六月三十日

Anthropic's Boris Cherny, who works on Claude Code, previewed a meaningful workflow shift coming in the next version: subagents will run in the background by default, so you can keep talking to Claude while your subagents grind away on their tasks. If you'd rather an agent run in the foreground and block, you just tell Claude. It's a small default change with a big ergonomic payoff — parallel agent work stops interrupting your main conversation.

Anthropic 的 Boris Cherny 负责 Claude Code。他预告了下一个版本里一个挺实在的工作流变化:subagent(子智能体)将默认在后台运行,这样你可以一边和 Claude 继续对话,一边让子智能体在后面埋头干活。如果你更希望某个 agent 在前台运行并阻塞主流程,直接告诉 Claude 即可。这只是一个默认行为的小改动,却带来很大的体验提升——并行的 agent 工作不再打断你和主对话的交流。

OpenAI's Thibault Sottiaux, who leads Codex and ChatGPT, published a detailed postmortem after users reported Codex burning through usage far faster than expected. There was no single culprit; several smaller problems compounded. On the actual-usage side, auto-review had become more proactive, another change triggered more subagent work, and background suggestions could run twice or retry too aggressively after failures — all of which OpenAI reverted. On the reporting side, auto-review was being mislabeled as GPT-5.4 usage and rate-limited or failed requests were wrongly shown as turns (they were never charged). As immediate relief, he reset everyone's usage limits and banked an extra reset for any usage over the next 24 hours, then shipped hotfixes across the CLI, desktop app, and usage backend with new monitoring to catch background-usage regressions sooner.

OpenAI 的 Thibault Sottiaux 负责 Codex 和 ChatGPT。在用户反映 Codex 额度被消耗得远超预期之后,他发布了一份详细的复盘。问题没有单一元凶,而是几个小毛病叠加所致。在"真实用量"这边:auto-review 变得过于主动,另一处改动触发了更多 subagent 工作,后台建议还可能重复运行、或在失败后过于频繁地重试——这些 OpenAI 都已回滚。在"用量统计"这边:auto-review 被错误地计为 GPT-5.4 的用量,被限流或失败的请求也被错误地显示成了对话轮次(其实这些从未计费)。作为即时补救,他重置了所有人的额度,并为接下来 24 小时内的用量额外储备了一次重置,随后在 CLI、桌面端和用量后台都推送了热修复,并加上了新的监控,以便更早发现后台用量的回归问题。

Separately, Sottiaux unveiled a replacement for Codex's coarse sandbox modes: reusable, inheritable permission profiles that bind OS-enforced file read/write/deny rules (even patterns like `**/*.env`) to per-domain network access and Unix sockets, plus fail-closed admin allowlists. The goal is genuine least-privilege scoping per task rather than blunt on/off sandboxing.

此外,Sottiaux 还公布了一套用来取代 Codex 粗粒度沙箱模式的方案:可复用、可继承的权限配置(permission profiles),把由操作系统强制执行的文件读/写/拒规则(甚至像 `**/*.env` 这样的通配模式)与按域名划分的网络访问、Unix socket 绑定在一起,再加上"默认拒绝"(fail-closed)的管理员白名单。目标是为每个任务做到真正的最小权限,而不是一刀切的沙箱开关。

AI creator Peter Yang shared a contrarian take on where the coding agents fall short: for writing and editing, plain vanilla Claude on the web is still the best, beating both Codex and Claude Code. His guess is that something in the coding agents' system prompts — tuned for terse, action-oriented engineering output — quietly makes them worse writers. The implication for builders: the harness you reach for matters, and the best tool for prose may not be the one optimized for code.

AI 创作者 Peter Yang 分享了一个反共识的观点,直指编码 agent 的短板:就写作和编辑而言,网页版那个朴素原味的 Claude 依然是最好的,胜过 Codex 和 Claude Code。他猜测,编码 agent 的系统提示词里有某些东西——为简短、行动导向的工程输出而调校——悄悄地让它们变成了更差的写手。对 builder 的启示是:你顺手抓起的那个工具(harness)很关键,写散文最好的工具,未必就是那个为写代码而优化过的。

Madhu Guru, a former product leader at Google who worked on Gemini, Veo, and Nano Banana, offered an underappreciated take: the rise of strong open-weight models like GLM will actually strengthen Google's position, not weaken it. His reasoning is that as more companies start experimenting with fine-tuning open-weight models, the value accrues to the infrastructure layer — enterprises want the flexibility to run and fine-tune open models on a managed platform with enterprise-grade reliability, security, and support, and Google Cloud is well positioned there. He adds the often-forgotten kicker that Google owns much of the compute stack underneath.

