Anthropic launches Claude Duet personal agent

Judo @ZhihaoJu77910Originally posted 2026.10.23

Open original on X

昨天，Anthropic 也加入了"全天候个人智能体"的牌桌。 它发布的研究预览版 Claude Duet，和你印象里的聊天机器人是两回事：它拥有独立的虚拟机和浏览器，你的设备关机了它也能在云端继续跑；每天整理跨应用的重要事项、主动提醒，用户离开后继续完成被委托的任务。 背景是这样的：OpenAI 的 Dots、Meta 的 Muse 已经抢跑数月，Anthropic 今天正式补齐个人 agent 战线。三家头部，至此在同一个赛道上撞车。 但三家赌的其实是同一件事的不同答案——个人智能体要从"被动对话工具"变成"主动工作的数字员工"。区别只在于，你愿不愿意把活委托出去。 Anthropic 这次把宝押在"可信委托"上： - 设备关机能持续运行（有独立虚拟机+浏览器） - 能调用 Claude Marketplace 6,000+ 个连接器（Gmail、Slack、QuickBooks、Stripe…） - 付款、发消息、签署文件、删内容前必须用户批准 - 线上支付走虚拟卡，不碰真实卡号，也不直接接触你的密码和登录 tokens 也就是说，它瞄准的不是"谁会聊天"，而是"谁能让用户放心地把财务、日程、发票交给 AI 代办"。从 OpenAI 到 Meta 到 Anthropic，这场战争的胜负手，大概率落在"安全感"和"可控性"上。 我的判断：个人智能体从单点产品进入平台竞争，三强格局初成。接下来看谁先解决"委托后的信任"——这比模型智商更难，也更值钱。 Yesterday, Anthropic joined the table for "always-on personal agents." Its research-preview app, Claude Duet, is a different beast from the chatbot you picture: it runs on its own virtual machine and browser, so it keeps working in the cloud even when your device is off. It compiles cross-app priorities every day, nudges you proactively, and completes delegated tasks after you step away. Here's the context: OpenAI's Dots and Meta's Muse have been sprinting for months. Today Anthropic finally fills in its personal-agent line. Three frontrunners, now colliding on the same track. But the three are betting on different answers to the same question — moving the personal agent from a "passive chat tool" to an "active digital employee." The real difference is whether you trust it with real work. Anthropic is betting on "trustworthy delegation": - Keeps running even when your device is off (dedicated VM + browser) - Calls on 6,000+ connectors in Claude Marketplace (Gmail, Slack, QuickBooks, Stripe…) - Requires your approval before payments, messages, signatures, or deletions - Pays via virtual cards that never expose your real number, and never touches your passwords or login tokens In short, it's not aiming at "who chats better," but "who users can safely hand over finances, calendars and invoices to." From OpenAI to Meta to Anthropic, this war will likely be decided on safety and control. My take: the personal agent has moved from single products to platform competition — a three-way landscape is forming. Next comes who solves "trust after delegation." That's harder than model IQ, and worth more.