TL;DR: Anthropic has opened the waitlist for Claude Duet, an agent that tells you what’s on your schedule every morning, monitors your apps throughout the day, and does work without you asking, though it does ask for your permission for sensitive tasks. I’ll tell you what it does, how they’ve handled passwords, and why I’m still hoping to see a dual system with Cowork…
In theory, an AI assistant waits for you to type to it. You open the app, ask it to do something, it responds, and then goes back to standing by, waiting for the next command.
With Duet, it’s the other way around—he’s the one who writes first.
Anthropic unveiled it on Wednesday, October 21, as a “research preview”, a new app in which Claude acts as a proactive agent. Every morning, it sends you a short summary of the most important items in your connected apps (email, calendar, and whatever else you’ve given it access to); it checks back every few hours and only notifies you when something needs your attention. And if you assign it a task, it works on it in the background and notifies you when it’s done, or when it needs you to review something.
Basically, it’s not waiting for you to call her anymore.
What exactly does it do (and where does it do it)?
The technical detail that caught my attention the most is where it runs. Duet doesn’t live on your laptop or your phone; it has its own virtual machine in the cloud, with its own browser, so it keeps working even if you close your computer or turn off your phone. It runs on Claude Opus 5.5 and is currently available as an app for iPhone and Android, as well as on claude.ai (TechCrunch reported this the day after the announcement).
The tasks Anthropic cites as examples are fairly typical office tasks: scheduling meetings, putting together a dashboard with data from various tools, preparing a sales forecast, prototyping a website… According to BetaNews, it’s designed for professionals and small business owners, and Anthropic itself states on its website that it still has room for improvement when it comes to personal errands, such as buying tickets.
One of the customers featured in the launch, the owner of a restaurant chain with about 110 employees, sums it up with a phrase I understand perfectly: “I don’t need to open up 100 tabs. “It no longer needs to open a hundred tabs. This is a testimonial chosen by the company, note that it’s not an independent study.
What they won't let him do
This is where I think Anthropic has done its homework, or at least has wanted to make it seem that way.
Before sending a message, making a payment, signing a document, or deleting data, Duet tells you what he’s going to do and waits for you to say yes. There are some things it simply cannot do, even if you authorize it, such as trading stocks or transferring money between accounts. And in their security article, they explain that the agent never sees your passwords, browser cookies, or access tokens: The actual token is added by an external intermediary when the request is made, and there is only an empty slot inside the agent’s machine.
I like how they’ve handled password access. With Cowork, for example, this required using secure vaults so the system could access these passwords, including API keys, etc. Here, the agent doesn’t even see them.
There’s another detail I really liked. The verification that what the agent is going to do matches what you wanted is done outside the model that performs the task (they put it exactly like that: “outside the model doing the work”). In other words, they don’t ask the agent that’s going to act whether it’s sure of what it’s going to do, which would be a bit like asking a student if they’ve passed the exam.
For purchases, it uses a linked card, the one from Stripe, which is limited to one store and one amount, so your actual card number never leaves that system. Claude’s own X account summed it up this way on the day of the announcement: “the important decisions stay with you”; the important decisions are still yours.
The Fine Print
That said, not everything is perfect.
For starters, it’s only available in the United States, only for Pro and Max plans, via a waitlist, and with invitations sent out in batches; Team and Enterprise plans are left out for now. It uses up the same quota as the rest of Claude, and those background checks count toward that quota as well, so Anthropic has decided to check your accounts every few hours rather than every minute (its support documentation notes that something urgent might come up between two checks, according to BetaNews).
And the line I liked best in the entire announcement is the one where the company acknowledges its limitations: the agent can still misread a thread or misjudge what’s urgent. The fact that the manufacturer itself says this is, honestly, a point in its favor.
Then there’s the usual issue. CBS News reported that some security professionals are already urging people to think carefully about what they share with these agents. An assistant who checks your email, calendar, and documents every few hours is very useful… but it’s also a huge security risk.
Do you remember a few weeks ago when we talked about Muse, Meta’s virtual assistant, and how Amazon shut the door on it? (Here’s that article. ) Meta launched Muse on September 8, and OpenAI released dots on the 29th, according to BetaNews, so Duet is the third entrant in the personal assistant race.
When I saw it, I thought of Cowork
One of the main drawbacks of apps like Cowork is that they run on your device. If your device isn’t turned on, you’re out of luck (though it’s true that Anthropic has started moving part of its operations to the cloud, we’ll get to that in a moment).
In my case, I’ve been able to work around this limitation by using Cowork’s features on my phone, but it’s not the same, you can get some work done, but it’s very limited.
The cloud update has a specific date. On October 6, two weeks before Duet, Anthropic updated Cowork: as of that day, new tasks for the Pro and Max plans run on their servers rather than on your computer, and scheduled tasks work even if no device is turned on. However, according to their own help page, the app can only access files on your computer while the desktop app is open on that machine.
So when I saw Duet, I thought Anthropic was proposing a collaborative system with Cowork or Code, but it doesn’t look like that’s the direction they’re heading in.
In the pages Anthropic has published about Duet, I haven’t found a single mention of Cowork. It’s a separate app, with its own cloud instance, designed for mobile devices and, for now, only for individual accounts. They do share the Claude account, though: Duet pulls in the memory, skills, and connected apps and uses the same quota. What I haven’t seen anywhere is them passing work back and forth between each other.
The fundamental difference is who takes the initiative. With Cowork, you ask it to do things or set them up in advance; with Duet, the agent decides what deserves your attention and brings it to you.
Will we see dual systems?
Will we really see dual systems set up, for example, with Cowork and Duet, to get the best of both worlds? Local data and applications on one hand, and the power of the cloud on the other.
I think so, although I don’t know if Anthropic will set them up or if we’ll have to set them up ourselves.
First of all, because the components are already in the same account: the memory, skills, and connectors are the same. Second, because Duet solves the biggest problem when an agent works alone, access to keys without the agent seeing them. And finally, because what Duet lacks is what Cowork has: access to what’s on your computer, with the small caveat that, for that to work, the computer has to be turned on.
Keep in mind this is just my take, not an announcement from anyone. As of now, these are two similar products that don’t talk to each other.
My thoughts on this matter
What interests me most about Duet is where it’s headed: from an assistant you ask questions to an assistant that asks you questions.
And that’s where the question stops being technical. If an agent decides every morning what’s urgent in your work life, someone has to decide beforehand what’s urgent for you, and that doesn’t come pre-programmed. I’ve been spending months teaching mine what matters to me and what doesn’t, and it still gets it wrong from time to time (Anthropic, by the way, says the same thing about theirs).
It’ll make its way to Europe, I imagine, and it’ll find its way into business plans. When it does… it’s worth being clear about which apps you’re going to open up to it and which ones you won’t.
What about you? Would you set up a system like that, with your data stored locally and the agent working in the cloud while you sleep?
Leave me your comments—I’d love to hear from you.
Have a great week!
