Anthropic’s new AI agent Claude Duet can work while users are away

# Anthropic Introduces Claude Duet, an AI Agent Designed to Manage Work Proactively

Anthropic has introduced Claude Duet, a new artificial intelligence agent designed to handle tasks proactively, as the company expands its efforts in the increasingly competitive market for AI assistants that can act on users’ behalf.

The product targets professionals and small business owners seeking help with organizing their workloads, managing connected applications and completing routine tasks with minimal instructions. Unlike conventional chatbots that primarily respond to individual prompts, Claude Duet is designed to monitor connected applications, identify issues requiring attention and continue working in the background.

Powered by Anthropic’s Claude Opus 5.5 model, the agent prepares a daily morning briefing highlighting important developments across connected accounts. It continues monitoring those applications throughout the day and alerts users when a decision or intervention is needed.

Anthropic says Claude Duet includes safeguards intended to prevent sensitive information from being shared without authorization. Information concerning health or personal finances is shared only when users explicitly instruct the agent to send it.

The system can also assist with online purchases through Link, Stripe’s digital wallet, and Shopify’s Shop Pay. According to Anthropic, Claude Duet confirms the selected item with the user, while Link requires approval for the payment. The purchase is then completed using a virtual card, keeping the user’s actual card number from being shared with the merchant.

Users can assign Claude Duet a wide range of responsibilities, including scheduling meetings, drafting follow-up emails, creating dashboards, preparing financial forecasts and developing web applications. The agent operates through its own virtual machine and browser, allowing it to continue working even when users step away from their devices.

The platform can also connect with more than 6,000 connectors and plugins available through the Claude Marketplace, giving it access to a broad range of business applications and digital services.

Claude Duet is being offered as a research preview through a waitlist for subscribers to Anthropic’s Pro and Max plans in the United States. The standalone application is accessible on iOS, Android and the web through Claude.ai.

Anthropic has emphasized that users remain in control of consequential actions. Claude Duet seeks approval before sending messages, making payments, signing documents or deleting information. By default, it also refuses requests to share highly sensitive information, including credit card numbers and government identification details.

Certain activities remain prohibited even when a user grants permission. These include trading securities and transferring money between accounts, reflecting the limits Anthropic has placed on the agent’s ability to perform sensitive financial actions.

The launch comes as AI developers increasingly shift their focus from tools that generate text or answer questions toward agents capable of coordinating multiple tasks and reducing the amount of direct supervision required from users.

However, greater autonomy also introduces concerns about privacy, security and unintended actions. An incident reportedly involving Meta’s Muse AI, which allegedly negotiated a Facebook Marketplace transaction, shared a user’s address and arranged a late-night pickup without approval, has highlighted potential risks associated with AI systems acting independently.

As these products become more capable of accessing applications and making decisions, the effectiveness of their safeguards will be central to user trust. Claude Duet’s combination of background monitoring, connected services and approval requirements illustrates Anthropic’s approach to building a more proactive assistant while retaining human oversight over important actions.

The product’s research-preview phase will provide an opportunity to assess how well it handles real-world workflows, protects sensitive information and balances automation with user control as competition in the AI-agent market intensifies.