Anthropic Launches Claude Duet, a Proactive Background Agent

Anthropic has launched "Claude Duet" in an initial research preview, a proactive AI agent for professionals and small business owners that completes tasks on demand, continues working in the background, and checks back with the user. The app is available on the web, iOS, and Android for Pro and Max subscribers in the United States, powered by the Opus 5.5 model.

The company stated that every morning, the agent sends a brief summary of key items across connected apps, monitors messages when something requires attention, and handles digital secretary-like administrative tasks such as scheduling meetings, building interactive dashboards, preparing financial or sales forecasts, and prototyping web applications. Users can interact with it via voice or text without prior setup or prompt formulas, leveraging Claude's memory, tools, and integrations, and returning results as a document, spreadsheet, or presentation for review. Published data indicates it works with more than 6,000 tools and extensions.

Anthropic says that "Duet" is built on its trusted agents framework, and that an automated classifier approves any action taken by the agent after checking messages for sensitive information leaks and verifying with the user before crucial steps. Internal teams at the company are already using it; Boris Cherny from the Claude Code team uses it while walking to review task lists, and Karan Ahuja, who leads the Duet team, uses it to schedule user interviews and even write code.

However, the current release is experimental, and it may misread conversation context or misjudge task urgency. Since the agent is designed around work, personal tasks like booking tickets do not receive the same priority, and it cannot view a browser. Anthropic says it will use feedback from early users to refine its performance before a broader rollout.

What Do These Terms Mean?

Proactive agent: An AI agent that initiates and follows up on tasks on its own instead of waiting for a prompt at every step, then checks back before important actions.

Automated classifier: A system that inspects and categorizes texts or requests, such as checking for sensitive information before allowing the agent to perform an action.

Trusted agents framework: A set of rules and safeguards that define an agent's permissions, monitoring methods, and requirement for human approval before sensitive steps.