Anthropic Launches Claude Duet Proactive AI Agent

Claude Duet Proactive Agent

Most chatbots wait for people to ask questions. Claude Duet starts conversations on its own. On October 21, Anthropic opened a research preview of Claude Duet. It is a separate app that checks a user's connected work apps, sends a summary each morning and keeps working on tasks on a cloud computer, even when the user's phone is turned off.

Key Takeaways

• Claude Duet is a new Anthropic app built around Claude Opus 5.5. It sends a daily summary, checks connected apps every few hours and carries out tasks using its own cloud computer.
• It asks for permission before sending a message, making a payment, signing a document or deleting anything. It will not trade securities or move money between accounts.
• The research preview is available by waitlist to Pro and Max subscribers in the US. People can use it on iOS, Android and claude.ai. It uses the limits included in their existing plans.

How a Day With Duet Works

Every morning, Duet sends a short summary of the most important things in a user's connected apps. During the day, it checks those accounts every few hours and sends a message only when something needs attention. Anthropic says checking all the time would use up a user's plan limits too quickly. Users can also tell Duet what they want updates about and how often they want to receive them.

Everything happens in one ongoing conversation, using text or voice. Duet carries over the memory, skills and connected apps from the user's Claude account, so there is not much to set up.

Users can ask it to handle small tasks, such as scheduling meetings, or bigger jobs, such as making dashboards, preparing financial or sales forecasts, creating early versions of products and building web apps. Duet works on a virtual computer with its own browser, so the user's laptop or phone does not need to stay on. It sends a notification when a task is finished or needs the user's review.

It can use more than 6,000 connectors and plugins to work with other services. For example, it can find Gmail and Slack messages that were missed, follow up on unpaid invoices in QuickBooks or Stripe, prepare payroll in Gusto, send agreements through Docusign, write follow-up messages using HubSpot and manage a Shopify store. It can also buy items for a business through Stripe's Link wallet using a virtual card, or pay at checkout with Shop Pay.

Anthropic designed Duet for professionals and small business owners. Amin Ebrahimi, who runs Shahs of Kabob in South Florida with about 110 employees, tested it on accounting, customer support and marketing tasks. “I don’t need to open up 100 tabs,” he said.

Where Duet Draws the Line

Duet can read information, write drafts and prepare work on its own. But before it sends a message, makes a payment, signs a document or deletes anything, it shows the exact action it plans to take and waits for approval. Some actions are not allowed even if the user approves them. These include trading securities and moving money between accounts. Users can disconnect any app at any time.

The agent never directly sees passwords or sign-in tokens. Its actions go through the same automatic mode classifier used to protect Claude Code. Its defenses against hidden instructions are based on Anthropic's research into AI agents that use web browsers. Red teams and an outside company tested the system for security weaknesses. Duet is also covered by Anthropic's public bug bounty program, which rewards people who find and report security problems.

A new classifier and a safety instruction check every message Duet is about to send. They block attempts to share highly sensitive information, such as card numbers and government ID numbers. Health or personal finance details are sent only when the user clearly tells Duet to send them.

Known Limits

Anthropic says Duet is an early version. It may misunderstand a conversation thread or make a poor judgment about what is urgent. It does not handle personal tasks, such as booking tickets, as well as work-related tasks. Users also cannot watch its browser while it works.

Duet runs in its own app, separate from the main Claude app. Its background checks count toward the usage limits of the user's plan. Large tasks, such as building a financial model or a slide deck, use more of those limits.

An Industry Turning to Agents That Act

Duet arrives weeks after two competitors made similar moves. OpenAI's Dots and Meta's Muse agents already offer assistants that stay active and contact users before being asked. Meta has also paired its agent with new models such as Muse Wonder. Hardware makers are pursuing the same idea, as shown by Amazon's Bee Frontier and Bee Eclipse listening devices. The competition is no longer just about how well chatbots answer questions. It is increasingly about how much everyday work an assistant can handle for people.

AI agents that send emails and spend money can cause bigger problems when they make mistakes. Anthropic says Duet follows its framework for trustworthy agents, but the industry as a whole is still learning how to make these systems safe. OpenAI recently rated two new cases of AI agents behaving incorrectly as critical. Google DeepMind's Demis Hassabis has also described recent attacks on AI agents as early warning signs. Asking users to approve actions, as Duet does, is one way to reduce risks. But these checks help only if users read them before tapping yes.

Written by Vytautas Valinskas