Skip to content

AI · 22 October 2026 · 5 min read

Anthropic’s Claude Duet will message you first. It still has to show you the payment before it sends it

A research preview opened on 21 October for Pro and Max subscribers in the United States, on a waitlist. The agent is built on Opus 5.5. It briefs you in the morning and keeps working when you leave. Trading securities and moving money between accounts stay off-limits even if you approve them.

Three phone screens showing a task list, a queue of drafts, and an email waiting for approval.
Anthropic’s product image for Claude Duet. A designed screen, not a customer’s phone.

Anthropic opened a research preview of Claude Duet on 21 October. It is a separate app, on iPhone, Android and the web, in which Claude is meant to come to you rather than wait to be asked. Each morning it sends a short brief of what it thinks matters across the apps you have connected. Through the day it watches those apps and messages only when it decides something needs you. You can also hand it a task: a meeting, a dashboard, a forecast, a prototype, a small web app. It is built around Claude Opus 5.5. For this preview it is included in Pro and Max plans in the United States, behind a waitlist, and it draws on the same usage limits as the rest of Claude. Background checks count toward those limits.

The permission is the part that is written down. Before it sends a message, makes a payment, signs a document or deletes anything, Anthropic says it shows you what it is about to do and waits. A few acts stay refused even if you approve them: trading securities, and moving money between accounts. A classifier is meant to reject requests for the strictest information, including card numbers and government identification. Purchases that are allowed go through Stripe’s Link or Shopify’s Shop Pay, and still with a confirmation. The company says the agent can reach more than 6,000 connectors in its marketplace. It also says this is an early version that can misread a thread or misjudge what is urgent, and that it is built for work. Booking a ticket or running an errand is, in Anthropic’s own words, not where it is strong. It has a browser in a virtual machine. It cannot yet show you that browser while it uses it.

A morning brief is not a finished job

An agent that writes first is a different product from a chat window. It is also a product whose mistakes arrive when you are not looking. Anthropic has put the consequential acts behind a screen you have to approve, and has taken securities and account-to-account transfers off the table entirely. That is a narrower envelope than the marketing sentence “it keeps working when you step away.” The work it keeps doing is the reading and the drafting. The sending is still yours.

What is not in the announcement is a rate. No count of briefs that were right, no share of drafts a person actually sent, no statement of how often the agent asked for an approval it should not have needed. The preview is a waitlist and a set of refusals. The refusals are the more concrete of the two.