七月十六日

Boris Cherny, who leads Claude Code at Anthropic, published a widely shared essay arguing that the old engineering habit of automating your own work is now the highest-leverage activity of the agent era. His three reasons: infra and DevX automation speeds up every agent in your fleet, not just you; moving fixes into lint rules, CI steps, and routines automates entire classes of busywork forever instead of solving them one-off; and most importantly, domain knowledge encoded as CLAUDE.md rules, skills, code comments, and memories lets engineers and non-engineers contribute to a codebase on day one. His sharpest line: if a PR gets rejected for not using the right framework or pattern, "these are failures of automation." His prescription: "Every team should be writing the CLAUDE.md's, REVIEW.md's, skills, and docs that enable agents to productively work in their codebase with zero additional context from the prompter."

Anthropic Claude Code 负责人 Boris Cherny 发表了一篇广为流传的长文，认为工程师"自动化自己工作"这个老习惯，在 agent 时代已经成为杠杆率最高的事情。他给出三个理由：基础设施和开发体验的自动化会加速你麾下的每一个 agent，而不只是你自己；把修复沉淀成 lint 规则、CI 步骤和例行流程，能一劳永逸地自动化整类杂活，而不是一次次单独解决；最重要的是，把领域知识编码成 CLAUDE.md 规则、skills、代码注释和 memories，能让工程师甚至非工程师第一天就能给代码库做贡献。他最犀利的一句话是：如果一个 PR 因为没用对框架或模式而被打回，"这些都是自动化的失败"。他开出的药方是："每个团队都应该去写 CLAUDE.md、REVIEW.md、skills 和文档，让 agent 无需提示者提供任何额外上下文，就能在代码库里高效工作。"

OpenAI's Codex and ChatGPT lead Thibault Sottiaux posted the findings from OpenAI's investigation into reports of GPT-5.6 unexpectedly deleting files: the incidents most commonly happen when full-access mode is on without sandboxing or auto review, and the model tries to override the $HOME env var to define a temporary directory, then mistakenly deletes the real $HOME. OpenAI is updating the developer message, guiding users toward safer permission modes, adding harness safeguards, and promising a detailed post-mortem. He also asked for feedback on usage limits — Codex Plus and Pro have had no 5-hour limit for a few days, only a weekly one — and, after the ChatGPT–Codex merge, polled what OpenAI should merge next.

OpenAI 的 Codex 和 ChatGPT 负责人 Thibault Sottiaux 公布了 OpenAI 对"GPT-5.6 意外删除文件"报告的调查结果：这类事故最常发生在开启了完全访问模式、且没有 sandbox 和 auto review 保护的情况下，模型试图覆盖 $HOME 环境变量来定义临时目录，结果误删了真正的 $HOME。OpenAI 正在更新 developer message，引导用户使用更安全的权限模式，并增加 harness 层面的保护措施，还承诺会发布一份详细的事故复盘。他同时征求关于用量限制的反馈：Codex Plus 和 Pro 已经取消 5 小时限制几天了，只保留每周上限；在 ChatGPT 和 Codex 合并之后，他还发起投票问大家 OpenAI 下一步该合并什么。

Google VP Josh Woodward announced Gemini Spark is rolling out to more Ultra subscribers worldwide with four upgrades: it can open and edit Google Docs, read comments in Sheets and Slides, runs more than 50% faster, and parallel-processes across multiple sources — with a Pro-tier access update coming soon. He also shared Google's first-ever Gemini Southeast Asia Report: active users in the region more than doubled in the last year, 70% of prompts are submitted in native languages, and 40% of prompts use only voice, image, or video.

Google 副总裁 Josh Woodward 宣布 Gemini Spark 正在向全球更多 Ultra 订阅用户开放，并带来四项升级：可以打开并编辑 Google Docs、能读取 Sheets 和 Slides 中的评论、速度提升超过 50%、支持跨多个来源并行处理，Pro 用户的开放计划也即将公布。他还分享了 Google 首份 Gemini 东南亚报告：该地区活跃用户在过去一年翻了一倍多，70% 的 prompt 使用本地语言提交，40% 的 prompt 只用语音、图片或视频。

Swyx of Latent Space and Cognition pushed back hard on a viral take downplaying computer-use agents (CUA), calling it a "Gell-Mann moment": he has tracked the field since the World of Bits paper in 2017, and says GPT 5.6 plus Superapp is better at computer use than everything that came before. He now has his nontechnical team run as much of their knowledge work as possible through CUA — payment portals, invoicing, vendor and attendee data requests. His warning: "underestimating capabilities is quite a dangerous category error if you are doing any ai decisionmaking."

