Anthropic opens research preview of Claude Duet, a proactive agent that works while you're away

Anthropic opens a research preview of Claude Duet

Anthropic has begun rolling out Claude Duet, a new application in which Claude operates as a proactive agent rather than an assistant that only responds when prompted. According to Anthropic's announcement, the app surfaces what needs a user's attention, carries out assigned tasks, keeps running while the user is away, and seeks confirmation before acting on their behalf. The company says it built the product for professionals and small business owners juggling more than they can track.

The preview ships as a standalone app on iOS and Android as well as on claude.ai. It is included in Pro and Max plans and shares usage limits with the rest of Claude. Access is gated behind a waitlist, and the initial rollout covers Pro and Max users in the United States.

Built on Claude Opus 5.5

Anthropic says Duet is built around Claude Opus 5.5 and is designed to exercise judgment about what needs the user, what can wait, and when to check in first. Each morning the agent sends a short brief covering the most important items across a user's connected applications, then monitors those services through the day and messages the user only when something requires attention.

Users can also assign tasks of varying scope, from administrative work such as scheduling meetings to larger projects like building dashboards, financial or sales forecasts, prototypes, and web apps.

Work continues on a virtual machine

A key distinction from ordinary chat is that Duet runs in a virtual machine with its own browser, so background tasks proceed without the user's laptop or phone being switched on. The user receives a notification when a task finishes or needs review.

To conserve plan limits, Duet checks connected accounts every few hours rather than continuously. Anthropic acknowledges it is still testing the agent's judgment about when to message users and will tune this based on tester feedback; users can also specify what they want to hear about and how often.

Approval gates and hard limits

Anthropic says Duet reads, drafts, and prepares work on its own, but before it sends a message, makes a payment, signs a document, or deletes anything, it shows the user exactly what it intends to do and waits for approval. A few actions are blocked even with the user's consent, including trading securities and moving money between accounts. Any connected app can be disconnected by the user.

On the security side, Anthropic says Duet follows its framework for trustworthy agents, the credential design used for Claude Managed Agents, meaning it never directly sees passwords or sign-in tokens, and the auto mode classifier that guards against steps the user did not request. Defenses against hidden instructions and prompt injection come from its browser-use research. The company says red teams and an external firm penetration tested the app, which is also covered by its public bug bounty program.

Anthropic is also shipping a dedicated classifier and safety system prompt that reviews each outgoing message. It is designed to refuse sharing the strictest tier of sensitive information, such as credit card numbers and government identification. Other sensitive data, including health or personal financial details, goes out only when the user has explicitly instructed the agent to send it.

Thousands of connectors and payments by virtual card

Duet can draw on the more than 6,000 connectors and plugins in the Claude Marketplace. Anthropic lists example workflows including triaging Gmail and Slack messages, chasing unpaid invoices in QuickBooks or Stripe, preparing payroll in Gusto, drafting and posting job listings to LinkedIn, sending agreements through Docusign, drafting HubSpot follow-ups from deal history, turning a product launch into a Canva social campaign, and managing a Shopify store.

For purchases, Duet can pay with Link, Stripe's digital wallet, using a virtual card so the real card number is never shared, and it can check out with Shop Pay. Support for signing in to sites using credentials saved in 1Password is listed as coming soon.

Known limitations

Anthropic is candid about the preview's rough edges. The company says Duet may misread a thread or miscalibrate what is urgent, that it is stronger at work tasks than personal errands such as booking tickets, and that it cannot yet show its browser to the user. Duet lives in its own app, separate from the main Claude client, and how the two eventually fit together will be shaped by early feedback. Background checks count toward shared plan limits, and larger jobs such as building a financial model or a deck consume more model turns than a typical conversation.

Why it matters

The release marks a shift from chatbots that answer questions toward agents that hold the initiative: Duet initiates contact, triages attention, and works unattended. That raises the stakes on judgment and safety, which Anthropic has tried to answer with approval gates, a leakage classifier, and hard refusals on financial actions. If the approach holds up, persistent agents handling email, invoicing, and payroll could reshape daily workflows for small businesses. But Anthropic itself warns that calibration is still being tested, and the waitlist-and-preview framing signals it wants real-world evidence before any wider launch.