Anthropic's Claude Duet Works in the Background

Host: Hello and welcome to AI Shadow Daily. I'm your host, and today we have one big story to dig into properly.

Co-host: And I'm the cohost. Today it's all about an AI agent that doesn't wait for you to type anything. It just gets on with things in the background.

Host: That's Claude Duet from Anthropic. We'll cover what it does, how it handles the scary bits like payments, how it stacks up against rivals, and what we still don't know.

Co-host: Then at the end we each pick one thing we'll be watching. Let's go.

Host: So let's start with the basics. Anthropic announced Claude Duet on the 21st of October, 2026, and it's being called a research preview.

Co-host: Research preview. I always translate that as, it's real and you can try it, but expect rough edges and expect things to change.

Host: That's a fair reading. It's an early release, so the company is still learning how people use it, and the product may well shift as a result.

Co-host: Okay, so what does it actually do? Because I've heard the word agent so much lately that it's starting to lose meaning.

Host: Fair. In plain terms, an AI agent is a piece of software that doesn't just answer a question. It takes steps on your behalf, like looking things up, reading, and acting, to get a job done.

Co-host: So less chatbot, more assistant who actually goes off and does stuff.

Host: Exactly. And here's the Duet twist. Each morning, it sends you a short brief on the most important items across your connected apps.

Co-host: So I wake up and there's a little summary waiting. That's quite appealing, honestly. Fewer tabs, fewer inboxes to wade through.

Host: And it doesn't stop there. According to Anthropic's own description, it then checks those apps every few hours.

Co-host: So it's not a one-off digest. It keeps an eye on things through the day.

Host: Right, and that's the key shift. Most AI tools sit idle until you speak to them. This one is meant to be working while you're not looking.

Co-host: Which is great when it works, and a bit nerve-racking when you think about it too hard. We'll get to that.

Co-host: Here's what I wondered. If it's checking things every few hours, does my laptop have to be open the whole time?

Host: No, and that's a nice detail. According to Anthropic, it runs on a virtual machine with its own browser, so it doesn't need your laptop or phone to be on.

Co-host: Okay, let me explain virtual machine for anyone who hasn't met it. It's basically a computer that lives inside another computer, usually in a data centre. You get a whole separate machine without owning the hardware.

Host: Lovely. So the agent has its own little computer, with its own browser, and it can go about its business while your devices are switched off or in your bag.

Co-host: That's a real difference from tools that borrow your own screen and your own browser. Those stop the moment you close the lid.

Host: Yes, and I think that's why the pitch works as a colleague rather than a gadget. A colleague doesn't need you standing next to them for the work to happen.

Co-host: Though a colleague who works while I sleep does raise the obvious question. What stops it doing something I'd regret?

Host: Which brings us neatly to the approval step. It's probably the most important design choice in the whole announcement.

Host: Anthropic says that before Duet sends a message, makes a payment, signs a document or deletes anything, it waits for your approval.

Co-host: So the four big ones, sending, paying, signing and deleting. Those are the actions that are hard to undo.

Host: Precisely. Think about it. A wrongly summarised email is annoying. A wrongly sent email, or a payment that has gone out, is a different matter.

Co-host: There's a general idea in this area called keeping a human in the loop. The software does the legwork, but a person makes the final call on anything with real consequences.

Host: And it's a sensible compromise. Full autonomy is risky, but an agent that asks permission for everything is just a nuisance. Drawing the line at those four actions is an attempt to find the middle.

Co-host: Now, I'll be a bit sceptical here. Asking for approval only helps if the request is clear. If I get a vague prompt at seven in the morning, I might just tap yes.

Host: That's a very human point, and honestly my own opinion is that the quality of those approval prompts will matter as much as the rule itself.

Co-host: And is there more detail coming on safety?

Host: There is. Engadget reports that Anthropic says it will share more on Duet's safety protections in the coming days.

Co-host: So the full picture isn't there yet. That's worth remembering. We know the approval rule, but not the rest of the safeguards.

Host: Quite. Until that detail lands, I'd treat the safety story as unfinished rather than settled.

Co-host: Okay, practical question. Can I get it today?

Host: Not straight away. Access is by waitlist, for Pro and Max subscribers in the US, according to MacRumors.

Co-host: So you need to be on one of those paid plans, be in the United States, and then join a queue.

Host: That's it. So if you're listening from outside the US, or on a free plan, this is one to watch rather than one to try this week.

Co-host: And I imagine a waitlist suits a research preview. They can let people in gradually and see what breaks.

Host: That's a reasonable guess, though it's our guess rather than anything Anthropic has said, so let's label it as that.

Co-host: Fair enough. Now, this isn't happening in a vacuum, is it? Other companies are chasing the same idea.

Host: No. Engadget calls Duet Anthropic's answer to Meta's Muse and OpenAI's Dots.

Co-host: So that framing is Engadget's take, not Anthropic's. But it tells you the big labs all want an assistant that lives in the background of your day.

Host: Yes, and it suggests the race has moved on from who has the cleverest chatbot to who can be genuinely useful when you're not typing.

Co-host: Which, as a user, I find interesting. The competition is now about trust and convenience as much as raw smarts.

Host: So let's pull back. Why should an ordinary listener care about a waitlisted research preview?

Co-host: For me it's the change in how we use AI. Instead of me going to it, it comes to me with a brief each morning.

Host: And that changes the stakes. An assistant that checks your apps every few hours has to be reliable, because you're no longer watching it work.

Co-host: Right, so there are things we simply don't know yet. How good are the briefs? Does it miss important items? How often does it ask for approval?

Host: None of that has been answered in what we have today. We know the design, the access and the approval rule, but not how it performs in real life.

Co-host: And the safety details are still to come, so that's another open question.

Host: Yes. Anthropic has promised more in the coming days, so we'll be paying attention.

Co-host: I think the honest summary is, promising idea, careful design on the risky actions, and plenty left to prove.

Host: Well put. A research preview is an invitation to test, not a finished verdict.

Host: Time to wrap up. Each of us picks one thing to watch. Cohost, you first.

Co-host: I'll be watching how people find the approval step in practice. Does it feel reassuring, or does it become something we all just click through?

Host: Mine is the safety detail. Anthropic says more is coming in the coming days, and I'd like to see what it actually contains.

Co-host: Good picks. That's a lot of waiting and watching, but that's the job.

Host: That's the show. The show notes list every source, and this episode was made with AI voices. Do follow AI Shadow for the daily videos. Thanks for listening, and see you tomorrow.

Co-host: Bye, everyone.