Anthropic launches Claude Duet: an AI agent that works in the background — but asks permission before it acts

Anthropic opened the research preview of Claude Duet this week, an "always-on" agent that monitors connected apps such as Gmail and Slack, keeps working while the user's devices are off, and sends a morning summary.

File photo: A quiet office workspace where a background AI agent monitors email and messages overnight.
File photo: Derrick Coetzee from Berkeley, CA, USA, CC0 1.0 · source

Anthropic launches Claude Duet: an AI agent that works in the background — but asks permission before it acts

Anthropic opened the research preview of Claude Duet this week, an "always-on" agent that monitors connected apps such as Gmail and Slack, keeps working while the user's devices are off, and sends a morning summary. Access is currently limited to a waitlist for US-based Claude Pro and Max users — and the company itself admits the agent may misread context or miss updates.

The news: a research preview with a waitlist

Claude Duet has been launched as a research preview and sits on a waitlist for Claude Pro and Max users in the US, according to reporting by Livemint. The app is available as standalone iOS and Android apps as well as on the web.

The dating is less clear-cut than the launch headlines suggest. A Japanese review, which its author states is based on Anthropic's official announcement, product page and help center, dates the launch to October 21, 2026. Most media coverage, including an AFP report that says only "Wednesday," was published on October 22. During the preview period, Duet costs nothing extra beyond the existing Pro or Max subscription, according to the Japanese review.

How it works

What sets Duet apart from a traditional chatbot is that it works proactively instead of waiting for questions. According to Anthropic's own information, as reported by Livemint, the agent runs on a virtual machine in the cloud with its own browser — meaning it can keep working even when the user has switched off their phone or computer.

The model is locked to Claude Opus 5.5, and users cannot choose a different model during the preview period. The agent carries over the user's previously saved memory, skills and connected apps from the Claude account. In practice, it is meant to function as a kind of digital secretary: it sends a morning brief with key updates from the connected services and checks the accounts every few hours, according to Business Today.

Integrations

Anthropic itself states that Claude Duet supports over 6,000 connectors and plugins via Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn and Shopify, covering communication, finance, recruiting, marketing and e-commerce. The figure comes from the company itself and has not been independently verified.

An MSN-republished article describes a Stripe integration with a virtual card, where purchases must be approved before the transaction is completed, so that owners retain control over spending. This account, too, is based on secondary reporting and has not been confirmed against Stripe's own information.

The safety mechanisms

The safety architecture, according to Anthropic's own information as reported by Livemint, rests on explicit approval: Duet can draft messages, documents and transactions, but the user is responsible for approving the actual actions. This applies to sending messages, making payments, signing documents and deleting information.

In addition, the agent is supposed to refuse certain actions outright: securities trading and money transfers between accounts cannot be carried out even with user approval.

What Anthropic itself admits

Anthropic acknowledges that the agent may misread conversations, misjudge priorities or miss updates between checks. The company also states that support for personal tasks, such as booking tickets, is weaker than for work-related tasks.

No independent testing of the product exists so far. All concrete claims about what Duet can do come from Anthropic itself via secondary reporting.

The broader market push

The launch came the same day Walmart launched its own AI agent tools for commerce, according to AFP. The news agency describes a market development in which OpenAI, Google, Anthropic and Meta are investing heavily in agents — an industry phase in which vendors are moving from conversational interfaces toward software designed to carry out tasks autonomously in the background.

Open questions

Several key issues remain unresolved:

  • Availability outside the US. The preview requires a US-based Claude account, and there is no information on when — or whether — Duet will become available in other countries, including Norway.
  • Pricing after the preview. The fact that the preview is free beyond the subscription says nothing about pricing at a potential full launch.
  • Integration into the main app. According to Anthropic's documentation, as cited by the Japanese review, Duet will function as a separate app during the preview period, and the company will decide how to integrate it into the main app based on early user feedback.

It also remains to be seen how reliable the agent is in practice, not least given Anthropic's own caveats about misinterpretation and missed updates. So far, only the company's own promises about the approval mechanisms — not independent documentation — stand behind the picture of an agent that works all night but touches nothing without asking first.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.