A proactive agent, behind a waitlist
Anthropic opened a waitlist on 21 October for Claude Duet, a research preview it describes as a proactive agent that “understands what needs your attention”, takes on tasks of all sizes, keeps working while the user is away and waits for approval when it matters. It runs on Claude Opus 5.5 and is available to Pro and Max subscribers in the United States, through a standalone app on iOS and Android and through the Claude website.
The product page sets out three behaviours. Duet briefs the user every morning, drafts next steps and then goes quiet until something needs attention. It takes on larger jobs, Anthropic’s examples being setting up a meeting or building a dashboard, and reports back when finished. And it carries over the memory, skills and connected apps from the Claude app, in one thread. Anthropic says it works with more than 6,000 connectors and plugins from the Claude Marketplace.
The limits are the design
What distinguishes Duet from the agent demonstrations of the past year is how much of the announcement is about what it will not do. TechCrunch reports that Duet can make online purchases through Stripe’s Link or Shopify’s Shop Pay, with sensitive steps such as payments requiring the user’s approval, but that it will not trade securities or move money between accounts. A new classifier screens messages for sensitive information, and by default the agent declines requests for credit card numbers and government ID.
Anthropic also tells TechCrunch that Duet is not built for errands or booking tickets, and that it cannot show its browser activity as it happens. The agent runs tasks in the background on its own virtual machine, which MacRumors says includes its own browser. That combination — real spending authority inside a sandbox the user cannot watch live — is the trade-off the preview is testing.
Usage counts against the subscriber’s existing plan limits rather than carrying its own quota, CNBC reported.
Entering a crowded category
Duet lands in a field that filled up over 2026. TechCrunch lists Meta’s Muse, OpenAI’s Dots, Hark, Instinct, Grok Bot and Fo as the products it competes with. Each is making a version of the same bet: that the useful form of a consumer assistant is not a chat window you open but a process that runs without being asked and interrupts you only when it has to.
That bet has a cost attached. An agent that acts while nobody is watching needs a story about what happens when it is wrong, and Anthropic’s answer here is a narrow permission surface plus an approval step on anything that spends. The company’s own framing is about judgment rather than capability: the product page’s section heading is “Built for judgment”, and the feature it leads with is knowing when to stay quiet.
What the preview will show
A research preview behind a waitlist, limited to paying subscribers in one country, is not a product launch, and Anthropic has not said when or whether Duet becomes generally available or reaches other markets. The company also has not published figures on how often the agent’s proactive suggestions are accepted, which is the number that would say whether the format works.
The things worth watching are narrower than the launch coverage. Whether the approval step survives contact with users who find it tedious. Whether the classifier that declines card numbers holds up against people who want to give the agent their card number. And whether an assistant that cannot show its browser work in real time is one that professionals are willing to let spend their money.
