Anthropic Releases Research Preview of Personal AI Agent Claude Duet

Meta's Muse has already delivered a rather impressive report card. As of October 20, about 40 days after launch, it had reached about 10.2 million downloads in the United States and about 10.9 million cumulative downloads worldwide.

On October 21, Anthropic officially released Claude Duet and launched a research preview.

After the launches of personal AI agents such as Grok Bot, Muse, and Dots, a new face has arrived on the scene!

What makes Claude's personal assistant different

Claude Duet's core positioning is that it can not only answer questions and generate content, but also proactively identify work needs, plan tasks, and call on different software to complete actual operations. Its main target users are professionals and small and medium-sized business owners.

Anthropic said Claude Duet can connect to more than 6,000 tools and services through connectors and plugins in the Claude Marketplace.

But to experience Claude Duet, it is not yet a matter of downloading the app and getting straight to work. During the research preview stage, the first batch is open to Claude Pro and Max individual subscribers in the United States; free users are not included in the first batch for now.

Go to the official Claude Duet product page to check the waitlist entry. After gaining eligibility, you can access Duet via iOS, Android, or the web.

After entering, authorize your email, calendar, and other supported apps according to the interface prompts. It is recommended to connect your most commonly used tools first rather than handing over all your work accounts at once, and you can also assign tasks to it.

Claude Duet uses the Opus 5.5 model and can break a complex task into many small parts, use a variety of software and tools to complete them, and finally deliver a complete piece of work, for example, telling you which emails need replies, which meetings are about to start, and which tasks have not yet been completed.

What is different is that it is more focused on professional scenarios, from financial forecasting, sales reports, and data dashboards to presentation decks and web application development, all of which it aims to cover.

To achieve these things, one important capability is running in the cloud.

Claude Duet has its own cloud virtual machine and browser environment. That is to say, even if the user turns off the computer or exits the application, it can continue to complete the work tasks previously assigned to it.

But so-called "all-weather work" does not mean it can scan you 24 hours a day without interruption. Anthropic says Duet will check the connected accounts at intervals and will also adjust the resources used by the model accordingly.

Their daily work depends on many software tools: using email to communicate with clients, managing sales leads in a CRM, checking bills with financial software, and then using various office software to organize reports. You can directly ask it to draft and send documents for you.

This is a scenario closer to daily life and actual use. In this scenario, Anthropic does not need to get to know you from scratch.

It can inherit the memory, skills, and already connected apps the user has accumulated in Claude, and can continue to use the previous work environment, maintaining a single conversation thread to advance work.

Duet's background work consumes the existing Claude subscription quota. Unlike ordinary chatbots, in order to achieve the goal of 24-hour work, Claude Duet needs to constantly call the model, use applications, operate the browser, and continuously plan and revise steps in complex tasks. These are all computing costs.

Claude Duet is sharing quota with the Pro and Max plans. That is to say, it can do background work, but it cannot keep working endlessly. The more information you ask it to check and the more complex the things it does, the more quota it may consume.

Putting a lock on personal agents

In October 2026, the Associated Press reported on the controversy caused by Meta Muse. A user in Canada asked the AI to help him sell a computer keyboard, but the agent, when the user did not expect it, pushed the transaction forward with the buyer and told the other party the home address.

After that, a strange buyer came to the user's home after 10 p.m.

Claude Duet is mainly used in professional fields. In this case, the target users' time savings and results are relatively easy to measure, and they also have a clearer willingness to pay.

Therefore, for it, security is even more important. Anthropic naturally knows this and tries to put a lock on its operating process.

According to the official statement, Claude Duet can read materials, organize information, produce drafts, and make preparations on its own, but when performing sensitive actions such as sending messages, making payments, signing contracts, or deleting data, it still needs the user's consent.

For example, if Duet is asked to find clients with overdue payments, compile the amounts owed, and draft collection emails, it can prepare these things first; but when it actually sends emails to clients or executes payment operations, it still needs to go through the corresponding confirmation process.

Claude Duet has also added credential isolation, sensitive information detection, and prompt injection protection measures to prevent the agent from being influenced by malicious instructions when receiving external information or from exposing important materials such as credit card numbers and identity documents without cause.

However, the existence of security mechanisms does not mean the risks have been eliminated.

Anthropic also stated that Claude Duet during the research preview stage may still misunderstand email or chat context and misjudge the urgency of tasks, and its performance in personal life scenarios such as booking tickets is not yet good enough.

A Crowded Track

Anthropic is not the only company that wants to let artificial intelligence work on its own.

Meta founder and CEO Mark Zuckerberg once said, "Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about."

OpenAI's product and engineering lead Tibo could not resist joking, "One of you is called Dot, the other Duet, and your feature roadmaps are so similar—why don't you get to know each other first?"

The three products all highlight long-term memory, proactive operation, and cross-application collaboration, but the angles they have chosen are not the same.

Meta Muse is more suitable for ordinary consumers, while OpenAI Dots works around users' long-term goals, allowing artificial intelligence to track a project and take on long-term responsibilities.

Claude Duet clearly leans more toward professional workplace environments, where it will first seek its value in work such as finance, sales, customer service, and complex document creation.

Judging from the current situation, the three companies are competing for different user groups, but the goal they pursue is the same: to be the first thing users think of in their digital affairs.

Those who master users' long-term goals, daily work tasks, and service invocation entry points will occupy a more important position in future software distribution, consumption decisions, and interactions with users.

However, there is still a gap between the ideas of the big Silicon Valley companies and the situations of real users. As an ordinary user, I just want to know how to solve the problems I least want to face with the least money and time.

This article comes from the WeChat public account "APPSO", author: the one who discovers tomorrow's products, published by 36Kr with authorization.