Anthropic Launches Claude Duet Proactive Agent for Professional Users

Anthropic has introduced Claude Duet, a new proactive agent application designed to monitor user applications, complete assigned tasks independently even when the user is away from their screen, and request approval before taking action on behalf of the user. This launch targets professional workers and small business owners who face more work than they can manage, with the service currently available as a research preview through a waitlist for Pro and Max plan subscribers in the United States.

The application operates as a standalone tool on iOS and Android devices, as well as through the claude.ai web interface. During this preview phase, Claude Duet is included within existing Pro and Max subscription plans, utilizing the same usage quotas as other Claude services. The system is powered by the Claude Opus 5.5 model, which Anthropic states provides the artificial intelligence with the judgment to decide which matters require immediate user attention, which can wait, and which necessitate prior consultation before execution.

Each morning, the artificial intelligence delivers a brief summary of critical items from connected applications, sending messages throughout the day only when specific attention is required. Users can assign tasks ranging from administrative duties like scheduling meetings to complex projects such as building dashboards using dozens of tools, creating financial or sales forecasts, developing prototypes, and constructing web applications. The agent leverages memory, skills, and connected applications from the user's existing Claude account, allowing continuous interaction within a single thread for both short responses and week-long projects without requiring additional configuration.

Tasks continue to execute even when the user is not present at their device because Claude Duet utilizes a virtual machine with its own browser, eliminating the need to keep laptops or phones active. Users simply send a short message describing the work, and the system determines which applications to use and what steps to take behind the scenes, notifying the user upon completion or when review is needed. To prevent rapid consumption of plan quotas, the agent checks connected accounts intermittently rather than continuously, with Anthropic currently testing and adjusting the timing of these checks based on feedback from trial users who can also customize what information they receive and how frequently.

While Claude Duet can read, draft, and prepare work autonomously, it must display its intended actions and wait for user approval before sending messages, making payments, signing documents, or deleting any content. Certain actions remain prohibited even with user consent, such as trading securities or transferring money between accounts, though users retain the ability to disconnect applications from artificial intelligence access at any time. The system integrates with over six thousand connectors and plugins available in the Claude Marketplace, enabling capabilities such as finding missed messages in Gmail and Slack to draft replies, checking overdue invoices in Intuit QuickBooks or Stripe to create collection messages, and preparing payroll in Gusto by inputting hours, bonuses, and reimbursable expenses.

Additional functions include drafting job postings for LinkedIn publication after user editing and approval, preparing contracts for review and signing via Docusign, composing follow-up emails based on customer history and deals in HubSpot, and converting product launch tasks into social media campaigns in Canva. For purchasing goods for businesses, the agent uses Link, the digital wallet from Stripe, where the artificial intelligence confirms items with the user before Link requests expense approval and payment is made via a virtual card. The system also supports checkout at online stores through Shop Pay from Shopify, while login functionality using data stored in 1Password is listed as coming soon and will require explicit user permission. Anthropic states that Claude Duet employs years of accumulated agent security frameworks, including trusted agent structures used in Claude Code and Claude in Chrome, login data management similar to Claude Managed Agents that prevents the system from directly viewing passwords or sign-in tokens, and an auto mode classifier that approves actions. A dedicated classifier for security system prompts examines every message to ensure the agent, which reads emails and sends messages on behalf of users, constantly decides where information should go, designed to reject requests to share highly sensitive data such as credit card numbers and government identity documents. Other sensitive information like health or personal financial data can only be transmitted when the user explicitly instructs Claude Duet to do so. Usage quotas for Claude Duet are shared with the user's existing Claude plan and differ from standard chat consumption, with background account monitoring counting toward the quota while major tasks like creating financial models or slides require multiple model rounds and significantly more background work. Techsauce reported this story regarding the launch and capabilities of the new agent system.