You don’t need to wait for a company to announce Italy as a supported market.

That’s a pretty compelling proposition.

Hermes: freedom comes with an operations manual

I’ve spent time with Hermes, and there’s a lot to like.

It works across messaging platforms including Telegram, Slack, WhatsApp and Signal. You can configure its gateway, choose which models to use, schedule tasks and control who can interact with it.

The security model includes allowlists, user pairing, approvals for dangerous commands and different forms of isolation.

For work that needs deliberate separation, this is useful. I can design distinct deployments and credential boundaries for different projects rather than keeping everything on a shared vendor computer.

But those boundaries still need to be configured properly. Self-hosting isn’t an automatic security guarantee.

There’s also the slightly less glamorous reality of owning your infrastructure.

When the gateway stops, you fix it. When a dependency breaks, you update it. When a scheduled task fails overnight, there isn’t a vendor support team magically solving the problem for you.

Suddenly you’re not just a designer using an AI assistant.

You’re also its system administrator.

And the cost of an agent isn’t just the subscription. Sometimes it’s the Sunday evening you spend figuring out why it stopped responding.

Hermes Agent

OpenClaw: the open alternative

OpenClaw approaches the same problem through an open-source ecosystem.

It gives you an assistant that can operate across the messaging apps you already use, with its own gateway, integrations, scheduled tasks and controls.

I like the idea of open infrastructure in this space.

If AI agents become as central to our working lives as email or cloud storage, having credible alternatives to closed platforms matters.

Especially in Europe.

But OpenClaw comes with much of the same operational responsibility as Hermes. The software may be free, but your hosting, model usage, configuration, security and maintenance aren’t.

And it’s worth being clear about something: self-hosting doesn’t magically solve GDPR or client confidentiality.

It changes who’s responsible.

You still need to think about where the models process information, who can access credentials, which external services receive data, and whether the whole setup fits your obligations.

Ownership is useful. It’s not a compliance certificate.

OpenClaw

So why am I actually using Grok Bot?

After all of that, the answer is almost embarrassingly simple.

Because I can use it. And I love it. And because it does what I need without turning me into a full-time infrastructure engineer.

Grok Bot is the hosted agent platform that currently makes the most practical sense for my personal workflow in Italy.

The entry-level Premium plan is advertised at $30 per month, with a weekly usage allowance. It gives me a persistent computer in the cloud, access to a browser and connected applications, and the ability to schedule work that continues when my own computer is closed.

An agent that only works while I’m actively sitting in front of it is still useful, but it’s not quite the same category of product.

What I want is something that can take a task, make progress on it and come back when it needs a decision.

I’ve already written about how I’d build a design studio around a network of agents doing the operational work that creative people generally don’t enjoy: research, administrative tasks, scheduling, preparation, follow-ups.

Grok Bot gets closer to that model without asking me to build and maintain the entire underlying system.

I can create different Bots for different roles, give them access to tools and establish rules around when they need my approval.

That’s genuinely useful.

But here’s where I’d be careful.

The trade-offs are real

The biggest limitation is that all your Grok Bots share the same persistent cloud computer.

Different Bots don’t mean isolated environments.

They can share files, browser sessions and credentials. The vendor’s own documentation explicitly warns against treating separate Bots as a security boundary.

For a personal workflow, that might be completely acceptable.

For different clients with strict confidentiality requirements, it can become a problem very quickly.

That’s one reason I still want a self-hosted option available for projects requiring stronger separation.

The second trade-off is pricing. The monthly subscription includes a weekly usage allowance, but additional work can generate usage-based charges.

It’s not quite the same as paying for a conventional SaaS seat and forgetting about it.

The third is control.

Grok Bot chooses which model to use for each task. I generally like not having to manage that decision, but I know plenty of people who would rather pick their own.

And then there’s data.

Grok Bot requires cloud data storage, and its hosted infrastructure is not a substitute for checking the data-processing terms of a client engagement. Its security documentation describes US-hosted computers, which matters when residency and contractual restrictions are involved.

This isn’t the environment I’d use indiscriminately for confidential client information.

Even the European availability story has an asterisk.

You can use the product here, but the feature that lets you tag @bot directly on X isn’t available to EU or UK accounts yet.

