Anthropic’s Claude Duet: The AI Agent That Gets to Work Before You Ask

For years, AI assistants have waited around for you to give them something to do. You open a chat window, type a prompt, and get an answer back. That loop worked fine for questions. It never worked very well for the messy pile of small tasks that fill a workday. On October 21, Anthropic took a different swing with Claude Duet, an AI agent that starts working before you ask it to and only checks in when something needs your attention. Right now it is a research preview open through a waitlist to Claude Pro and Max subscribers in the US.

What Claude Duet actually does

The basic pitch is simple: instead of a chatbot sitting idle between your questions, Duet behaves like a proactive assistant. Each morning it sends you a brief that covers the items across your connected apps that need your attention. You hand it tasks by message, and it runs them on its own computer in the cloud, using your apps the way you would, then reports back when the work is done or when a decision needs you.

It is built on Anthropic's Opus 5.5 model, and the memory, skills, and connectors you already have in the Claude app carry over. So if you have already connected your calendar, email, or other tools, Duet can see the same context. Anthropic says the agent asks for approval before sensitive actions such as sending messages or making payments. That matters, because an agent that can act on its own is only useful if you can trust the moments when it pauses.

Why the timing makes sense

The launch did not happen in a vacuum. Meta has Muse, an always-on assistant line of products, and OpenAI has Dots, its own always-on agent feature. The Verge framed Claude Duet as Anthropic's answer to both. OpenAI Japan's president even said on October 20 that Japan ranks second in the world in weekly active users of Dots. The industry is clearly moving from "ask me anything" chat windows to assistants that live alongside your workday.

Anthropic seems to be betting that restraint is the differentiator. Where rivals emphasize always-on presence, Duet's pitch is the opposite: it works on your behalf and reaches out only when something needs you. That is a design choice aimed at a real complaint. Always-on assistants can feel like another feed demanding attention. Duet promises the opposite experience, which is silence until something matters.

How to get access

This part is narrow. Claude Duet is available only as a research preview through a waitlist, and only for Claude Pro and Max subscribers in the US. It ships as a standalone app on iOS and Android, and it also runs on the web at claude.ai. Anthropic has not announced when it might expand beyond the United States or open up to free-tier users.

The research preview label means you should expect rough edges. Preview agents sometimes misunderstand tasks, take longer routes to finish them, or ask for clarification at odd moments. If you join the waitlist, treat Duet as a test drive rather than a finished product you can rely on for anything time-sensitive.

The practical questions to ask before using it

A proactive agent raises a different set of worries than a chatbot. A chatbot can only act when you tell it to. Duet can run tasks on its own cloud computer using your apps, which means the security model has to be stronger. Anthropic's answer is the approval step before sensitive actions. That is a sensible line, but the details matter: which actions count as sensitive and how approvals reach you.

There is also the question of cost. Running tasks in the cloud means compute time. Since this is a research preview for paying subscribers, Anthropic is presumably absorbing the cost while it studies how people use it.

Another thing worth watching is how Duet handles the morning brief. A brief that consistently surfaces the right three things is worth a lot. A brief that dumps twenty notifications on you is just another inbox.

Where this fits in the agent race

Zoom out and the picture is clear. OpenAI announced on October 21 that its models and Codex are coming to Google Cloud. Walmart announced an Agent Gateway that lets approved AI agents search its inventory and buy on a customer's behalf. Everyone is building plumbing for agents to act in the real world.

Claude Duet sits on the consumer and professional end of that same trend. The underlying bet across the industry is that the next phase of AI is not better answers but reliable action. Whoever makes agents that do the right thing without constant supervision wins the next wave. Anthropic's version of that bet is an agent that talks less and does more, and that checks with you only when it genuinely needs you.

Whether that bet pays off depends on the boring details: the accuracy of the morning briefs, the reliability of task execution, and the trustworthiness of the approval flow. If those three land, Duet could become the kind of tool that quietly reorganizes how people start their workday. If any one of them wobbles, it becomes a demo people try once. For now, the waitlist is the way to find out which one it is.

Frequently Asked Questions

What is Claude Duet?

Claude Duet is Anthropic's proactive AI agent, released October 21, 2026 as a research preview. Instead of waiting for prompts, it sends you a morning brief of items needing attention across your connected apps, runs tasks on its own cloud computer, and contacts you when work is done or a decision is required. It is built on Anthropic's Opus 5.5 model.

Who can use Claude Duet right now?

Access is through a waitlist limited to Claude Pro and Max subscribers in the United States. It is available as a standalone app on iOS and Android, and on the web at claude.ai. Anthropic has not announced a timeline for expanding access to other countries or to free-tier users.

Is Claude Duet safe to use with my apps?

Anthropic says Duet asks for your approval before sensitive actions such as sending messages or making payments, and your existing memory, skills, and connectors from the Claude app carry over. As a research preview, it is still being tested, so it is wise to start with low-stakes tasks and review what the agent does before granting it access to anything critical.