Claude Duet aims to become the proactive assistant working in the background

Anthropic launches Claude Duet in research preview, a proactive agent capable of monitoring tasks, preparing actions, and requesting validation before acting. Access is via a waitlist, for Pro and Max subscribers in the United States.

Anthropic has launched Claude Duet, a proactive AI agent presented as a work assistant capable of spotting what requires immediate attention, executing tasks in the background, and requesting confirmation before any sensitive action. The service is offered in research preview, via a waitlist, to Claude Pro and Max subscribers in the United States. The goal is clear: to evolve Claude from a useful chat to a tool that truly keeps pace with a workday, without handing over the keys to the vault. The approach remains cautious, as Anthropic keeps humans at the center of validations.

An agent designed for already overloaded days

Claude Duet primarily targets professionals and small businesses juggling too many tools, tabs, and reminders. Anthropic describes an agent built around Claude Opus 5.5, capable of sending a brief summary of priorities from connected applications each morning, then monitoring these accounts throughout the day. When something requires a response, Duet alerts the user rather than acting alone.

The use case is not limited to messaging or scheduling. According to examples given by Anthropic and 9to5Google, the agent can help prepare dashboards, financial or commercial forecasts, prototypes, and even web applications. All this happens through a single thread, in text or voice, designed to handle both a quick request and a project that stretches over a week.

The line between assistance and autonomy remains very clear

Anthropic emphasizes this point: Claude Duet can read, write, and prepare work autonomously, but it must display the proposed action and wait for a green light before sending a message, making a payment, signing a document, or deleting anything. Transfers between accounts and securities trading operations remain prohibited. The framework is therefore closer to a very disciplined co-pilot than an agent left alone in the wild.

Technically, Duet uses the architecture of Claude Managed Agents, with separation of identifiers and action capabilities. Anthropic also explains having added a dedicated classifier and a safety prompt to block sharing of credit card numbers or government IDs. Sensitive information, such as health or personal financial data, requires explicit instructions. The company says it has tested the system with red teams and an external security firm, while including it in its bug bounty program.

What Duet can do, and what it cannot do yet

The service relies on more than 6,000 connectors and plugins via the Claude Marketplace, according to Anthropic. Among the announced integrations are Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva, and Shopify. Duet can also initiate professional purchases via Stripe Link or Shopify Shop Pay, with validation steps and virtual card management for Link. Support for 1Password and richer visual editing in Shopify are announced for later.

The downside of the system lies in its current limitations. Anthropic warns that Duet may misread some threads or overestimate the urgency of a task, that personal uses are less advanced, and that the user cannot yet see the browser used by the agent. Background checks consume usage limits of the plan, and heavy projects require more exchanges with the model. In other words, the agent promises time saved, but it does not work for free.

A limited preview, but revealing of the agent battle

Claude Duet remains reserved for Pro and Max subscribers in the United States, via a waitlist, on a standalone app for iOS and Android as well as on the Claude website. During this phase, it remains separate from the main application. Anthropic indicates that initial feedback will help decide how Duet and the classic Claude app should coexist in the long term.

This release is part of a broader sequence where proactive agents become a clear competitive focus among major generative AI players. 9to5Google compares Duet to Muse and Dots, OpenAI’s and Meta’s offerings in the same field. For Anthropic, the real issue is not just answering a question, but following work that progresses while the user does something else. It is less spectacular than a lab demonstration, but much more useful in an overflowing schedule.

Key points

  • Claude Duet is in research preview in the United States.
  • Access is reserved for Claude Pro and Max subscribers.
  • Anthropic cites Claude Opus 5.5 as the service’s base.
  • More than 6,000 connectors are announced via Claude Marketplace.
  • Human validation remains mandatory before final action.
  • Anthropic says it works with red teams and a bug bounty.

By the numbers

  • 2 plans — Pro and Max, the only eligible plans at launch.
  • 6,000+ — connectors and plugins announced by Anthropic.
  • 1 waitlist — access mode during the preview.
  • iOS and Android — platforms of the standalone app.

To read