Claude Duet Keeps Working on Your Tasks Even When Your Laptop Is Off

Anthropic has announced Claude Duet, an AI assistant that works on its own virtual computer and flags anything that needs attention. The first test version is rolling out through a waitlist to Pro and Max subscribers in the US. Sending messages and making payments require approval first, according to Anthropic’s announcement.

Claude Duet runs in the cloud, so your laptop or phone doesn’t need to stay on. Users give instructions by text or voice and get a notification when the work is done—or when the assistant needs a decision.

Anthropic says Duet can prepare documents, track down unpaid invoices and build dashboards. It uses the memory, skills and connected apps in a user’s existing Claude account. Conversations stay together in a single, ongoing message thread.

That builds on earlier features that let Claude turn business data into dashboards and animations. The key difference is when the assistant gets to work: business owners no longer need to reopen a conversation every time they want to move a task forward.

Duet can independently read information, draft content and prepare work. But before sending a message, making a payment, signing a document or deleting anything, it must show what it plans to do and wait for approval. Trading stocks and transferring money between accounts are prohibited, even with permission.

That distinction matters in practice. You can still edit a draft payment reminder. Once sent, however, the message reaches a customer immediately and could affect the relationship. For tasks like these, human oversight belongs at the finish line—not just at the start.

In its principles for trustworthy AI agents, Anthropic explains why getting that balance right is difficult. An agent that asks for approval at every turn quickly becomes useless. But a system that always decides for itself may misunderstand what the user wants.

An assistant that reads emails and websites also encounters information from strangers. Bad actors can hide instructions in that content to steer the AI away from its original task. This type of attack is called prompt injection. AI Wereld has previously covered the security risks of AI browsers.

Anthropic’s research into attacks on browser agents explains why the risk is greater: these agents don’t just process text. They can also fill out forms, click buttons and download files. A deceptive instruction can therefore trigger a real-world action.

The company stresses in its safety framework that multiple layers of defense are no guarantee. Connected tools, their permissions and the working environment all affect the risk. For organizations, the key question is what data and actions an assistant actually needs. Access to an entire inbox is a very different choice from access to a single, clearly defined project.

According to Claude Duet’s product page, Team and Enterprise subscribers are not yet included in the test. A separate app will be available for iOS and Android, and Duet can also be used on the web. Anthropic has not announced a launch date for the Netherlands.

During the test, Duet shares Pro and Max usage limits with the rest of Claude. Background tasks therefore use capacity that would otherwise be available for regular chats. Bigger jobs, such as building a financial model or presentation, demand more from the model.

Anthropic also warns that Duet may misread conversations or misjudge urgency. Users will need to decide whether the assistant genuinely takes work off their hands—or simply creates more notifications and corrections. In the end, that matters more than how many tasks it can start at once.