Anthropic has launched a research preview for Claude Duet, an AI agent designed to perform tasks autonomously. The tool operates on a dedicated virtual machine with its own browser, allowing it to function independently of the user's local hardware. It integrates with platforms like Gmail, Slack, and QuickBooks to provide daily summaries and manage workflows such as scheduling meetings or building dashboards.
The agent utilizes the Claude Opus 5.5 model and retains memory and connected app data from the user's existing account. While designed to work proactively, the system requires user approval for sensitive actions, including sending messages, making payments, or signing documents. Anthropic has implemented security measures, such as a classifier to block the sharing of sensitive data like card numbers, and restricts the agent from trading securities or moving funds.
The release has prompted public commentary regarding AI autonomy, including a reaction from Binance founder Changpeng Zhao, who noted the shift in control dynamics. Anthropic acknowledges potential risks, citing a previous incident where a different model sent a false report to law enforcement. The company maintains that the agent may still misinterpret urgent tasks or context during operation.
Access to the preview is currently limited to Pro and Max subscribers in the United States via a waitlist, with no additional fees. The company has not announced a timeline for a wider rollout or availability for Team and Enterprise plans. This launch follows the recent release of the Haiku 5.5 model and precedes the company's anticipated November listing, where reports suggest a potential valuation near 1,5 trillion USD.
