Anthropic introduced Claude Duet — an agent that autonomously tracks work tasks and executes them

Anthropic introduced Claude Duet — a new AI agent that can autonomously track a user’s work tasks, offer assistance, and perform tasks in connected apps.

This was announced by the company.

Unlike a regular chatbot, it doesn’t require a separate prompt for every action: the assistant can proactively notify about an important event or an issue it has noticed.

What Claude Duet can do

Claude Duet can prepare a daily brief of the most important tasks based on information from connected apps. For example, it can check email and calendar, find messages that require replies, or remind you about tasks that deserve attention.

Throughout the day the assistant continues to monitor work apps. According to Anthropic, it checks them roughly every few hours and sends notifications when it notices something important. Users can also set which events they want to be notified about and how often.

Besides this, Claude Duet can perform tasks of varying complexity: from organizing meetings and drafting emails to creating financial forecasts, analytics dashboards, and web applications.

The assistant supports over 6,000 integrations and plugins from Claude Marketplace. In particular, it can:

• find missed messages in Gmail and Slack and draft responses;
• prepare documents for signing via DocuSign;
• create job postings and, after approval, publish them on LinkedIn;
• prepare marketing materials in Canva and manage a store on Shopify.

The service has a dedicated app for iOS and Android. On desktop it can be used via claude.ai. 

Currently Claude Duet is available only in a limited test launch for Pro and Max users in the US. To get access you need to sign up for the waitlist.

Can Claude Duet act without permission

Claude Duet is based on the Claude Opus 5.5 model. The assistant uses the memory, settings, and connected apps from the user’s Claude account, so the user doesn’t need to set them up again.

You can interact with it via text or voice. Claude Duet has its own virtual machine and browser, which allows it to continue working in the background even if the user’s phone or computer are turned off. After completing a task it sends a notification with the result.

At the same time Anthropic has put limits on actions that could have consequences for the user. Before sending messages, paying for purchases, signing documents, or deleting information, Claude Duet must obtain permission. The assistant shows exactly what it intends to do and waits for confirmation.

Some operations remain unavailable even with the user’s consent. In particular, Claude Duet cannot trade securities or transfer money between accounts.

The company also claims additional protection for sensitive information. According to it, the assistant does not get direct access to passwords and login tokens. A separate system inspects messages to prevent the transmission of bank card numbers and identity documents.

Other sensitive information, including health or financial data, the agent can only send by direct instruction from the user.

Which other companies are developing similar AI agents

Similar solutions already exist at other tech companies. In particular, OpenAI in September 2026 introduced Dots — AI agents that can work around the clock on user tasks. They have their own cloud computer, use connected apps, remember the work context, and independently report results or decisions that require human attention.

Microsoft is developing Copilot Cowork, which can perform multi-step tasks in Microsoft 365. For example, it can work with emails in Outlook, organize meetings, prepare documents and presentations, and analyze corporate data. The agent continues to perform tasks in the background and coordinates important actions with the user.

Unlike agents that primarily execute assigned tasks, Claude Duet and OpenAI Dots also emphasize proactivity: they can notice important events in connected services themselves and reach out to the user without an explicit prompt.

Also read: Spotify removed over 50 million AI-generated tracks from recommendations: how the service labels such content