九月二十二日

Box CEO Aaron Levie argues that personal agents capable of completing complex tasks end to end could become a major channel for commerce: people will start with routine chores, then entrust agents with more complex purchases and services as lower friction increases total spending. He sees opportunity both for agent providers and for new commerce, local-service, and B2B layers designed for agent interaction. Levie also predicts agents will use software far more heavily than people do, making familiar CRM, ERP, and data primitives more important rather than obsolete. In that environment, platforms that secure agent access, guard destructive actions, manage context, and orchestrate business logic become a strategic control layer.

Box CEO Aaron Levie 认为，能够端到端完成复杂任务的个人 agent 可能成为重要的商业交易入口：用户会先把日常琐事交给 agent，随后随着摩擦降低，再逐步委托更复杂的购买和服务，从而带动更高的总支出。这不仅给 agent 提供商带来机会，也会催生专门面向 agent 交互的电商、本地服务和 B2B 新层级。Levie 还预测，agent 使用软件的强度将远高于人类，因此 CRM、ERP 和数据平台等基础能力不会过时，反而会更重要。在这种环境中，能够保护 agent 访问、约束破坏性操作、管理上下文并编排业务逻辑的平台，将成为关键的控制层。

Y Combinator President and CEO Garry Tan says Capy has become his preferred new agentic coding tool because it can track multistep workflows and produce large pull requests faster than Codex or Claude Code on their own. His example shows it tackling an ambitious GBrain bug-fix wave through explicit task delineation, automatic parallelization, and a GitHub pull-request and CI workflow. The claim is a hands-on endorsement rather than a benchmark, but it points to orchestration across a large unit of work, not just single-task code generation, as the differentiator he values.

Y Combinator 总裁兼 CEO Garry Tan 表示，Capy 已成为他近期最喜欢的 agentic coding 工具，因为它能跟踪多步骤工作流，并比单独使用 Codex 或 Claude Code 更快地完成大型 pull request。他分享的案例中，Capy 通过清晰拆分任务、自动并行执行，以及 GitHub pull request 与 CI 工作流，处理了一轮规模很大的 GBrain bug 修复。这是基于实际使用的推荐，而非基准测试，但它清楚表明：他看重的差异化能力不是单任务代码生成，而是对大规模工作单元的整体编排。

AI tutorial creator Peter Yang warns that targeted display advertising faces a structural problem when agents browse websites and complete tasks without a human ever seeing the ads. He also shares two examples of agents handling financial follow-through: ChatGPT Finances detected a hotel charge after a points booking and helped secure reimbursement, while Yang used ChatGPT to cancel his parents' hotel booking, monitor replies, and continue following up until the money was returned. His broader behavioral claim is that the primary interface is shifting away from email and text toward persistent conversations with agents.

AI 教程创作者 Peter Yang 警告说，当 agent 代替用户浏览网站并完成任务，而人类根本看不到广告时，依赖定向展示广告的商业模式会面临结构性问题。他还分享了两个 agent 负责财务跟进的例子：ChatGPT Finances 发现了一笔使用积分预订酒店后仍被收取的费用，并协助追回款项；Yang 则让 ChatGPT 为父母取消酒店预订、持续监控回复并跟进，直到退款到账。他更广泛的判断是，用户的主要交互界面正在从 email 和短信转向与 agent 的持续对话。

Vercel CEO Guillermo Rauch says Vercel AI Gateway now exposes Jev over HTTP in addition to its type-safe AI SDK API, widening the ways developers can call the system. In a separate hands-on report, he says Grok 4.7 quickly solved a difficult reverse-engineering problem involving a running binary. The latter is an individual usage report, not a controlled comparison.

Vercel CEO Guillermo Rauch 表示，Vercel AI Gateway 除了提供类型安全的 AI SDK API，现在也支持通过 HTTP 使用 Jev，让开发者有更多调用方式。在另一条实际使用反馈中，他称 Grok 4.7 很快解决了一个涉及运行中 binary 的高难度逆向工程问题。后者是个人使用反馈，并非受控对比评测。

FPV Ventures partner Nikunj Kothari argues that products boasting about maximizing token use often deliver poor experiences because good products depend on curation and ongoing gardening, not on handing agents every possible input and expecting them to sort it out. He contrasts that design discipline with the low setup required by consumer browser agents such as Instinct and Muse, while praising Codex on Mac for computer-use workflows. His practical request for the next mass-market agent launch is straightforward: clearly show the work it can do and make first use easy.

FPV Ventures 合伙人 Nikunj Kothari 认为，那些强调把 token 用量拉满的产品往往体验不佳，因为好产品依赖持续的筛选与维护，而不是把所有可能的输入都丢给 agent，再期待它自行理清。他将这种产品纪律与 Instinct、Muse 等消费级浏览器 agent 的低配置门槛放在一起讨论，同时肯定 Codex on Mac 在 computer-use 工作流中的能力。他对下一款大众 agent 产品的要求很直接：清楚展示它能完成什么工作，并让首次上手足够简单。

OpenClaw builder Peter Steinberger says reports that Meta itself uses OpenClaw are inaccurate: Meta built its own agent, although it was inspired by OpenClaw. The distinction matters because inspiration and direct adoption imply different evidence about a system's architecture and deployment.

OpenClaw 开发者 Peter Steinberger 澄清称，关于 Meta 本身使用 OpenClaw 的说法并不准确：Meta 构建了自己的 agent，只是受到了 OpenClaw 的启发。这个区别很重要，因为“受到启发”和“直接采用”代表的是完全不同的系统架构与部署证据。

#### Claude in Chrome is generally available

Claude Blog: “Claude in Chrome is generally available,” published August 26, announces availability on every paid Claude plan and autonomous browser actions checked against the user's request. It works with existing logins on internal dashboards, legacy systems, and vendor portals. The defenses combine model training, probes that screen tool results for prompt injection, and a classifier that checks actions before execution; users can turn automatic approval off.

Claude Blog 的《Claude in Chrome is generally available》发表于 8 月 26 日，宣布向所有 Claude 付费方案开放，并支持在核对用户请求后自主执行浏览器操作。它可以借助现有登录状态访问内部仪表盘、遗留系统和供应商门户。防护由模型训练、扫描工具返回结果中 prompt injection 的探测器，以及执行前检查操作的分类器组成；用户也可以关闭自动批准。

On its stronger red-team evaluation, attacks reaching models without added safeguards succeeded against Opus 4.5 in 17.6% of cases and Opus 5 in 3.8%. With probes and the safety classifier, no attacks succeeded against Sonnet 5 or Opus 5, while Fable 5 recorded 0.3%, all described as low severity. These are evaluation results, not a guarantee: the article cautions, “Prompt injection remains a moving target.” Enterprise admins can restrict access to approved domains. Local files and other applications still require the desktop app; other Chromium browsers and mobile are not supported yet.

在更强的红队评测中，进入未加额外防护模型的攻击，对 Opus 4.5 和 Opus 5 的成功率分别为 17.6% 和 3.8%。加入探测器和安全分类器后，针对 Sonnet 5 和 Opus 5 的攻击均未成功，Fable 5 的成功率为 0.3%，文中称成功案例均为低严重度。这些是评测结果，并非安全保证；文章提醒：“prompt injection 攻击仍在不断变化。”Enterprise 管理员可以将访问范围限制在获批域名内。本地文件和其他应用仍需通过桌面应用操作，目前尚不支持其他 Chromium 浏览器和移动端。