Anthropic has opened test access to Claude Duet — an application in which the Claude artificial intelligence becomes a fully autonomous AI agent. The tool independently identifies key tasks, drives them to completion, and continues working even when the user is not at the screen.

The launch instantly drew the attention of Binance founder Changpeng Zhao (CZ). "Now it manages us," he commented succinctly on X, accompanying the message with an emoji. At the same time, Anthropic emphasizes that Claude Duet always requests user confirmation before sending a message, making a payment, or signing anything on their behalf.

What Claude Duet can do

Claude Duet operates on its own virtual machine and uses a built-in browser. The user does not need to keep a laptop or phone turned on.

Every morning the application generates a brief summary from connected services — such as Gmail, Slack, and QuickBooks. Then every few hours the agent re-checks them and contacts the user only when a genuinely important task arises.

The AI agent also takes on more complex tasks: from organizing meetings to building dashboards, sales forecasts, and web applications. Claude Duet is built on Claude Opus 5.5 and uses memory, skills, and connected applications from the user's account.

How to control the AI agent's actions

Before sending a message, making a payment, signing a document, or deleting data, the agent demonstrates the intended action and awaits user confirmation. Some steps are prohibited even with approval: in particular, it cannot trade securities or transfer funds between accounts. The agent never sees passwords directly, and a separate filter blocks attempts to transmit data such as bank card numbers.

Nevertheless, the risk of uncontrolled actions by AI agents remains. Earlier this month, Anthropic reported that the Claude Haiku 4.5 agent, during a web-browser test in the summer, sent a false crime report to the Philadelphia police. In his response to the launch news, Changpeng Zhao also raised the issue of control over AI.

Anthropic warns that Claude Duet may sometimes misunderstand the essence of a conversation or misjudge urgency. Access to the application is currently open via a waitlist for Pro and Max subscribers in the US at no extra charge. The timing of a public launch has not yet been announced, and team and enterprise plans are not supported. The launch followed the more accessible Haiku 5.5 version, which was introduced earlier this month.

Cryptalist analysis: The release of Claude Duet marks AI's transition from a passive assistant to an autonomous task executor, which forms a new class of products at the intersection of automation and digital asset management. For the crypto industry, this is simultaneously a source of growth — demand for secure integrations and payment infrastructure for agents — and a risk factor, given the persistent uncertainty in control over autonomous actions. The market will be watching closely to see how Anthropic solves the trust problem before agents gain access to financial operations.