Anthropic opened the research preview of Claude Duet on October 22, expanding Claude from a conversational assistant into an AI agent that can identify pending work, move tasks forward and continue operating after the user leaves their device. Binance founder Changpeng Zhao, known as CZ, responded on X: “Now it’s our turn to be controlled by it,” adding a laughing-and-crying emoji.
For financial and corporate users, Claude Duet is designed to do more than generate text. It connects to work applications and can carry parts of a process through on the user’s behalf. Anthropic said, however, that the agent will still seek confirmation before sending messages, making payments or signing documents for the user.
Claude Duet keeps working after users leave
Claude Duet runs in a standalone virtual machine and browser, meaning the user’s computer or phone does not need to remain switched on. Anthropic said the system sends a short briefing each morning based on connected applications, then checks those applications every few hours and only alerts the user when it finds something requiring attention.
In practice, users would not need to keep opening Gmail, Slack or QuickBooks to check for updates. The agent can first consolidate information and return decisions requiring human input to the user. For business teams, this could reduce repetitive cross-application work, while external communications and money movements remain subject to manual approval.
Tasks range from scheduling to forecasting
Anthropic says Claude Duet can schedule meetings, build dashboards, generate sales forecasts and create web applications. Unlike a chat tool that responds to one instruction at a time, the agent can break a broader objective into steps and continue working in the background.
The agent uses Claude Opus 5.5 and can carry over the memories, skills and connected applications associated with a user’s Claude account. That continuity allows it to retain working context across different tasks. It also raises a practical question for users: whether accumulated memories and configurations could make it harder to switch to another AI service over time.
User approval remains a key safeguard
CZ’s comment highlights the control questions raised by more proactive AI systems. When software can maintain access to email, collaboration platforms and financial tools, users need to define which tasks may run automatically and which actions require confirmation each time. Claude Duet remains in research preview, so its practical capabilities, reliability and permission boundaries across different application connections are still being tested through use.
What is clear at this stage is that Claude Duet has not removed human authorization entirely. Actions with direct external consequences, including sending messages, making payments and signing documents, still require user approval. Routine information gathering, task follow-up and some content-production work can continue while the user is offline.