Anthropic has announced Claude Duet, an AI agent designed to monitor connected applications and carry out work for professionals and small business owners. The service is launching as a research preview, with access limited to eligible Claude Pro and Max subscribers in the United States through a waitlist.
Duet is built around Claude Opus 5.5 and is available through standalone iOS and Android apps as well as Claude on the web. It enters the market alongside competing agents including Meta’s Muse and OpenAI’s Dots, as AI developers expand beyond conversational assistants into software that can perform tasks across other applications.
Anthropic describes a workflow in which Duet delivers a morning summary of important developments from a user’s connected apps, then continues monitoring those services throughout the day. Rather than simply responding when prompted, it is intended to contact users when it determines that an item needs their attention.
The agent can also receive assignments directly. Anthropic’s examples range from scheduling meetings to developing dashboards that involve dozens of tools. Other proposed uses include producing financial or sales forecasts, creating prototypes and building web applications. Those examples describe the product’s intended scope, rather than a guarantee that every assignment will be completed successfully.
Duet draws on memory, skills and application connections already associated with Claude. It also has access to a virtual machine equipped with a browser, giving it an environment in which to perform work. That setup distinguishes an agent workflow from a conventional chatbot exchange: instead of only suggesting steps or drafting an answer, the system can interact with tools to pursue a requested outcome.
Anthropic has included approval checkpoints for consequential actions. Before taking steps such as sending a payment or deleting something, Duet is designed to message the user for permission. These checkpoints place a human decision between the agent’s proposed action and its execution, an important distinction when software has access to business systems or sensitive records.
The company has also added a dedicated classifier and safety system prompt intended to prevent disclosure of the most sensitive category of information, including credit card numbers. Such safeguards address a different problem from action approvals: an agent may need to use information to complete a task without being permitted to share that information elsewhere.
The research preview comes with acknowledged limitations. Anthropic says this early version can misunderstand a conversation thread or incorrectly judge how urgent an issue is. Those weaknesses are particularly relevant to a product expected to filter messages and decide when to interrupt its user. The company also says Duet is less capable at personal errands, such as booking tickets, than at its intended professional tasks.
For now, joining the waitlist does not make Duet broadly available to every Claude user. Access requires a Pro or Max subscription and is restricted to the United States during the preview.
Sources: The Verge (AI)