Anthropic Launches Claude Duet, a Proactive AI Agent for Professionals and Small Businesses
Anthropic has announced Claude Duet, an AI agent aimed at professionals and small business owners, and is launching it as a research preview, an early test version for a limited group of users. The Verge reported the announcement on October 21, 2026.
Access works through a waitlist, and in the US the product requires a Pro or Max subscription. Approved users can run Duet as a standalone app on iOS and Android or through Claude on the web. The agent is built around Claude Opus 5.5.
Duet is designed to work proactively. Each morning it sends a brief on the most important items across the user's connected apps and messages, and it contacts the user directly only when something needs attention. To complete tasks, it draws on memory, skills, and connected apps from Claude. It also has a virtual machine, a sandboxed computer environment running inside a host system, with a browser at its disposal.
The agent requests approval before consequential steps, such as sending a payment or deleting something. Anthropic has also added a new dedicated classifier, a model that screens incoming requests, along with a safety system prompt designed to reject requests to share the strictest tier of sensitive information. Credit card numbers are cited as an example.
Anthropic describes Duet as an early version. It may misread a thread or miscalibrate what is urgent, and it is weaker at personal errands such as booking tickets. The company says its urgency classification is still imperfect.
The Verge characterizes Duet as a competitor to Meta's Muse and OpenAI's Dots AI agents. The three products target the same gap: a persistent assistant that watches several communication channels and takes multi-step action on the user's behalf.
Anthropic describes itself as a public benefit corporation dedicated to securing the benefits of AI and mitigating its risks. Its homepage characterizes the company as an AI safety and research company working to build reliable, interpretable, and steerable AI systems. In March, CIO Dive reported that Anthropic had given Claude computer access from a mobile device under a research preview available to Pro and Max users. Duet extends that trajectory into an ongoing, multi-application role.
In my view, the approval gate and the classifier are the design choices worth watching most closely. Earlier assistants mostly produced wrong answers that a person could catch before anything happened. An agent that operates a browser inside a virtual machine and can reach payment flows turns errors into actions, and a misjudged step is harder to reverse.
The broader context here is that the safety question is shifting from what a model says to what it does on a user's behalf.
The upside, as I see it, is concrete. A small business owner who starts the day triaging inboxes and calendars could regain time if the morning brief reliably surfaces only the items that need a decision. Whether it does so will depend on the urgency calls holding up outside a controlled preview, which is precisely what a research preview is designed to test.