Madhu Guru 曾是 Google 的产品负责人,参与过 Gemini、Veo 和 Nano Banana。他抛出了一个被低估的观点:像 GLM 这样强大的开源权重(open-weight)模型崛起,实际上会强化而非削弱 Google 的地位。他的逻辑是,随着越来越多公司开始尝试微调开源权重模型,价值会沉淀到基础设施这一层——企业想要的是在一个具备企业级可靠性、安全性和支持的托管平台上,灵活地运行和微调开源模型,而 Google Cloud 在这方面占位很好。他还补了一句常被忽略的关键:Google 还掌握着底层很大一部分的算力栈。

Anthropic's Thariq, who works on Claude Code, gave a candid look at his actual writing process, which doubles as a lesson in iteration: do some engineering work, talk to a bunch of people about it, brainstorm and research with Claude, write a post, give one or two talks on it, rewrite the post, give another talk, rewrite the intro, wake up at 6am and rewrite it again, then finally post. The takeaway is that good writing isn't a single draft — it's many passes shaped by talking, teaching, and building, with AI woven in as a research and brainstorming partner.

Anthropic 的 Thariq 负责 Claude Code。他坦诚地展示了自己真实的写作流程,这本身也是一堂关于"反复打磨"的课:先做一些工程工作,再找一堆人聊它,和 Claude 一起头脑风暴并做调研,写出一篇帖子,围绕它做一两场分享,重写帖子,再做一场分享,重写开头,清晨六点醒来再改一遍,最后才发出来。要点是:好文章不是一稿写成的——它是在交谈、讲授和构建中反复打磨出来的多次迭代,而 AI 被编织进来,充当调研和头脑风暴的伙伴。

Vercel CEO Guillermo Rauch announced a concrete platform upgrade: functions on Vercel can now be 20x larger. He paired it with a teaser — "You can deploy anything and everything to Vercel. More tomorrow" — signaling a broader push to remove size and runtime limits so developers can ship heavier workloads without leaving the platform.

Vercel CEO Guillermo Rauch 宣布了一项实打实的平台升级:Vercel 上的 function(函数)现在可以做到原来 20 倍的大小。他还顺势放出预告——"你可以把任何东西、所有东西都部署到 Vercel,明天还有更多"——透露出一个更大的方向:不断拆掉体积和运行时的限制,让开发者无需离开平台就能部署更重的工作负载。

Box CEO Aaron Levie laid out what he sees as one of the central debates in AI, and it hinges entirely on one variable: how close to the frontier open-weight models can stay. If a closed stack is perpetually ahead by a wide margin, then a vertically integrated, gatekeeping-in-the-US approach can work — you always control who gets the best technology, and demand stays high enough that it favors you. But if open weights can remain a close second to frontier intelligence, the equation reverses: a highly regulated approach lets you keep the frontier market while the vast majority of tokens flow to an alternative stack — model plus underlying hardware — controlled and monetized by someone else. Your entire opinion on how to regulate AI, he argues, falls out of your belief about that one gap.

Box CEO Aaron Levie 阐述了他眼中 AI 领域最核心的争论之一,而这一切都取决于一个变量:开源权重模型能在多大程度上紧咬前沿。如果闭源技术栈永远以巨大优势领先,那么"垂直整合 + 在美国设卡把关"的路线就行得通——你始终掌控谁能拿到最好的技术,而需求又足够旺盛,局面对你有利。但如果开源权重能够稳居前沿智能之后的"紧随第二位",那么等式就反转了:高度管制的路线会让你守住前沿市场,但绝大多数 token 会流向另一套技术栈——模型加上底层硬件——而这套栈由别人掌控、由别人变现。他认为,你对"该如何监管 AI"的全部看法,都源自你对那一道差距的判断。

Builder Zara Zhang shipped and open-sourced a small but practical tool: a Chrome extension that turns your "read later" list into dedicated reading time on your calendar. Save 5 articles and it auto-books a 30-minute "reading block" on your Google Calendar with the links included, so you'll actually sit down and go through them. There's no account and no server — everything runs locally.

