AI agents are becoming Internet-scale infrastructure load, not just app features. Cloudflare CEO Matthew Prince says bot traffic has already passed human traffic, forcing a rethink of capacity, security, and who pays for the web. AI agents 正在变成 Internet-scale 的基础设施负载，而不只是应用功能。Cloudflare CEO Matthew Prince 说 bot traffic 已经超过 human traffic，这会迫使行业重新思考容量、安全，以及谁来为 Web 买单。
The ad-funded Internet looks structurally exposed in an agent-heavy world. If agents inspect thousands of pages but do not click ads, publishers, platforms, and infrastructure companies need new units of value beyond impressions and clicks. 广告资助的 Internet 在 agent-heavy 世界里显得结构性脆弱。如果 agents 会检查成千上万个页面但不点击广告，publisher、platform 和 infrastructure company 都需要在 impression 和 click 之外找到新的价值单位。
Apple's Foundation Models framework is turning local intelligence into a routing layer. Claude's new Swift package lets apps use on-device models for fast typed tasks, then hand off harder reasoning, coding, web search, and analysis to Claude. Apple 的 Foundation Models framework 正在把本地智能变成 routing layer。Claude 的新 Swift package 让 app 可以用 on-device model 处理快速 typed tasks，再把更复杂的 reasoning、coding、web search 和 analysis 交给 Claude。
Product teams are still searching for the right division of labor between AI and humans. Linear's Nan Yu argues that using the product remains the best bug-finding method, while code review should focus more on architecture, APIs, and technical debt. 产品团队仍在寻找 AI 与人的合理分工。Linear 的 Nan Yu 认为，用产品、主动把它弄坏，仍然是最好的 bug-finding 方法；code review 更应该聚焦 architecture、API 和 technical debt。
AI developer tools are moving from raw capability toward judgment and operational fit. Posts from Anthropic and OpenAI builders highlight interest in better model judgment, Codex gaps, Claude Code workflows, and clearer usage limits for power users. AI developer tools 正在从原始能力走向 judgment 和 operational fit。Anthropic 与 OpenAI builders 的帖子都在指向更好的模型判断、Codex 的短板、Claude Code workflow，以及 power users 对用量窗口的可见性需求。
OpenAI's Thibault Sottiaux asked builders what still feels surprisingly weak in Codex and should have been solved already. The useful signal is not a launch, but the product posture: Codex is mature enough that the team is now probing for long-standing workflow failures users still tolerate.
OpenAI 的 Thibault Sottiaux 问 builders：Codex 还有哪些 surprisingly weak 的地方，本来早该做好。这里的有用信号不是发布新功能，而是产品姿态：Codex 已经成熟到团队开始主动寻找用户长期忍受的 workflow failure。
https://x.com/thsottiaux/status/2073551549494596079
Linear head of product Nan Yu argued that the best way to catch bugs is still to use the product and actively try to break it. He framed code review as a better tool for architecture and API design than for discovering most bugs, because the real value is controlling technical debt rather than pretending static inspection catches everything.
Linear head of product Nan Yu 认为，抓 bug 的最好办法仍然是使用产品，并主动尝试把它弄坏。他把 code review 定位成更适合检查 architecture 和 API design 的工具，而不是发现大多数 bug 的工具，因为真正的价值是控制 technical debt，而不是假装静态读代码能抓住所有问题。
https://x.com/thenanyu/status/2073410299680428445
Anthropic's Cat Wu pointed to Claude Fable 5's improved judgment by noting that it chose propensity score matching for retention analysis without being explicitly prompted. The point is practical: better AI work is not only about generating more output, but about choosing appropriate methods across analytics, writing, and debugging.
Anthropic 的 Cat Wu 用 Claude Fable 5 在 retention analysis 中主动选择 propensity score matching 作为例子，说明模型 judgment 在变好。重点很实际：更好的 AI work 不只是生成更多内容，而是在 analytics、writing、debugging 等任务里选择更合适的方法。
https://x.com/_catwu/status/2073439890482794966
Vercel CEO Guillermo Rauch shared a visualization of token spend across lifetime Vercel AI Gateway usage, which aggregates trillions of tokens from millions of developers each month. His read: Anthropic remains dominant in the data, but open-weight AI is rising and model-lab share is visibly shifting over time.
Vercel CEO Guillermo Rauch 分享了 Vercel AI Gateway lifetime usage 中 token spend 的可视化。这个 gateway 每月聚合来自数百万开发者的 trillions of tokens。他的观察是：Anthropic 在数据中仍占优势，但 open-weight AI 正在上升，不同 model lab 的份额变化已经清晰可见。
https://x.com/rauchg/status/2073563586270781674
Peter Steinberger previewed an OpenClaw improvement that shows exactly when usage resets expire, aimed at users trying to manage their AI usage windows more precisely. It is a small product detail, but it reflects a broader pattern: heavy agent users increasingly need operational dashboards, not just chat boxes.
Peter Steinberger 预告了 OpenClaw 的一个改进：显示 usage reset 何时过期，方便用户更精确地管理 AI usage windows。这是一个小产品细节，但反映了更大的趋势：重度 agent 用户需要 operational dashboard，而不只是 chat box。
https://x.com/steipete/status/2073482942513565713
Every CEO Dan Shipper joked about "Codex in ChatGPT," a lightweight post but still a signal of how coding agents are being normalized inside general AI surfaces rather than living only in separate developer tools.
