"你问它答"的时代，可能正在结束
昨天，Anthropic 做了一件不太"Anthropic"的事
当地时间 10 月 21 日，Anthropic 开放了一款新产品的研究预览——Claude Duet。
这款产品最大的不同，用 Anthropic 自己的话说："Claude 不再等你打开对话框，而是主动来找你。"
它的定位是主动型智能体——每天早上给你发一条摘要，白天盯着你连接的所有应用，只有真正需要你处理的事才会打扰你，剩下的活儿它自己在后台干，干完再向你汇报。
首批面向美国 Pro 和 Max 订阅用户开放研究预览，iOS、Android 和 claude.ai 网页端都能用。
图片来源于：Claude Duet官网
三个核心模块，看懂 Duet 在干什么
① 每天一条简报 + 全天候主动提醒
Duet 会基于 Claude Opus 5.5，每天早上整理一份跨应用的简报。白天它会持续监控你连接的账户（注意是每几小时轮询一次，避免耗光配额），只在真的有事时才发消息。你可以自定义"想听什么、多久听一次"。
② 你下线，它继续干
Duet 有自己的虚拟机和浏览器，不依赖你的手机或电脑开机。发一条消息描述任务，它在后台自己决定用哪些应用、走哪些步骤，做完再通知你。复杂任务如搭建几十个工具的看板、财务预测、原型开发、Web 应用，都能接。
③ 关键动作，必须经过你
Duet 可以自己读、自己写、自己准备——但发送消息、付款、签文件、删除任何东西之前，会把要做什么完整展示给你，等你批准。证券交易、账户间转账这两类动作，即使你同意也直接禁止。
图片来源于：Claude Duet官网
一个小企业主的实测，值得看
Anthropic 给了几位早期用户的反馈，最有信息量的是两位——
Amin，南佛罗里达一家拥有约 110 名员工的烤肉店老板，用 Duet 做账务、客服和营销，逐店处理发票。他的原话：
"它把所有东西汇到一起，我不用再开 100 个标签页了。"
Jess，房产经纪人，Duet 主动发现她日历上一个带看排错了日期——她根本没问。她的评价："这正是我想不问就被发现的那类事。"
这两条反馈指向同一个判断：小团队真正需要的不是更聪明的模型，而是一个能主动替你盯盘的"数字副手"。
图片来源于：Claude Duet官网
接入生态：6000+ 连接器，覆盖企业刚需
Duet 可以调用 Claude Marketplace 里的 6000+ 连接器和插件，官方明确列举的场景包括：
- Gmail 和 Slack 未读整理 + 草拟回复
- 查 QuickBooks / Stripe 的未付发票 + 草拟催款
- 在 Gusto 里跑工资，自动填工时、奖金、报销
- 起草职位描述 → 你改完 → 一键发 LinkedIn
- 准备合同 → Docusign 发起签署
- 用 HubSpot 的联系人 + 交易历史起跟进邮件
- 把产品发布变成 Canva 上的社媒 campaign
- 建立和管理 Shopify 店铺
支付这块的细节很讲究：用 Stripe 的 Link 数字钱包付款，先用虚拟卡支付，真实卡号不外泄；后续还将支持用 1Password 存的登录信息登录网站。
安全这块，Anthropic 写得最厚
这是整个发布里最值得 B 端读者细读的一段。Anthropic 给 Duet 配了四层防线：
- 凭证隔离：沿用 Claude Managed Agents 的设计，Duet 永远不直接接触你的密码或登录令牌；
- 自动模式分类器：和 Claude Code 用同一套防护，拦截你"没让它做"的动作；
- 提示词注入防御：来自浏览器使用场景的研究成果；
- 敏感信息分类器：全新上线——信用卡号、政府身份证号直接拒绝发送；健康、个人财务这类敏感信息，只有你明确指示才放行。
此外，Anthropic 内部红队 + 外部公司渗透测试 + 公开漏洞赏金计划，三套机制同时跑。
这套安全设计，是目前所有消费级 AI 智能体里披露得最完整的一份。
几条必要的边界说明
研究预览阶段，这些限制值得知道：
- 仅限美国市场，仅 Pro 和 Max 订阅用户，需要加入 waitlist；
- 和 Claude 其他功能共享用量额度，后台检查也计入额度，复杂任务消耗更快；
- 核心功能围绕工作场景设计，个人事务（订票之类）表现一般；
- 还看不到它的浏览器操作过程；
- 目前是独立应用，和主 Claude 应用分开，未来会不会合并 Anthropic 没说；
- Anthropic 自己也承认：它仍可能误读线索、误判紧急程度。
对国内团队的启发
Duet 不是 Claude 主应用的升级版，而是 Anthropic 对"下一代工作界面"的一次押注——从被动响应转向主动协作，从单次对话转向长期任务。
这个方向，国内做企业级 AI 助手的产品也在走，只是路径不同。
以 CoCo这类面向企业的智能助手为例，它在设计逻辑上和 Duet 形成了几个清晰的对应点：
一个是消费级市场的激进尝试，一个是企业级市场的稳妥落地——方向一致，节奏不同。具体能力边界以 CoCo 产品内权限面板展示为准，建议选型前做一次实测。
过去三年，我们习惯了 AI 是一个"对话框"——你问，它答。
Claude Duet 试图打破这个范式：AI 变成一个有主动性、有判断力、懂分寸的"同事"——它知道什么该自己做、什么该问你、什么绝对不能碰。
这个范式能不能跑通，还要看研究预览期的用户反馈。
但它至少给出了一个信号：下一代 AI 产品的竞争焦点，已经不在"谁更聪明"，而在"谁更值得托付"。
Coco官网：coco.sinoaus.net
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The age of "You Ask Him" may be over.
