Anthropic launches Claude Duet: a proactive agent that must ask before it pays, sends, or deletes
Access is initially via a waitlist for Pro and Max subscribers in the US, and the launch comes just under two months after Meta's Muse — but with an approval gate that separates what the agent may do on its own, what it must ask about, and what it is barred from.
Anthropic opened the research preview of Claude Duet on October 21, 2026, a proactive AI agent that messages the user, works in connected apps, and continues tasks after the app is closed. Access goes through a waitlist for Pro and Max subscribers in the US, and the launch places the company in what has become a race this fall around proactive personal agents — following Meta's Muse and OpenAI's Dots.
What Duet actually does
Duet is described by Anthropic as a "proactive agent": it brings forward whatever needs the user's attention, completes tasks, keeps working when the user walks away, and checks before acting on the user's behalf, the company writes in its launch statement as cited by Shattered.
Technically, the tasks run on a virtual machine with its own browser. That means the user's iPhone or computer does not need to be on — the agent works in the background independently of the device. It checks connected accounts every few hours. Tasks can range from something as simple as scheduling a meeting to building a dashboard, a financial forecast, or a web app, as iClarified describes with reference to Anthropic's launch materials. Duet is offered as a standalone app for iOS and Android, as well as via claude.ai on the web, and according to Shattered runs on the Opus 5.5 model.
The approval gate — and the hard limits
The most distinctive part of the design is the approval requirement. Duet can read, draft, and prepare work on its own, but must show the user what it intends to do and wait for approval before it sends a message, executes a payment, signs a document, or deletes anything. Some actions are blocked even with permission: trading securities and moving money between accounts. Approved purchases use Stripe's Link wallet with a virtual card, according to iClarified.
In practice, Anthropic thus distinguishes between three levels: things the agent does alone (read, prepare), things it must ask about (send, pay, sign, delete), and things it is not allowed to do at all (securities trading, account transfers).
At the same time, the product page acknowledges the limitations: "Claude Duet is a research preview. It can misread a thread or miscalibrate what's urgent, and it can miss something that arrives between check-ins," it says, as relayed by AI Weekly. Anthropic writes in its launch statement that "it can still misjudge what's urgent, and we'll continue adjusting it based on feedback from early users."
One practical detail may matter to subscribers: the background jobs draw on the same usage limits as chat. A Duet monitoring an inbox therefore uses the same Pro or Max quota as a chat window, according to AI Weekly.
6,000 connectors — a claim, not a test
The product page states that Duet "works with more than 6,000 connectors and plugins," with logos for among others Stripe, 1Password, Shopify, Gmail, Google Calendar, Slack, Intuit QuickBooks, DocuSign, Google Drive, HubSpot, Gusto, Microsoft 365, Notion, Canva, and Ramp. This is, however, a claim from the product page, relayed by AI Weekly and iClarified — neither has tested whether the connectors actually work in the preview.
Who is behind it, and how the launch is positioned
Mike Krieger, co-lead of Anthropic's Labs team, said in the launch video that the agent "can have discretion and judgment around when to be proactive, when to act on your behalf," according to Business Insider.
Development is led by Karan Ahuja, who according to his LinkedIn led product management at Meta Superintelligence Labs until July, before starting at Anthropic in August. On X, he has posted that he used Duet to build Duet, and that the product is made for professionals and small business owners, Business Insider writes with reference to Ahuja's posts.
The race for the proactive agent
Duet enters a field already in motion. Meta's Muse has, according to estimates from the analytics firm Apptopia, passed 14 million installs as of October 21 — an estimate, not an official figure from Meta. Apptopia also found that Claude users who installed the Muse app used Claude more than before, but still fewer minutes per day than they spent on the Meta product. This too is third-party estimates relayed via Business Insider, not independently confirmed findings.
OpenAI launched Dots at DevDay earlier this fall, according to Business Insider; Shattered says it was "just under three weeks" ago, but this is not confirmed by other sources in the case. Business Insider points out that the waitlist means Duet is not yet competing for the mass market against these products — Anthropic is thus choosing a controlled rollout rather than a broad launch.
What remains uncertain
All the product details above are secondary: they come from media outlets quoting the same product page and launch statement. Cryptobriefing wrote at its time of publication that the information came from a social media source and had not been independently verified through Anthropic's official channels; at the same time, several major outlets reported consistently the same day, including direct quotes from the company, which argues against the warning. There is no independent testing of the connectors, performance, or security, and no date has been given for full launch — only the waitlist for US Pro and Max subscribers. Any availability outside the US has not been confirmed.