Michael JohnsonAn agent that reviews connected work apps, identifies important events and continues tasks in the background is Anthropic’s latest product. The company introduced Claude Duet on October 21, 2026 and opened a research preview via a waitlist for Pro and Max subscribers in the US. The app is available on iOS, Android and the web.
Duet is designed for professionals and small-business owners who need to bring together email, calendars and work services. Anthropic describes it as a proactive agent: it sends a morning summary, monitors connected accounts and takes on requests ranging from drafting a reply to preparing a report or dashboard. The practical question for users is which tasks they can delegate now and how they retain control over the outcome.
What the agent does between messages
Duet uses memory, skills and connected apps from a Claude account. Users assign a task in a chat, and the agent carries it out on a virtual machine with its own browser; the user’s computer or phone can be switched off. According to Anthropic’s description, accounts are checked every few hours, and notifications arrive when the agent thinks a person needs to step in.
Among the scenarios cited by the company are finding missed messages in Gmail and Slack, checking for unpaid invoices in QuickBooks or Stripe and preparing reminders, and preparing a timesheet for review. Anthropic also says Duet can prepare hiring materials, contracts, customer emails and Canva campaigns. These are product capabilities claimed by the company; it notes that the preview is still being tuned to recognize urgency.
Confirmation before sending or paying
Anthropic says Duet shows users an action and waits for approval before sending a message, making a payment, signing a document or deleting data. For app access, the company says it uses a cloud environment and handles credentials through the Claude Managed Agents model, in which the agent does not directly see passwords or login tokens. Certain actions, including securities trading and transferring money between accounts, remain prohibited even with user consent.
For businesses, the key distinction is between preparation and execution: the agent can gather data, prepare a draft or fill out a transaction, but consequential external actions require human review. Before connecting work accounts, it is useful to define which data and actions are permitted, who owns the accounts and who will review approval requests. A connected app can be disconnected.
Access and usage limits
In the preview, Duet uses the same limits as the rest of Claude, and background checks count toward usage too. Anthropic warns that large tasks—for example, a financial model or presentation—require more model calls and background work. It is therefore easier to test the agent with a limited set of tasks: start with sorting notifications and drafts, track how often it asks the user for input, and assess usage per completed task.
The company also specifies limitations of the early version: Duet may misunderstand a conversation or its priority, and a new event may occur between periodic checks. Personal requests such as booking tickets are currently a poorer fit than work tasks; users cannot yet see the agent’s browser. Access for Team and Enterprise plans has not been announced.
At launch, Claude Duet is best viewed as an assistant for monitoring and preparing work actions, rather than an autonomous operator. Pro and Max subscribers in the US can join the preview waitlist and try it on reversible tasks, while keeping human approval in place for messages, payments and documents. Such a test will quickly show whether summaries and background tasks justify the shared usage.