Yesterday, Anthropic did something less "Anthropic."
On October 21st, local time, Anthropic opened a research preview of a new product, Claude Duet.
This product is the biggest difference, in the words of Anthropic himself: "Claude will not wait for you to open the dialogue box, but will come to you voluntarily."
Its location is the active AI agent — sending you a summary every morning, staring at all applications of your connection during the day, only if you really need to deal with it, and the rest of the work is done in the backstage itself and then I'll report to you.
The first open research previews for U.S. Pro and Max subscribers are available on iOS, Android and claude.ai.
Image by Claude Duet.
Three core modules, see what Duet is doing.
1 Daily Briefing + 24/7 Active Reminder
Duet, based on Claude Opus 5.5, organizes a cross-application briefing every morning. In the daytime, it monitors the accounts that you connect to (note that it rotates every few hours to avoid the use of quotas) and only sends messages when there is a real problem. You can define "what you want to hear, how often."
Two, you get off the line. It goes on.
Duet has its own virtual machine and browser, not dependent on your cell phone or computer. Send a message describing the mission, and it decides on its own applications and steps to be taken and will inform you when it is done. Complex tasks such as panels with dozens of tools, financial projections, prototype development, and Web applications are available.
Critical move, must pass you.
Duet can read it himself, write it, prepare it for itself -- but before sending messages, paying, signing documents, deleting anything, you will be presented with a complete picture of what you want to do, pending your approval. The two types of action, securities trading, inter-account transfers, are prohibited even if you agree.
Image by Claude Duet.
It's worth looking at.
Anthropic gave feedback to several early users, the most important of which was two --
Amin, a barbeque owner in South Florida with about 110 employees, used Duet for accountkeeping, customer service and marketing, processing invoices. His exact words:
"It transfers everything together and I don't have to open another 100 tabs."
Jess, the realtor, Duet, volunteered to find out that she had a wrong date on her calendar - she didn't ask. Her assessment: "That's exactly what I wanted to be found without asking."
These two feedbacks point to the same judgment: what small teams really need is not a smarter model, but a "digital second mate" who can keep an eye on you.
Image by Claude Duet.
Access to ecology: 6,000+ connectors, covering just-in-kind business
Duet can call 6000+ connectors and plugins in Claude Marketplace, which are explicitly listed by the authorities:
- Gmail and Slack Unread + Draft Response
- Check outstanding invoices for QuickBooks / Stripe + draft reminders
- Running wages in Gusto, filling in hours, bonuses, reimbursement
- Drafting job descriptions.
- Prepare the contract.
- Follow up mail with HubSpot contacts + trade history
- Turn product publish into a social media campaign on Canva
- Establishment and management of the Shopify shop
The details of the payment are very detailed: payments are made in the Link digital wallet of Stripe, first in a virtual card, and the real card number is not leaked; subsequent access to the website will be supported with the login information stored in 1 Password.
Safe, Anthropic.
It's the most valuable part of the entire publish that the readers read. Anthropic has four layers of defense for Duet:
- Segregation of documents: following Claude Managed Ages design, Duet never had direct contact with your password or login;
- Automatic mode classifier: use the same kind of protection as Claude Code to intercept your "not letting it do" moves;
- prompt Injecting Defense: Research results from the use of scenes in browsers;
- Sensitivity information sorter: completely new access — credit card number, government ID number, direct refusal to send; sensitive information, such as health, personal finances, is only released if you give clear instructions.
In addition, Anthropic's internal red team + external company penetration tests + open-failure reward scheme, three mechanisms running simultaneously.
The security design is the most complete of all current consumer-grade AI AI agent disclosures.
Several necessary boundary statements
The research preview phase, these limitations are worth knowing:
- Only the United States market, only Pro and Max subscribers, need to be added to the waterlist;
- Share quota with Claude for other functions, and back-office checks are also included in quota, which consumes more complex tasks;
- The core functions are designed around the workspace, and personal matters (such as booking) are generally performed;
- The browser operation is not yet visible;
- It's a stand-alone application, separate from the main Claude application, and will not merge Anthropic in the future;
- Anthropic himself admits that it may still misread the clues and misjudge the degree of urgency.
Inspired by national teams
Duet is not an upgraded version of Claude's main application, but a bet from Anthropic on the next generation of work interfaces -- moving from passive response to active collaboration, moving from a single conversation to a long-term mission.
In this direction, the domestic company-level AI assistant product is also going along, except for different paths.
In the case of firm-oriented smart assistants such as CoCo, the design logic and Duet form several clear points of correspondence:
One is a radical attempt at consumer-level markets, and the other is the sound landing of enterprise-level markets — one direction and one pace. The specific capacity boundary is based on the presentation of permission panel in the CoCo product and is recommended to be measured before selection.
For the past three years, we've been used to AI being a "dialogue" -- you ask, it answers.
Claude Duet is trying to break the paradigm: AI becomes an active, judgmental, well-informed "coworker" -- it knows what to do, what to ask you and what to never touch.
It's a model that runs, and it depends on user feedback from the research preview.
But it gave at least one signal: the next generation of AI's product's competitive focus is no longer "who's smarter," but "who's more worthy of trust."
Coco official network: coco.sinoaus.net
Auto-fetched and translated by AI Atlas. All rights belong to the original author.