Claude Duet: Anthropic’s New A.I. Agent Takes On Dots, Muse and Grok Bot

Set Trending Topics as a preferred source on Google.

Now Anthropic has an assistant that never sleeps, too: The Claude maker has just unveiled Claude Duet, an A.I. agent that works around the clock in the background, keeps an eye on the inbox and calendar, and only gets in touch when it needs something. With it, Anthropic is catching up with OpenAI, Meta and SpaceXSI, which have already launched their always-on agents. Unlike its rivals, though, Duet is aimed mainly at work rather than leisure.

Who Can Use Duet

Duet is launching as a so-called research preview, an early test version. For now, access is available only through a waitlist and only to customers with a Pro or Max subscription in the United States. Team and Enterprise customers are “not included yet,” according to Anthropic, which gave no timeline. Users in Europe are left out for now, as they are with many other new A.I. agents.

Duet runs as a standalone app for iOS and Android and in the browser at claude.ai. Anthropic is not charging extra for it, but the agent draws on the usage allowance of the subscriber’s plan, including for background work that users never see directly. Under the hood, it runs on Claude Opus 5.5. Internally, the project was reportedly called Claude Loop.

What Duet Is Supposed to Do

The principle is similar to that of its rivals: Duet works on its own computer in the cloud and keeps going long after the phone and laptop have been switched off. Its main features:

• Morning brief: Every morning, Duet sums up what is coming up, then only gets in touch when something needs the user’s attention.
• Big tasks: The agent is designed to set up meetings, build dashboards or prepare presentations, documents and spreadsheets for review.
• Business focus: As examples, Anthropic cites tracking down unpaid invoices and drafting reminders, preparing payroll and getting contracts ready to send.
• Thousands of integrations: Duet can tap into more than 6,000 connectors and plug-ins, including Gmail, Slack, Microsoft 365, HubSpot, Stripe, QuickBooks, Docusign and Shopify.
• Memory: Duet automatically carries over memories, skills and connected apps from the user’s Claude account. Everything is managed in a single chat thread, by text or voice.
• Approvals: Before Duet sends messages, makes payments, signs documents or deletes data, a human has to approve. Securities trading and transfers between accounts are reportedly blocked entirely, even with approval.

Anthropic’s Answer to the Competition

Anthropic has been under pressure to act for some time. OpenAI has introduced Dots, an agent that also runs around the clock on its own cloud computer and reaches out on its own, for example when the wrong attachment is about to go out with an email. Meta scored a hit with Muse: The app racked up more than 10 million downloads in the United States in just 40 days. And Elon Musk’s A.I. unit SpaceXSI has been fielding Grok Bot since August, which, awkwardly, now relies on Claude Opus 5.5 itself.

While Meta is targeting consumers who want to book trips or go shopping with Muse, Duet fits Anthropic’s strategy. The company makes most of its money from business customers; according to its I.P.O. prospectus, just two large customers accounted for about 20 percent of revenue in the first half of the year, as we reported. Accordingly, Duet is built around invoices, payroll and contracts, but in this first phase, through the Pro and Max plans, it is aimed more at freelancers and small businesses than at large corporations. Personal tasks such as booking tickets are reportedly less developed than the business workflows.

What Problems Duet Could Cause

An agent with access to the inbox, the books and payments is not without risk. Anthropic itself warns that Duet may misread a thread or misjudge how urgent something is. It can also miss something that arrives between two check-ins, because the agent only looks at connected services every few hours. According to the trade publication AlphaSignal, early testers reported a property viewing entered on the wrong day and a client call scheduled shortly before a flight.

There are other open issues as well:

• Security: Giving an agent access to email and payment systems opens the door to prompt injections, hidden instructions on websites or in messages. Anthropic counters with isolated credentials, filters for actions the user did not request and a bug bounty program.
• Overeager agents: Only a few days ago, Anthropic disclosed in its own report that Claude models in internal tests had worked around obstacles instead of stopping, for example by submitting forms on real websites.
• Little visibility: Users cannot watch in real time what the agent is doing in its cloud browser.
• Usage limits: Pro customers in particular are likely to hit their usage limits faster, because the regular check-ins and big tasks constantly consume computing power.