Imagine waking up to a short note from your AI. It’s checked your inbox, spotted an invoice that needs a decision, and drafted a sales forecast while you slept. That’s the pitch behind Claude Duet, the new background agent Anthropic announced on 21 October.
It’s not a chatbot you open and type into. Instead, it keeps working when you close the app. And that’s a pretty big shift in how we’re supposed to use AI.
So what is Claude Duet, exactly?
Claude Duet is a separate app from Anthropic, available on iOS, Android and the web through claude.ai. Right now it’s a research preview, which basically means an early version the company is still testing in public. Only US subscribers on Claude’s Pro and Max plans can join the waitlist.
According to TechCrunch, Anthropic is aiming it at professionals and small businesses that want help automating and managing everyday tasks. Under the hood, it runs on Claude Opus 5.5.
The idea is simple. You connect your apps once. Then Duet watches them, figures out what needs doing, and either does it or asks you first. Anthropic’s own one-line summary, quoted by Engadget, is that Duet “checks with you before it acts on your behalf.”
If you’re fuzzy on what makes something an “agent” rather than a chatbot, we wrote a plain guide on the definition of an AI agent. The short version: it pursues a goal on its own instead of waiting for your next message.
How Claude Duet actually works
Here’s the clever bit. Duet runs on its own virtual machine in the cloud, so it has its own computer with a browser. That means you don’t need to keep your laptop or phone switched on while it works.
Each morning, it sends you a short briefing of action items pulled from your connected apps. During the day, it checks those apps every few hours. Then it messages you only when something needs your attention or a decision. You can also choose how often it pings you and which topics it covers.
Why every few hours and not every minute? Anthropic was refreshingly honest about that one. As Engadget reported, the company said “checking nonstop would use up your plan’s limits much faster.” Duet’s work counts against the same usage limits as your other Claude tools, so a busy agent eats into your allowance.
The connections list is long, too. TNW reports more than 6,000 integrations, including Gmail, Slack, QuickBooks, Stripe, HubSpot and Shopify. Example jobs range from scheduling meetings to preparing sales forecasts and building simple web apps.
What Duet won’t do without asking
Giving an AI access to your email and payments sounds a little scary. So it’s worth looking at the guardrails, which are the rules that stop the agent from going rogue.
First, Duet asks for approval before sensitive actions. TNW lists sending messages, making payments, signing documents and deleting files. It also shows you its plan before acting. Next, Duet refuses some things even if you say yes. It won’t trade securities, and it won’t move money between your accounts.
There’s a privacy layer as well. TechCrunch says a new classifier screens each message to stop sensitive information from leaking out. By default, Duet also declines requests for credit card numbers and government ID. And according to Engadget, it never directly sees your passwords or sign-in tokens.
That said, guardrails only matter if they hold up in the real world. We’ve already seen how quickly AI agent hacking is getting cheaper, so we’d treat any agent with access to your accounts as a security decision as much as a productivity one.
Where Claude Duet still falls short
It’s a preview, and it shows. TNW notes that Duet can misjudge how urgent a message is. Personal errands, like booking tickets, are still a work in progress. You also can’t watch its browser while it works, which some people will find unnerving.
There’s a flip side to the convenience as well. Duet carries over your memory, skills and app connections from your existing Claude account. That saves setup time, but it also means the agent starts out knowing a lot about you.
Dan Shipper of Every, who had early access, called it a professional tool. But Shipper also said its email drafts were weaker than a rival agent’s. Another complaint was blunt: “There’s no real voice mode yet, just transcription.”
Anthropic seems to know the timing piece isn’t solved. “We’re continuing to test its judgment about when to message you,” the company said, per Engadget. Honestly, that’s the hardest part. An assistant that pings you too often is annoying, while one that stays quiet at the wrong moment is worse.
Why every AI company wants an agent right now
Duet didn’t arrive in a vacuum. OpenAI introduced its own personal agent, Dots, in September, and Meta has launched Muse. Engadget points out that Dots currently sits behind OpenAI’s Pro tier, which starts at $100 a month. TechCrunch adds a whole crowd of other personal agents, from Instinct to Grok Bot.
So the race has clearly moved from “who has the smartest chatbot” to “whose AI can do your chores.” We saw the same shift with the ChatGPT Chrome extension that clicks around your tabs. The big labs all want to be the assistant that sits in the background of your workday.
Meanwhile, the business logic is easy to follow. An agent that runs all day keeps you inside one company’s apps, plans and usage limits. That’s sticky in a way a chat window never was. It also helps explain the spending. Anthropic recently signed a $45 billion computing deal, according to TNW, which gives you a sense of the scale these companies are betting at.
Okay, but should you care about Claude Duet?
If you run a small business, maybe yes. Duet can even make purchases through Stripe’s Link wallet or Shop Pay, so it’s clearly built to handle real work. If your team has no dedicated operations person, an agent that drafts the forecast, flags the overdue invoice and books the meeting could cover real gaps. That’s time a founder could spend on customers instead. It could matter for small teams in Armenia and other emerging markets too, once access widens beyond the US.
But we’d wait for two things before handing it the keys. First, real-world reviews once more people get access. Second, the extra safety details Anthropic says it’ll share “in the coming days.”
So here’s the real question Claude Duet raises. It’s less “can AI do my admin?” and more “how much do I trust it to act when I’m not looking?” Your answer to that will probably shape how you work over the next few years.