Builder Zara Zhang 做了一个小而实用的工具并开源了:一个 Chrome 扩展,把你的"稍后读"清单变成日历上专门的阅读时间。存够 5 篇文章,它就会自动在你的 Google 日历上预约一个 30 分钟的"阅读时段",并把链接一并附上,这样你就真的会坐下来把它们读完。没有账号,也没有服务器——所有东西都在本地运行。

She also surfaced a sharp observation, attributed to a PM at Anthropic: "The market value of writing has gone way up." Clear writing and articulation are now required both for building a product (steering models effectively) and for building an audience — and she notes that writing well has historically been undervalued in tech, and that it's time to change that.

她还转述了一个犀利的观察,出处是 Anthropic 的一位 PM:"写作的市场价值已经大幅上升。"如今,清晰的写作和表达,既是构建产品(有效地驾驭模型)的必备能力,也是积累受众的必备能力——她指出,写作能力在科技圈一直被低估,而现在是时候改变这一点了。

Anthropic's official Claude account announced that Claude in Microsoft Foundry is now generally available, hosted on Azure. Azure customers get Claude Opus 4.8 and Claude Haiku 4.5 with Azure authentication, billing, and commitment retirement. Inference runs on Azure infrastructure operated by Anthropic, with prompt caching and extended thinking supported today and more capabilities on the way.

Anthropic 官方的 Claude 账号宣布,Microsoft Foundry 中的 Claude 现已正式可用(GA),托管在 Azure 上。Azure 客户可以用上 Claude Opus 4.8 和 Claude Haiku 4.5,并享受 Azure 的认证、计费以及承诺额度抵扣(commitment retirement)。推理运行在由 Anthropic 运营的 Azure 基础设施上,目前已支持 prompt caching 和 extended thinking(扩展思考),更多能力还在路上。

An update on recent Claude Code quality reports

Anthropic published an unusually candid postmortem tracing a month of "Claude got worse" reports to three separate, overlapping changes — none of which touched the API or inference layer. First, on March 4 it lowered Claude Code's default reasoning effort from high to medium to cut latency; users found it less intelligent, and the default was reversed on April 7 (now xhigh for Opus 4.7, high for everything else). Second, a March 26 caching optimization meant to clear stale thinking *once* contained a bug that cleared it on *every* turn for the rest of a session, making Claude "forgetful, repetitive" and quietly draining usage limits through cache misses; it was fixed April 10. Third, an April 16 system-prompt line — "keep text between tool calls to ≤25 words... final responses to ≤100 words" — hurt coding quality by 3% in later ablations and was reverted April 20. Because each change hit a different slice of traffic on a different schedule, "the aggregate effect looked like broad, inconsistent degradation." Notably, when back-testing its Code Review tool against the offending pull requests, Opus 4.7 found the caching bug while Opus 4.6 didn't. Going forward, Anthropic is gating system-prompt changes behind per-model evals, ablations, soak periods, and gradual rollouts, having more internal staff run the exact public build, and resetting usage limits for all subscribers.

Anthropic 发布了一份罕见坦诚的复盘,把持续一个月的"Claude 变笨了"反馈,追溯到三处各自独立、又彼此重叠的改动——没有一处涉及 API 或推理层。第一,3 月 4 日,它把 Claude Code 的默认推理强度从 high 降到 medium 以缩短延迟,用户觉得变笨了,该默认值在 4 月 7 日被改回(现在 Opus 4.7 默认 xhigh,其余模型默认 high)。第二,3 月 26 日一项本应只清理*一次*陈旧思考的缓存优化,藏着一个 bug,导致在一个会话的剩余过程中*每一轮*都清理,让 Claude 变得"健忘、重复",还通过缓存未命中悄悄消耗额度;此问题于 4 月 10 日修复。第三,4 月 16 日加进系统提示词的一句——"工具调用之间的文字控制在 25 词以内……最终回复控制在 100 词以内"——在后续的消融实验中让编码质量下降了 3%,于 4 月 20 日回滚。由于每处改动按不同的时间表命中了不同的流量切片,"叠加起来的效果看上去就像是大范围、无规律的退化"。值得一提的是,在用其 Code Review 工具回测这些问题 PR 时,Opus 4.7 找出了那个缓存 bug,而 Opus 4.6 没有。展望未来,Anthropic 将对系统提示词改动设卡,要求过逐模型评测、消融实验、浸泡期(soak period)和灰度发布,并让更多内部员工使用与公开版完全一致的构建,同时为所有订阅用户重置额度。

Scaling Managed Agents: Decoupling the brain from the hands — by Lance Martin, Gabe Cemaj, and Michael Cohen

