Claude Duet: Anthropic’s New AI Agent Comes to You — Here’s What It Actually Does
Almost every AI tool you use today is a waiter: it stands there politely until you ask for something. Useful, but it means the work only happens when you remember to start it. On October 21, 2026, Anthropic opened a research preview of something built the other way around. It’s called Claude Duet, and the whole idea is that it comes to you – watching what’s on your plate, working quietly in the background, and checking in when something actually needs a human.
If that sounds familiar, it should: this is the same proactive-agent wave we’ve already seen from other labs. But Anthropic’s take has a distinct personality – it’s built for work, it’s cautious by design, and it leans hard on knowing when to interrupt you and when to leave you alone. Here’s what Claude Duet does, how it keeps itself safe, what it still can’t do, and whether it’s worth joining the waitlist.
What Claude Duet actually is
Claude Duet is a standalone app where Claude works as a proactive agent rather than a chatbot. Anthropic built it for professionals and small business owners – people with more landing on them each day than they can realistically track. It runs on Claude Opus 5.5, the company’s most capable model, which Anthropic says is what gives Duet “the judgment to decide what needs you, what can wait, and when to check with you first.”
The clever structural choice is that Duet doesn’t live on your phone or laptop. It runs on its own cloud computer, with its own browser, so it keeps working after you close the app or shut your laptop. It also inherits the memory, skills, and connectors from your existing Claude account, so it starts out already knowing your context instead of from a blank slate. The app is available on iOS, Android, and at claude.ai.
How it works: morning briefs and background tasks
The rhythm is the part that makes Duet feel different. Each morning, it scans your email and calendar, highlights the priority items, and can even start the first steps – like drafting replies for you to review. Through the day, it checks your connected accounts every few hours so things don’t pile up unnoticed.
For bigger jobs, you hand it a task and it goes away and works, returning the result as a file: a research brief, a financial model, a slide deck, a dashboard. To do that work it can reach into more than 6,000 connectors and plugins in Anthropic’s marketplace – including Gmail, Slack, QuickBooks, HubSpot, Docusign, and Gusto – and for any site that doesn’t have a connector, it falls back to its cloud browser and clicks around like a person would. That’s a genuinely broad reach, and it’s why Anthropic positions Duet less as an assistant and more as a kind of always-on chief of staff.
The approval gates – the part that matters most
An agent that can send emails, move through your accounts, and spend money is only as good as its brakes. This is where Duet is deliberately conservative. Before it takes any major action – making a payment, signing a document, or deleting something – it stops and waits for your approval. And some actions are blocked outright even if you approve them, such as moving money around or accessing someone else’s account.
Payments are handled carefully, too. Duet can buy through Stripe’s Link wallet and Shopify’s Shop Pay, but every purchase needs your sign-off and runs on a one-time card that only works for that single transaction – so a mistake can’t turn into a recurring charge or a leaked card number. Anthropic also says Duet’s credential handling is modeled on its managed-agent work and that the agent never gets direct access to your passwords or account login tokens. For a tool you’re inviting into your inbox and bank-adjacent accounts, those guardrails are the whole ballgame.
What it can’t do (yet)
This is a research preview, and Anthropic is refreshingly upfront that it’s early. Duet can misread a conversation or misjudge how urgent something is, so you’ll want to keep an eye on what it flags and what it quietly decides can wait. It’s also tuned for work, which means its personal-errand skills – things like booking travel – are weaker than its professional ones.
There are practical limits as well. Access is currently US-only, for Pro and Max subscribers, through a waitlist that opens in waves; Team and Enterprise plans aren’t supported yet. And while Duet costs nothing extra during the preview, it draws from the same usage limits as your normal Claude chats – including the five-hour session and weekly caps. The catch worth knowing: the background work Duet does on its own, like your morning brief and those periodic account checks, also counts toward those limits, so you may hit them sooner than you expect. Anthropic is giving each new user one full limit reset, available until November 30, 2026.
How Duet compares to Dots, Spark, and Muse
Duet isn’t arriving in a vacuum – 2026 has become the year of the proactive personal agent, and every major lab now has an entry. The shape is strikingly consistent across all of them: an assistant that watches your accounts, works in the background, and acts on your behalf with your permission. We’ve already walked through the others, and it’s worth reading them side by side: OpenAI’s version in our breakdown of the always-on Dots agents, Google’s in our look at the 24/7 Gemini Spark assistant, and Meta’s errand-running Muse, and what you’re handing it.
So what makes Duet Duet? Three things stand out. First, it’s unapologetically made for work rather than errands, where some rivals chase the “run my whole life” pitch. Second, the approval-first, one-time-card posture is central to the design, not an afterthought – a reflection of Anthropic’s safety-forward identity. Third, it runs on Opus 5.5 and is betting that the hard part of a proactive agent isn’t doing the task, but having the judgment to know when to involve you. Whether that judgment holds up in daily use is exactly what a research preview is meant to find out.
Should you get on the waitlist?
If you’re a US Pro or Max subscriber who juggles a lot of email, calendar, and small-business admin, Duet is worth a spot on the waitlist – the morning-brief-plus-background-work pattern is genuinely well suited to that life, and the guardrails are reassuringly strict. A few things to keep in mind when your invite lands:
Start with low-stakes delegation. Let it triage your inbox and draft replies before you trust it with anything that spends money or signs your name. Watch how well it judges urgency for a week or two.
Mind the usage meter. Because its background checks eat into your normal Claude limits, decide how aggressively you want it working on its own – and use that free reset before November 30 to find your comfort zone.
Keep the approvals meaningful. The one-time card and the stop-and-ask gates only protect you if you actually read what you’re approving. The point of a proactive agent isn’t to stop paying attention – it’s to spend your attention on the handful of things that truly need it. On that, at least, Duet and you are on the same page.
Sources & further reading:
- Claude Duet: Anthropic Previews an AI Agent That Messages You First – VKTR
- Get started with Claude Duet – Anthropic Support
Related Reading
- OpenAI Dots: What These New Always-On AI Agents Actually Do for You
- Claude Opus 5 Is Here — What Anthropic’s New Model Means for You
- Anthropic Just Put Claude to Work for Your Financial Advisor — Here’s What It Means for You