Anthropic’s Claude Duet Puts a Proactive Agent on Work Apps

Anthropic has started letting a limited number of people try Claude Duet, an AI assistant that can take action without waiting for someone to tell it what to do each time. It checks connected work apps, prepares a morning update and keeps working even after someone closes the app. Instead of waiting for every new instruction, Duet can point out things it thinks need attention and prepare work for the user to check.

For now, people must join a waiting list to try Duet. It is only available to subscribers in the United States who have Anthropic’s Pro or Max plans. They can use it on iPhones, Android phones and the web. Duet runs on Claude Opus 5.5 and uses the same usage limits as the person’s regular Claude account. People on Team or Enterprise plans cannot use it yet.

From answering questions to looking for work that needs doing

Anthropic says Duet is an AI assistant made for people who work in offices, other professionals and small-business owners. It can use connected email accounts, calendars and business tools to put together a daily update, write draft replies, arrange meetings and prepare bigger pieces of work, such as dashboards that show important information and forecasts that predict what might happen. The company says Duet can work with more than 6,000 connectors and plugins in its Marketplace, which help it connect to other tools.

The biggest change is that Duet can keep working on its own. It runs on a virtual computer in the cloud, with its own web browser. This means it can continue working even when someone closes their phone or laptop. According to Anthropic, Duet checks connected accounts every few hours rather than watching them all the time. The company says these background checks count toward the user’s plan limits. Duet can then send a notification when it thinks a task is finished or needs the user to make a decision.

These are things Anthropic says Duet can do. No independent performance test available to the public was found.

• What it can do: Prepare updates, drafts and files.
• Who can use it: People in the United States with a Pro or Max account who receive an invitation from the waiting list.
• Limits: Duet checks accounts from time to time, not constantly. It must ask permission before sending messages, making payments, signing documents or deleting things.

Users must approve important actions

Anthropic says Duet waits for the user to give permission before it sends a message, makes a payment, signs a document or deletes something. Some actions are not allowed at all, even if the user gives permission. These include buying or selling investments such as stocks and moving money between accounts. Users can also disconnect an app from Duet. Anthropic’s help centre says users cannot watch Duet’s browser as it works or take control of that browser themselves.

Anthropic says Duet has a separate system that checks for possible problems, protections against hidden instructions designed to trick the AI, and ways of handling login details that are meant to stop the AI model from seeing passwords and sign-in tokens. The company also says its security testing teams, known as red teams, and an outside company tested the system for weaknesses that attackers could use. Duet is also covered by a programme that rewards people for finding security bugs. Anthropic says it is still improving these protections and plans to share more details. However, the launch materials reviewed for this article did not include a public report from an independent testing company or the results of tests looking for security weaknesses.

These risks are not limited to Duet. An AI assistant that reads messages and websites might come across instructions written to trick it into doing something wrong. Giving it access to email, calendars or payment tools also means it can see private and sensitive information. The Associated Press has reported warnings from privacy experts that people may not be able to take back information after an AI assistant has already seen it. Anthropic’s own instructions warn that Duet might misunderstand a conversation or wrongly decide how urgent something is. The company also warns that no protection against instructions designed to trick AI works perfectly.

A small early test, not a full launch for workplaces

During this early testing period, people with Pro and Max plans can use Duet without paying extra. However, the work Duet does in the background counts toward the same session and weekly usage limits as regular Claude. This means bigger tasks can use up more of a customer’s available allowance. Anthropic’s help centre says it is letting people in gradually, in groups. Joining the waiting list does not mean someone will get access on a particular date.

Duet also has several limits right now. It is only available in the United States and cannot be used with Team or Enterprise plans. It does not yet work with Microsoft 365, and some connections to other apps have restrictions. Duet checks accounts at certain times rather than constantly. Users cannot see what its browser is doing live. It is also not as good at personal tasks, such as booking tickets. It can send an email after the user approves it, but it cannot make phone calls or send text messages.

For people and businesses in Kenya, Duet is not available yet because the early test is limited to the United States. This means its importance in Kenya is currently about what the technology might offer in the future, rather than something people can use today. Before using it for sensitive work, businesses will need to look out for availability in their region, rules about how data is handled, controls for workplace administrators and support for the tools they already use.

Anthropic joins other companies building AI assistants that keep working

Claude Duet is entering a market where major AI companies are moving beyond chatbots that simply answer questions. OpenAI’s Dots product also works on a computer in the cloud and connects to work apps. OpenAI says it is gradually making Dots available to people on Pro, Business Premium and Enterprise plans in places where the product is supported. Meta’s Muse is designed to be a personal AI assistant that works across Meta’s apps and is gradually becoming available in the United States. Each company describes its own safety and privacy protections, but these claims have not been measured against one shared, independent test.

Reuters and TechCrunch have described Duet as Anthropic’s move into a growing competition to build AI assistants that can do things between instructions from users. The most important question is not which company makes the biggest promises. It is whether people will trust these systems to keep accessing their accounts, and whether they can clearly understand what the systems read, what information they keep, when they take action and how to stop them. Duet’s waiting list, checks every few hours and rules requiring permission show that Anthropic is still testing how these systems should work safely. They do not mean the main problems have been solved.