AI Builders Digest — 2026-10-23

X / TWITTER

Anthropic: Claude Duet moves from chat to proactive work

Anthropic introduced Claude Duet, an early research preview of an always-on personal agent powered by Claude Opus 5.5. It can judge what needs the user, what can wait, and when to ask first, while carrying over Claude memory, skills, and connected apps. The company emphasizes that passwords stay hidden, a classifier guards sensitive information, and consequential decisions remain with the user; access is currently waitlist-only for US Pro and Max subscribers.

Anthropic 推出 Claude Duet：一款由 Claude Opus 5.5 驱动的常驻型个人 agent，目前处于早期研究预览。它会判断哪些事项需要用户介入、哪些可以等待、何时应先征求确认，并继承 Claude 账户中的 memory、skills 与 connected apps。安全设计包括不直接读取密码、用分类器防止敏感信息泄露，以及把重要决策留给用户；目前仅向美国 Pro 和 Max 用户开放候补名单。

Sources: product and availability · security design · Opus 5.5 and context

Every CEO Dan Shipper: Duet feels like the first “pro” personal agent

After two weeks of testing, Every CEO Dan Shipper says Claude Duet shifted his usage from mostly ChatGPT to roughly a 50/50 split. He praises its separate app, work/life channels, persistent todo list, and document creation, but says it is not yet a chief of staff: it cannot direct computer-based Claude sessions, its email drafts trail Dot, and voice is transcription-only. Shipper also argues every company needs a “frontier team” to test fast-moving AI capabilities without distracting the rest of the organization.

Every CEO Dan Shipper 在测试两周后表示，Claude Duet 已让他的使用习惯从以 ChatGPT 为主变成两者约各占一半。他认可独立 app、工作/生活频道、持续 todo list 和文档创作能力，但也指出它还不是“幕僚长”：无法调度电脑上的 Claude sessions，邮件草稿不如 Dot，语音也只有转写。他同时主张，每家公司都需要一支 “frontier team”，专门探索快速变化的 AI 能力，而不打断主业务团队。

Sources: Claude Duet field test · why companies need a frontier team

Box CEO Aaron Levie: AI creates a new cyber regime

Box CEO Aaron Levie predicts cybersecurity will become one of the largest enterprise investment areas. Agents will scan every log and line of code, discover and patch vulnerabilities, and also accelerate attacks, potentially driving 100 times more token usage than expected. His key labor point is counterintuitive: autonomy will create more demand for people who direct agents, review findings, patch systems, and respond to incidents.

Box CEO Aaron Levie 预测，cybersecurity 将成为企业最大的投资领域之一。Agent 会检查每一条日志和每一行代码、发现并修补漏洞，同时也会加速攻击，token 消耗量可能达到当前预期的 100 倍。他对就业的判断颇具反直觉性：自主化不会消灭人才需求，反而会增加对 agent 调度、结果复核、系统修补和事件响应人员的需求。

Sources: cybersecurity jobs · the new cyber regime

Vercel CEO Guillermo Rauch: security response becomes an always-on product feature

Vercel CEO Guillermo Rauch says the company deployed firewall-level mitigations for a Node.js remote-code-execution vulnerability before the official patch, automatically protecting all deployments while runtime upgrades rolled out. He also highlighted pre-coordinated WAF defenses against “php2shell” and urged teams to practice rapid patch-and-redeploy workflows. Separately, he recommends DeepSeek v4.1 Flash as a low-cost, zero-data-retention model for subagents executing plans created by stronger models.

Vercel CEO Guillermo Rauch 表示，公司在 Node.js 远程代码执行漏洞的官方补丁发布前，就已在 firewall 层部署缓解措施，并在滚动升级 runtime 的同时自动保护所有 deployments。他还提到行业针对 “php2shell” 提前协作部署 WAF 防护，并建议团队练熟快速 patch 与 redeploy。另一方面，他推荐用低成本、zero-data-retention 的 DeepSeek v4.1 Flash 作为 subagent，执行更强模型制定的计划。

Sources: Node.js RCE mitigation · php2shell response · DeepSeek for subagents

OpenClaw and OpenAI builder Peter Steinberger: AI found a high-impact PHP flaw

Peter Steinberger reports that the PHP vulnerability expected to cause major disruption was discovered with GPT-5.6 Sol. The signal is less about one model benchmark and more about AI-assisted security research crossing into consequential real-world vulnerability discovery.

