Claude Duet Is a Proactive Agent — and Its Approval Rules Are the Template

Anthropic’s Claude Duet, announced October 21, 2026, is a research preview of an agent that works without being asked — and the way it limits itself is the part worth studying. Anthropic describes Duet as an app for professionals and small business owners who juggle more tasks than they can track. It sends a daily morning brief across connected apps, checks those accounts every few hours, and messages the user only when something needs attention.

The work runs on Duet’s own virtual machine with a browser, so the user’s device does not need to be on. It can reach more than 6,000 connectors and plugins in the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto and HubSpot, and it can buy things through Link with a virtual card. Tasks range from scheduling meetings to building dashboards, forecasts, decks, prototypes and web apps.

The guardrails are the design. Duet requires approval before it sends messages, pays, signs or deletes, and it cannot trade securities or move money between accounts. Anthropic says it never directly sees passwords or sign-in tokens, a new classifier blocks sharing of sensitive data such as card and government ID numbers, and the product was red-teamed and externally penetration tested. Anthropic also lists limitations: it can misread threads, it is weaker on personal errands, and it has no browser view yet.

For anyone ordering an AI adoption plan, this is a useful template. Autonomy is granted by action type — read, draft, spend, send, delete — not switched on as a single setting. The approval list marks where a mistake costs money or trust. A restaurant group owner with about 110 employees told Anthropic it means he does not need to open 100 tabs. That is the outcome to ship toward: agents that watch the inbox and the books continuously, with humans approving only the handful of actions that cannot be undone.

Access is by waitlist, starting with US Pro and Max subscribers.