Anthropic detailed the architecture behind Managed Agents, its hosted service for running long-horizon agents through interfaces "meant to outlast any particular implementation." The core move is decoupling the "brain" (Claude plus its harness) from the "hands" (sandboxes and tools) and the "session" (an append-only event log). The original design crammed all three into one container — which, in the pets-vs-cattle analogy, made the server a "pet" you couldn't afford to lose: if a container failed, the session was lost, and credentials sat in the same container as untrusted agent-generated code, so a prompt injection only had to convince Claude to read its own environment. The fix makes the harness call the container like any tool — `execute(name, input) → string` — so the container becomes cattle; if it dies, the harness catches a tool-call error. The harness itself becomes stateless and recoverable via `wake(sessionId)`, because the durable session log lives outside it and is interrogated with `getEvents()`. Tokens are made unreachable from the sandbox (git tokens wired into the remote at init, MCP OAuth held in a vault behind a proxy). The payoff: provisioning containers only when a tool call actually needs one dropped p50 time-to-first-token roughly 60% and p95 over 90%, while enabling "many brains, many hands" — brains can even pass hands to one another.

Anthropic 详述了 Managed Agents 背后的架构。这是它用于运行长周期 agent 的托管服务,所依赖的接口"旨在比任何具体实现都更长寿"。核心思路是把"大脑"(Claude 及其 harness)、"双手"(沙箱和工具)和"会话"(只追加的事件日志)三者解耦。最初的设计把三者全塞进一个容器——用"宠物 vs 牛群"的比喻来说,这让那台服务器成了一只你输不起的"宠物":容器一旦挂掉,会话就丢了;而且凭据与不受信任的、由 agent 生成的代码同处一个容器,因此一次 prompt injection 只需说服 Claude 去读取自己的环境变量即可。修复办法是让 harness 像调用任何工具一样去调用容器——`execute(name, input) → string`——于是容器变成了"牛":挂了就挂了,harness 捕获一个工具调用错误即可。harness 自身也变成无状态、可通过 `wake(sessionId)` 恢复,因为持久化的会话日志存在它之外,并通过 `getEvents()` 来检索。凭据被设计成沙箱内不可触及(git token 在初始化时直接接入远端,MCP 的 OAuth 令牌存在代理后面的保险库里)。回报是:只在工具调用真正需要时才去配置容器,把首字延迟(TTFT)的 p50 降低了约 60%、p95 降低超过 90%,同时实现了"多大脑、多双手"——大脑之间甚至可以互相传递双手。

New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels

Claude Managed Agents can now operate in a sandbox you control and connect to your private MCP servers, keeping both agent execution and the services it reaches within your enterprise perimeter. The agent loop — orchestration, context management, error recovery — stays on Anthropic's infrastructure, while tool execution moves to your own configured environment, so files and repositories never leave and you control compute sizing and the runtime image. Self-hosted sandboxes is in public beta and MCP tunnels in research preview. You can bring any sandbox client or start with supported providers: Cloudflare (microVMs and isolates with zero-trust secrets injection), Daytona (long-running, stateful, pause-and-restore computers), Modal (sub-second startup, scaling to hundreds of thousands of concurrent sandboxes), and Vercel (VM security plus VPC peering with millisecond startup). MCP tunnels let agents reach internal databases, private APIs, and ticketing systems via a lightweight gateway that makes a single outbound connection — no inbound firewall rules or public endpoints. Real deployments cited include Amplitude's Design Agent on Cloudflare, Clay's GTM agent Sculptor on Daytona, and Rogo's institutional-finance analyst agent on Vercel.

Claude Managed Agents 现在可以在你自己掌控的沙箱中运行,并连接到你的私有 MCP 服务器,从而把 agent 的执行和它所访问的服务都圈在你的企业边界之内。agent 主循环——编排、上下文管理、错误恢复——仍然留在 Anthropic 的基础设施上,而工具执行则迁移到你自己配置的环境里,这样文件和代码库都不会外流,算力规格和运行时镜像也由你掌控。自托管沙箱已进入公开测试,MCP tunnels 处于研究预览阶段。你可以接入任意沙箱客户端,也可以从受支持的提供商起步:Cloudflare(microVM 与 isolate,配合零信任的密钥注入)、Daytona(长时运行、有状态、可暂停可恢复的"计算机")、Modal(亚秒级启动,可扩展到数十万并发沙箱),以及 Vercel(VM 级安全外加 VPC peering,毫秒级启动)。MCP tunnels 让 agent 能够通过一个轻量网关访问内部数据库、私有 API 和工单系统,该网关只发起一条对外连接——无需入站防火墙规则,也没有公网端点。文中列举的真实部署包括:Amplitude 跑在 Cloudflare 上的 Design Agent、Clay 跑在 Daytona 上的 GTM agent Sculptor,以及 Rogo 跑在 Vercel 上的机构金融分析师 agent。