Every CEO Dan Shipper 开玩笑说 "Codex in ChatGPT"。这条很轻，但仍然说明 coding agents 正在被纳入通用 AI surface，而不是只存在于独立 developer tools 里。
https://x.com/danshipper/status/2073586548545638459
The MAD Podcast with Matt Turck: Cloudflare CEO: The Internet's Business Model Is Dead
The Takeaway: Cloudflare CEO Matthew Prince thinks AI agents have pushed the Internet past a historical boundary: machines now generate more traffic than humans, and the business model that funded the web is not built for that.
The Takeaway: Cloudflare CEO Matthew Prince 认为，AI agents 已经让 Internet 跨过一个历史边界：机器产生的流量超过了人类，而支撑 Web 多年的商业模式并不是为此设计的。
Prince says Cloudflare expected automated traffic to overtake human traffic around 2027, then watched the timeline collapse into the first half of 2026. The driver is not old-school scraping alone. It is agentic behavior: a human shopping for a camera might visit five sites, while an agent might visit 5,000 to complete the same job. That means the web could face radically higher load even when the number of human users barely changes.
Prince 说，Cloudflare 原本预计 automated traffic 会在 2027 年左右超过 human traffic，但实际时间线压缩到了 2026 年上半年。驱动因素不只是传统 scraping，而是 agentic behavior：一个人买相机可能访问 5 个网站，一个 agent 可能为了完成同一个任务访问 5,000 个网站。这意味着即使人类用户数量没有大幅变化，Web 也会承受截然不同的负载。
The infrastructure implication is blunt: more servers, more network capacity, more CPUs, more GPUs, more memory, and more security work. Prince compares the coming shift to COVID-era traffic spikes, then says those may look small if agent traffic compounds. The business implication is sharper. The web has been funded largely by advertising for nearly three decades, but, as Prince puts it, "bots don't click on ads." If agents consume content, compare products, and make decisions without producing human ad inventory, the old unit economics break.
基础设施含义很直接：更多 servers、network capacity、CPUs、GPUs、memory，以及更多 security work。Prince 把即将到来的变化和 COVID 期间的 traffic spike 对比，并认为如果 agent traffic 持续复合增长，那次 spike 可能很快显得很小。商业含义更尖锐。过去近三十年，Web 很大程度靠广告资助，但 Prince 的一句话很关键："bots don't click on ads." 如果 agents 消费内容、比较商品、做出决策，却不产生 human ad inventory，原有 unit economics 就会失效。
The most interesting part is Prince's view that brand itself changes in an agentic world. A brand used to be a shortcut for human expectations. Agents have near-infinite patience to inspect every option, so reputation, discovery, payments, and authorization may need new protocols. Cloudflare's likely role is to sit in the middle: deciding which machines are allowed in, helping sites charge or govern access, and making the agent-heavy Internet survivable.
最有意思的是 Prince 对 brand 的判断。在 agentic world 里，brand 不再只是人类预期的 shortcut。Agents 有近乎无限的耐心检查所有选项，所以 reputation、discovery、payments 和 authorization 都可能需要新协议。Cloudflare 可能会处在中间层：判断哪些机器可以进来，帮助网站收费或管理访问，并让 agent-heavy Internet 继续可用。
https://www.youtube.com/watch?v=UN47z_opfmo
Claude Blog: Building intelligent apps for Apple platforms with Claude in the Foundation Models framework
Anthropic is releasing a Swift package that lets Apple developers call Claude through Apple's Foundation Models framework. The practical idea is model routing inside native apps: use Apple's on-device models for fast local tasks like summarization, extraction, and typed Swift values, then hand off harder work to Claude when the request needs multi-step reasoning, code generation, web search, or data analysis.
Anthropic 发布了一个 Swift package，让 Apple developers 可以通过 Apple's Foundation Models framework 调用 Claude。核心思路是在 native apps 里做 model routing：用 Apple 的 on-device models 处理 summarization、extraction、typed Swift values 等快速本地任务，再把需要 multi-step reasoning、code generation、web search 或 data analysis 的复杂请求交给 Claude。
The interesting product design is that Apple's framework can turn user intent into typed outputs through @Generable annotations before Claude ever sees the request. That gives the Claude API cleaner inputs than raw user text and lets the app stream Claude's response back into the same SwiftUI view. Anthropic says the package handles "streaming, tool calls, and structured responses."
有意思的产品设计是，Apple's framework 可以先通过 @Generable annotations 把用户意图转成 typed outputs，再交给 Claude。这样 Claude API 收到的是比 raw user text 更干净的输入，app 也可以把 Claude 的 response stream 回同一个 SwiftUI view。Anthropic 说这个 package 会处理 "streaming, tool calls, and structured responses."
The examples are concrete: a journaling app can create local prompts, then ask Claude to find patterns across months of entries; a study app can define a term locally, then ask Claude to explain why it matters in a broader course context. Claude support is slated for Apple's Foundation Models framework across iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27.
示例很具体：journaling app 可以先在本地生成 prompts，再让 Claude 从数月 entries 中寻找 patterns；study app 可以先在本地解释术语，再让 Claude 说明这个概念为什么和课程其他内容相关。Claude support 将用于 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27 上的 Apple's Foundation Models framework。
https://claude.com/blog/claude-for-foundation-models
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders