Skip to content
Indicops
AgentsGlobal4 min read14 views

Anthropic previews Claude Duet, a proactive agent that works from its own virtual machine

Anthropic has opened a waitlist for Claude Duet, a research preview in which Claude runs on a dedicated virtual machine, monitors connected apps and messages the user only when something needs attention. Built on Claude Opus 5.5, it ships as a standalone app for US Pro, Max and Ultra subscribers with hard limits on financial actions.

By Indicops Intelligence DeskSource: Anthropic

Anthropic has unveiled Claude Duet, a research preview that inverts the usual chatbot relationship. Rather than waiting for a prompt, Duet runs continuously on its own virtual machine with its own browser, sends a short morning brief across the user's connected apps, and interrupts during the day only when it judges that something needs a human decision. The laptop or phone does not need to stay on.

The product is built around Claude Opus 5.5 and inherits the user's memory, skills and connected apps from their Claude account, with a single ongoing thread for text or voice. It checks connected accounts every few hours to conserve usage limits and draws on more than 6,000 connectors in the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, DocuSign, HubSpot and Shopify. Purchases can be made through Stripe's Link wallet using a virtual card.

Anthropic has been unusually explicit about guardrails. Duet shows actions before sending messages, making payments, signing documents or deleting anything, and waits for approval. Some actions remain off-limits even with approval, including trading securities or moving money between accounts. A classifier blocks requests to share the most sensitive data categories, credentials are handled so the model never sees passwords, and the system has been penetration-tested by an external firm and placed under a public bug bounty.

Access begins with a waitlist for US subscribers on Pro, Max and Ultra plans, and the preview is bundled into those plans rather than priced separately. Anthropic concedes the limits plainly: Duet "still might misread a thread or miscalibrate what's urgent", personal errands such as booking tickets are less developed, and it cannot yet show its browser to the user.

Why it matters

Duet is the clearest statement yet from a frontier lab that the next product surface is not a chat window but a persistent, semi-autonomous colleague. The architecture choices, a separate VM, approval gates and a hard list of forbidden actions, are likely to become the reference design that enterprise buyers ask every agent vendor to match.

Indicops view

Indian enterprises and departments should read Duet as a template rather than a product to buy today; it is US-only and consumer-facing. The lessons transfer immediately, though: run agents on isolated infrastructure, log every action, gate anything irreversible behind human approval, and define a non-negotiable list of actions the agent may never take. We apply exactly this pattern when building agents over government workflows, where the irreversible actions are disbursements and record changes rather than card payments.

Back to the newsroom