So apparently no agent launch is complete without at least one feature I can’t use from Milan.

The Grok Bot FAQ · Pricing · Security documentation

The agent I actually want doesn’t exist yet

Here’s the interesting thing about looking at all six products together.

I don’t think any one of them has solved the problem completely.

But each is showing us a piece of what a mature agent experience could become.

I’d take the approval model from OpenAI dots, because the difference between asking for help and authorising an action needs to be crystal clear.

I’d borrow Meta’s approach to isolating the agent environment and controlling network access.

I’d take Duet’s work-first experience, shared memory and organisation around ongoing responsibilities.

I’d keep the ownership and portability that make Hermes and OpenClaw attractive.

And I’d preserve the simplicity of getting started with Grok Bot, because the best infrastructure in the world means very little if using it becomes a second job.

Put all of that together and you start to see the product I actually want.

Not six disconnected chat windows.

Not fifteen agents with elaborate job titles pretending to be a company.

One reliable place to delegate work, with sensible permissions, clear boundaries, enough memory to understand what matters, and the ability to move my workflows if I change platforms.

An assistant that knows when to act, when to ask, and when to get out of the way.

That, to me, is where agentic experience design gets genuinely interesting.

Availability is a design decision, too

There’s a broader point here for anyone building digital products.

We tend to talk about AI availability as a commercial or regulatory footnote. Something that sits at the bottom of the launch announcement, separate from the product experience.

I think that’s a mistake.

Availability is part of the user experience.

If I can’t access a product, I can’t learn its interaction patterns. I can’t experiment with its boundaries. I can’t discover where it fails. I can’t develop the habits that make it valuable.

And I certainly can’t recommend that a team adopt it on the strength of an impressive demo.

For designers working in Europe, this matters.

We’re watching an entire category of software evolve through products that some of us can’t yet touch.

I’m not going to speculate about the precise reasons for each company’s regional rollout. There may be different legal, operational and commercial considerations, and those decisions belong to the vendors.

But the practical result is the same.

The experience we get to design around is not always the experience being demonstrated onstage.

And that changes which tools we choose.

Five questions before you give an AI agent your work

I’ve stopped asking which agent is the smartest.

At least, that’s no longer my first question.

Before committing to one, here’s the checklist I’d use:

1. Can I actually use it? Not in theory, not through a VPN, but from my country, with an account and subscription I can legitimately maintain.
2. Whose computer is doing the work? Is it a shared machine, a dedicated environment or infrastructure I control? What other tasks and credentials can access it?
3. When does it ask permission? I want to know exactly what happens before it sends a message, spends money, deletes a file or publishes something.
4. What does it really cost? Subscription, usage, hosting, maintenance and the time required to keep it working.
5. Can I leave? If a better agent arrives tomorrow, which of my workflows, skills and data can I take with me?

That last question is becoming increasingly important to me.

AI products are evolving too quickly to build an entire working life around the assumption that one vendor will always be the right choice.

I’d rather own portable skills, clear workflows and a few good integration patterns than become dependent on one beautiful interface.

So, which one would I pick?

Today, for my personal setup in Europe, it’s Grok Bot.

Not because I think it has won the agent race. Not because I believe every product decision behind it is correct.

And definitely not because I’m comfortable giving a hosted agent unrestricted access to every part of my working life.

I use it because it’s available to me, it works for my lifestyle and work life, it’s useful, and the convenience is worth the compromises for the tasks I choose to give it.

I keep Hermes in the picture for situations where I want more infrastructure control and deliberately separated environments. OpenClaw remains another credible self-hosted route.

And I’m watching dots, Muse and Duet very closely.

If they become available to me, I’ll happily reconsider my setup. That’s the whole point.

There’s something slightly absurd about living through one of the most exciting periods in software history and having to begin every product evaluation by checking whether my country is on a list.

But that’s where we are.

The agent I want most may not be the agent I can use today.

And until that changes, I’ll choose the one that can actually do the work, on terms I understand, with boundaries I’m comfortable defending.

Because the real question isn’t which AI is most impressive in a demo.

It’s which one I’m willing to trust with my Tuesday morning.

And that decision, at least for now, is still mine.