The Takeaway: Saving Intel is a "crawl, walk, run" job — strengthen the balance sheet, simplify the product, obsess over foundry yield and trust, and bet on new materials — run by a venture investor at heart who still wants a 10x.

Lip-Bu Tan is the legendary semiconductor investor behind Walden, the former CEO of Cadence for thirteen years (where he delivered roughly 76–85x returns to shareholders), and, at 66, now CEO of Intel. He took what he calls "the hardest job in the industry" not because he needed it but to save an iconic company critical to the US — a resolve tested when, one early morning, President Trump asked him to resign over a conflict of interest. "I had to convince myself, first of all, I don't need this job. I do it purely to save Intel." His turnaround playbook is sequential: first strengthen a "really horrible" balance sheet (welcoming the US government as a major shareholder, plus Jensen Huang's $5B investment that "become $25 billion now," and SoftBank), then simplify the product line and listen relentlessly to customers, with all engineering reporting directly to him.

On the macro shift, he argues agentic AI and inference have made CPUs suddenly hot — the old 1:8 training CPU-to-GPU ratio is moving toward 1:4 or even 1:1. Foundry, he stresses, is fundamentally a trust business where yield, defect density, and cycle time decide whether customers hand you their wafers; Intel is pushing 14A (1.4nm) and already planning 1nm and 0.7nm, while collaborating weekly with Elon Musk on his TeraFab. As CMOS runs out of steam, Tan is betting on new materials — gallium nitride, silicon carbide, indium phosphide — plus glass and even artificial-diamond packaging. His investor's frame stays constant: find the real bottleneck, target a hyperscale first customer, and back teams that adapt, since "nine of the 10 company I invest, halfway they change their business plan because market have changed." Fourteen months in, with a 6x return to shareholders already, he insists "it's just a beginning."

核心要点: 拯救 Intel 是一份"爬、走、跑"的活儿——夯实资产负债表、精简产品、死磕代工厂的良率与信任、押注新材料——而操盘者,是一位骨子里依然想要 10 倍回报的风险投资人。

Lip-Bu Tan(陈立武)是 Walden 背后那位传奇的半导体投资人,曾执掌 Cadence 长达十三年(任内为股东带来约 76–85 倍回报),如今 66 岁,出任 Intel CEO。他接下这份自称"全行业最难的工作",并不是因为他需要它,而是为了拯救这家对美国至关重要的标志性公司——这份决心曾受到考验:某个清晨,特朗普总统以利益冲突为由要他辞职。"我得先说服我自己:第一,我并不需要这份工作。我做它,纯粹是为了拯救 Intel。"他的扭转计划是分步走的:先夯实"相当糟糕"的资产负债表(欢迎美国政府成为大股东,加上 Jensen Huang 那笔"如今已变成 250 亿美元"的 50 亿美元投资,以及 SoftBank),再精简产品线、不遗余力地倾听客户,并让所有工程团队直接向他汇报。

谈到宏观变化,他认为 agentic AI 和推理让 CPU 突然变得抢手——过去训练场景里 1:8 的 CPU 比 GPU 比例,正在向 1:4 甚至 1:1 移动。他强调,代工本质上是一门信任生意,良率、缺陷密度和周期时间决定了客户是否把晶圆交给你;Intel 正在推进 14A(1.4 纳米),并已在规划 1 纳米和 0.7 纳米,同时每周与 Elon Musk 就他的 TeraFab 展开合作。随着 CMOS 逐渐到达极限,陈立武押注新材料——氮化镓、碳化硅、磷化铟——以及玻璃乃至人造金刚石封装。他作为投资人的思维框架始终不变:找到真正的瓶颈,锁定一个超大规模(hyperscale)的首位客户,并支持那些懂得变通的团队,因为"我投的 10 家公司里有 9 家,走到一半就因为市场变了而修改了商业计划"。上任十四个月、已为股东带来 6 倍回报的他坚持说:"这才只是个开始。"