Claude Duet Is Here: Anthropic's AI Agent That Works While You're Away

Anthropic says Claude Duet supports over 6,000 connectors and plugins. But its public Marketplace shows only about 1,350, Microsoft 365 isn't supported in Duet, and the AI that's supposed to work proactively only checks your accounts every few hours. So how much of Anthropic's new AI agent is actually working, and how much is still a promise? Launched October 21, Claude Duet is Anthropic's new proactive AI agent powered by Claude Opus 5.5. Unlike a traditional chatbot, Duet can contact you first, monitor connected accounts, prepare morning briefings, follow up on forgotten conversations, and complete delegated tasks using its own cloud computer—even after you close the app. But there's a catch: every background check consumes your existing Claude usage allowance. Claude Duet is currently a research preview for U.S. Claude Pro and Max subscribers, with plans starting at $20 per month. There's no separate Duet subscription fee during the preview, but its background activity shares the same usage limits as regular Claude. And the deeper you look, the more interesting the trade-offs become. Anthropic advertises more than 6,000 connectors and plugins, but that number hasn't been independently verified. Duet names integrations including Gmail, Google Calendar, Google Drive, Slack, Shopify, Stripe, QuickBooks, HubSpot, DocuSign, Canva, and Gusto. Yet some integrations have restrictions, and others aren't available at all. We break down what Claude Duet can actually access, why its proactive monitoring isn't truly real-time, and how its cloud computer handles complex workflows without requiring your phone or laptop to stay online. Anthropic's demonstrations show Duet building business dashboards, preparing investor meetings, tracking customer churn, managing invoices, and coordinating work across multiple business systems. One example, Marlo Pulse, combines information from sales, payroll, customer support, banking, and CRM tools. But these are Anthropic's demonstrations—not independently verified reliability benchmarks. Then there's one of Duet's most consequential capabilities: spending money. Through Link by Stripe and Shop Pay, Duet can find products, request your approval, and complete purchases using merchant-specific, amount-limited virtual cards. It must pause before payments, messages, signatures, and deletions. Securities trading and account-to-account money transfers remain prohibited. So how autonomous is this agent really? We also examine early independent testing from Dan Shipper at Every and Claire Vo. Their experiences reveal promising interfaces and useful automation, alongside slow responses, limited proactive nudges, weaker email drafts, connector restrictions, and missing desktop functionality. Security raises another important question. Duet processes emails, websites, and other third-party content, making prompt injection a genuine concern. Anthropic describes safeguards for credentials, sensitive information, and high-risk actions, but hasn't published Duet-specific reliability or security failure rates. Finally, we compare Claude Duet with competing AI agents from OpenAI, Microsoft, Google, and Meta. From webhook-triggered automation and local computer control to persistent cloud execution and pricing, each takes a different approach to building an AI assistant that works without constant prompting. The central question isn't whether Claude Duet can perform impressive tasks. Anthropic has demonstrated that it can. It's whether Duet can consistently recognize what matters, make the right decisions, and complete real work reliably when nobody is watching. That's the part nobody has independently measured yet. Subscribe to BitBiased for evidence-based AI news, technical breakdowns, cybersecurity analysis, and honest coverage of the next generation of AI agents. CHAPTERS 00:00 Claude Duet: 6,000 Connectors, but What's Actually Working? 00:38 What Claude Duet Actually Is 01:31 The AI That Contacts You First 03:00 Why It Only Checks Every Few Hours 04:08 Does "6,000 Connectors" Mean 6,000 Working Apps? 05:23 A Cloud Computer That Keeps Working After You Close the App 06:39 The Real Work Anthropic Shows It Doing 07:56 Can It Spend Your Money? 09:16 What the First Independent Testers Found 10:24 What Can Go Wrong, and What Protects You 11:55 How Duet Stacks Up Against Rival Agents 12:50 What Has Actually Been Proven #ClaudeDuet #Anthropic #ClaudeAI #AIAgents #AIAutomation