Anthropic 向美国 Pro 和 Max 订阅用户开放 Claude Duet 智能体预览
Anthropic Opens Claude Duet AI Agent Preview for US Pro and Max Subscribers
Anthropic 面向美国 Pro 和 Max 订阅用户开放 Claude Duet 智能体研究预览,提供 iOS、Android 独立应用及 claude.ai 入口,基于 Claude Opus 5.5 模型。该智能体可安排会议、起草邮件、制作仪表盘、生成财务与销售预测、开发原型和网页应用,并通过 Claude Marketplace 连接 Gmail、Slack、QuickBooks、HubSpot 等 6000 多个工具。它运行在自带浏览器的虚拟机中,每隔几小时检查关联账户,发送消息、付款、签署文件或删除信息前需用户批准,证券交易和账户间转账即使获批也不允许。
推荐理由
精选AI 评分 81
原文列出 Duet 可执行的任务范围与需人工审批的边界,读者可据此判断这类常驻智能体的可用范围。
主题
正文
“We built Claude Duet for professionals and small business owners with more on their plates than they can track,” Anthropic said in its announcement.
Available as a standalone app on iOS and Android, as well as through claude.ai, Claude Duet is built around the company’s Claude Opus 5.5 model. According to Anthropic, it prepares a morning briefing based on connected applications and sends notifications when it identifies something requiring attention.
The agent can schedule meetings, draft emails, build dashboards, prepare financial and sales forecasts, develop prototypes and create web applications. It can also connect with more than 6,000 tools through the Claude Marketplace, including Gmail, Slack, QuickBooks and HubSpot, the company said.
Claude Duet uses a virtual machine with its own browser so it can keep working when users step away from their devices. It checks connected accounts every few hours and sends updates when tasks are complete or require review. The agent also carries over existing memory, skills and connected apps from a user's Claude account.
Anthropic said Claude Duet prepares work independently but requires approval before sending messages, making payments, signing documents or deleting information. Securities trading and transfers between accounts remain prohibited, even with user approval.
The company said Duet does not directly access users’ passwords or sign-in tokens. It also uses safeguards designed to detect prompt injection, in which malicious instructions embedded in content attempt to manipulate an AI system.
The AI company said a dedicated classifier and safety prompt are designed to block requests to share highly sensitive information, including credit card numbers and government identification.
Users can share health and personal financial information when they explicitly instruct the agent to send it. These are the company’s stated safeguards, and Anthropic said it continues to develop them.
Anthropic said early testers have used Claude Duet for accounting, customer support and marketing. Amin Ebrahimi, owner of Shahs of Kabob, a South Florida restaurant business with about 110 employees, used the agent to work through invoices across stores.
“It brings it all together,” Ebrahimi said. “I don’t need to open up 100 tabs.”
The research preview shares usage limits with the rest of Claude, with background checks and complex tasks counting towards those limits.
Anthropic acknowledged that Duet may misread conversations or misjudge urgency, while its capabilities for personal errands such as booking tickets remain less developed.