OpenClaw 与 OpenAI builder Peter Steinberger 表示，预计会造成重大影响的 PHP 漏洞是借助 GPT-5.6 Sol 发现的。真正值得关注的不是一次模型跑分，而是 AI 辅助安全研究已经进入能够发现现实世界重大漏洞的阶段。

Source: original post

FirstMark VC Matt Turck: capital is concentrating around AI-native enterprise workflows

FirstMark VC Matt Turck highlighted two large Series E rounds in one week: $250 million for Pigment and $550 million for Legora. He frames Pigment as infrastructure for continuous financial planning in a world shaped by tokens, APIs, and agents, signaling investor conviction that AI-native enterprise workflows can support category-scale companies.

FirstMark VC Matt Turck 强调同一周出现两笔大型 Series E：Pigment 融资 2.5 亿美元，Legora 融资 5.5 亿美元。他把 Pigment 定位为 token、API 和 agent 时代的持续财务规划基础设施，反映出投资者相信 AI-native 企业工作流能够孕育品类级公司。

Sources: two Series E rounds · Pigment thesis

SPC General Partner Aditya Agarwal: AI may democratize the joy of making

SPC General Partner and Bevel Health co-founder Aditya Agarwal observed that his son spent about three hours watching Star Wars but around 20 hours building games inspired by it, then barely played them. His broader hypothesis is that AI can make creation itself as accessible and enjoyable as consumption, extending the intrinsic pleasure programmers get from building to a much wider population.

SPC General Partner、Bevel Health 联合创始人 Aditya Agarwal 观察到，儿子看《Star Wars》约 3 小时，却花约 20 小时制作相关游戏，而且做完后并没有玩多久。他提出的更大假设是：AI 会让“创造”像“消费”一样易得且有趣，把程序员在构建过程中体验到的内在乐趣扩展给更多人。

Source: original post

Builder Zara Zhang: the personal website is becoming the new resume

Builder Zara Zhang argues that a personal website is now the new resume, while demonstrating one-prompt creation and a photo-to-3D-popup-card workflow. The common thread is that AI is lowering the cost of expressive proof-of-work: showing what you can make increasingly matters more than listing credentials.

Builder Zara Zhang 认为个人网站正在成为新的简历，同时展示了单 prompt 创作以及把照片变成 3D 弹出卡片的工作流。共同信号是，AI 正在降低表达型作品的生产成本：与其罗列资历，直接展示你能做出什么正变得更重要。

Sources: personal website thesis · one-prompt creation · photo to 3D card

PODCASTS

AI & I by Every: How a Frontier Team Can Help Your Company Keep Up With AI

The Takeaway: A small frontier team lets a company explore fast-changing AI capabilities without forcing everyone else to chase every new model and tool.

Every’s approach pairs “pirates,” who run many fast and messy experiments, with “architects,” who turn the few useful discoveries into maintainable systems. A wrangler then organizes the learning, connects experiments to themes, and translates them into products, content, or consulting. The team also uses personal benchmarks to replace vague model impressions with repeatable evaluations, and tracks token spend by person, model, and work category to ask whether experiments are producing enough value. Dan Shipper captures the operating model clearly: “The frontier team's job is to keep up as models change. Everyone else's job is to use what they learn.”

核心结论： 一支小型 frontier team 能让公司持续探索快速变化的 AI 能力，同时避免让其他员工追逐每一个新模型和工具。

Every 的方法把两类人配对：“pirates” 快速、大量、略显混乱地做实验，“architects” 则把少数有效发现转化为可维护的系统。再由 “wrangler” 整理学习成果、提炼主题，并将其转化为产品、内容或咨询能力。团队还用个人 benchmarks 把模糊的模型体验变成可重复评估，并按人员、模型和工作类别追踪 token 成本，判断实验投入是否产生足够价值。Dan Shipper 对这一机制的概括很直接：“frontier team 的工作是跟上模型变化，其他人的工作是使用他们学到的东西。”

Source: How a Frontier Team Can Help Your Company Keep Up With AI

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders