AI Affairs, home

Thursday 22 October 2026

Technology

Anthropic launches Claude Duet agent in a US research preview

The assistant runs on Claude Opus 5.5 and can keep working after its app closes, but access starts with a waitlist for Pro and Max subscribers.

A hand holds a smartphone displaying Claude and several other AI app icons.
Photo: Solen Feyissa via Pexels

22 October 2026 — Anthropic launched Claude Duet on 21 October as a research preview of an AI agent that can work across connected apps and continue after its own app is closed. Access is opening through a waitlist for Pro and Max subscribers in the United States, 9to5Mac reported. The release puts an assistant designed to keep track of work between conversations into the hands of invited subscribers, rather than customers on Anthropic’s Team or Enterprise plans.

Key points

  • Claude Duet runs on Claude Opus 5.5 and works on its own cloud-based computer.
  • Invited US Pro and Max subscribers can use it on iOS, Android and the web; background work counts towards plan limits.
  • Anthropic says the agent asks for approval before sending messages, making payments, signing documents or deleting data.

Claude Duet works on a cloud computer

Anthropic says Duet uses a computer in the cloud to carry out work after the app is closed. The distinction matters: closing a window does not stop a task when the computer doing it is elsewhere. The company describes an agent that can check connected email and calendars, prepare work for review and return when it needs a decision. Users can give it instructions by typing or speaking.

Anthropic says Duet can draw on memories, skills and connections already associated with a Claude account, and search earlier conversations. A main conversation holds continuing work, while separate channels can share the same memory. For a larger request, the company says the agent can choose which connected apps to use and produce an editable file, such as a research brief, slide deck or dashboard. Where a service has no connector, Duet can use its cloud browser.

Checking unpaid invoices could lead to reminders drafted for review, using the Stripe connection Anthropic describes. Sending one would still require approval under the company’s stated rules. That separates the background work of finding something that needs attention from the decision to contact someone about it.

Anthropic also describes a morning brief assembled from email and calendar entries, followed by updates on continuing tasks. Trending Topics reported that Duet supports more than 6,000 integrations, including Gmail, Slack, Microsoft 365, HubSpot, Stripe, QuickBooks, Docusign and Shopify. Those connections give the agent places to look and work; the separate approval rules govern some of the actions it can take with what it finds.

Pro and Max invitations start in the US

Anthropic is making Duet available through standalone iOS and Android apps and on the Claude website for subscribers who receive an invitation, 9to5Mac reported. The company says invitations will go out in waves. It calls the release a research preview, and says Team and Enterprise customers are not included yet, although it plans to expand access to those plans.

Duet carries no additional charge for eligible subscribers, according to Anthropic, but its work uses their existing plan allowances. That includes activity in the background, when a subscriber is not directly chatting with it. The company warns that people using Duet could reach their limits sooner than they would through chat alone — a practical constraint for an agent designed to continue working between requests.

The release places Anthropic alongside OpenAI’s Dots and Meta’s Muse, other always-on agents described by Trending Topics. Anthropic’s examples for Duet include unpaid invoices, payroll preparation and contracts. In this first phase, however, access comes through its Pro and Max subscriptions rather than its plans for teams and larger organisations.

Duet requires approval for consequential actions

Anthropic says Duet is intended to seek a person’s approval before sending correspondence, making payments, signing documents or deleting data. The company gives another example of that division of labour: the agent can use Stripe to find unpaid invoices and draft reminders, leaving a person to decide whether a message should go out. Trending Topics reported that securities trading and transfers between accounts are blocked entirely, even if a user would approve them.

Those controls sit alongside limits to the agent’s background monitoring. Anthropic warns that Duet may misread an email exchange or get the urgency of a task wrong, Trending Topics reported. The agent can also miss an arrival between checks because it looks at connected services every few hours. A morning brief or later update would therefore reflect what Duet found when it checked, rather than a continuous view of every change.

Duet has a Tasks tab for following work in progress and identifying tasks that need attention. Anthropic says it posts updates in the main conversation when work is complete or a decision is required. Only the user can mark a task as finished.

Sources

Topics: Agents, Foundation models