Latent Space 和 Cognition 的 Swyx 强硬回击了一个唱衰 computer-use agent (CUA) 的热门观点，称这是他的"Gell-Mann 时刻"：他从 2017 年的 World of Bits 论文起就一直跟踪这个领域，并表示 GPT 5.6 加上 Superapp 的电脑操作能力已经超过此前所有产品。他现在要求自己的非技术团队尽可能把知识型工作跑在 CUA 上，包括支付门户、开发票、供应商和参会者的数据请求。他的警告是："如果你在做任何 AI 决策，低估能力是一个相当危险的类别错误。"

Anthropic Claude Code team member Thariq distilled his prompting philosophy into one line: "ideal prompting technique is: thin prompts, thick artifacts + context, thin skills." Quoting Boris Cherny's automation essay, he added a maxim of his own: "software engineering is the profession of automation."

Anthropic Claude Code 团队的 Thariq 把他的 prompt 哲学浓缩成一句话："理想的 prompt 技巧是：轻 prompt，重 artifacts 和上下文，轻 skills。"在转发 Boris Cherny 那篇自动化长文时，他补了一句自己的格言："软件工程就是自动化这门职业。"

Box CEO Aaron Levie shared notes from a dinner with IT leaders of large enterprises on agent adoption: change management is still the biggest blocker, and embedding engineers directly into business functions as internal forward-deployed engineers is working; cross-functional agent workflows create hard data-modeling and permissions problems, so agentic systems need their own roles and privileges; coding budgets run $1,000–$5,000 per developer per month while non-coding AI budgets remain far smaller; there's consensus that all enterprise software must become headless, with "huge warning for existing software vendors" that don't play nice with agents; and frontier models are now finding sophisticated security risks by chaining vulnerabilities, leaving companies with long patch backlogs. Separately, he cheered new US open-weights model releases, predicting a future of frontier orchestrators combined with cheaper workhorse models.

Box CEO Aaron Levie 分享了他与多家大型企业 IT 负责人共进晚餐、讨论企业 agent 落地的笔记：变革管理仍是最大阻碍，而把工程师以内部 FDE (forward-deployed engineer) 的形式直接嵌入业务部门的做法正在奏效；跨部门的 agent 工作流带来棘手的数据建模和权限问题，因此 agent 系统需要拥有自己的角色和权限；编程方面的预算达到每位开发者每月 1,000 到 5,000 美元，而非编程类 AI 预算仍然小得多；大家一致认为所有企业软件未来都必须走向 headless，对那些不愿与 agent 良好协作的传统软件厂商来说是"巨大的警告"；前沿模型现在还能通过串联漏洞发现复杂的安全风险，让企业排起了长长的补丁待办清单。另外，他为美国实验室新发布的开放权重模型叫好，预测未来会是前沿模型做编排、搭配更低成本的主力模型。

Vercel CEO Guillermo Rauch shared Vercel Sandbox growth numbers: DAUs growing 100% month over month, more than 3.5 million sandboxes created per day, powering Notion, Airtable, Meta, Zapier, CodeRabbit, and others. He also showcased Web Analytics API use cases, like asking your agent to correlate visitors and custom events ("purchase", "checkout") with deployments and performance, or building custom frontends that plot analytics alongside Stripe and Resend data.

Vercel CEO Guillermo Rauch 公布了 Vercel Sandbox 的增长数据：日活以每月 100% 的速度增长，每天创建超过 350 万个 sandbox，客户包括 Notion、Airtable、Meta、Zapier、CodeRabbit 等。他还展示了 Web Analytics API 的玩法，比如让你的 agent 把访客和自定义事件（"purchase"、"checkout"）与部署和性能变化关联起来，或者构建自定义前端，把分析数据和 Stripe、Resend 的数据画在一起。

AI educator Peter Yang called the disconnect between ChatGPT Live and Codex OpenAI's biggest missed opportunity: on a walk, ChatGPT Live couldn't pull up his Google Doc until he manually triggered the Documents plugin — at which point it suddenly had the right context. He wants the voice assistant to have full access to all the plugins, tools, and browser use that Codex has, so it can reply to emails, schedule meetings, edit docs, and ship code during a live conversation. "Why build such a great voice assistant but have it not be able to do anything?"

AI 教育者 Peter Yang 认为 ChatGPT Live 和 Codex 互不连通是 OpenAI 错失的最大机会：他散步时让 ChatGPT Live 调出自己的 Google Doc，它说做不到；直到他手动触发 Documents 插件，它才突然有了正确的上下文。他希望这个语音助手能完整接入 Codex 拥有的所有插件、工具和浏览器操作能力，从而在实时对话中回复邮件、安排会议、编辑文档、发布代码。"为什么造出这么棒的语音助手，却让它什么都干不了？"

Builder Zara Zhang argued that if you want agents to actually work inside a company, you have to design the company so they can read it — citing Shopify's internal agent that had no private chat function at all, only public channels, with peer learning as the side effect. She also reflected that because she never learned programming the traditional way, using coding agents has been purely an act of creativity and self-expression: "GitHub is basically my Substack."

Builder Zara Zhang 提出，如果你想让 agent 真正在公司内部发挥作用，就必须把公司设计成 agent 能"读懂"的样子。她举了 Shopify 的例子：其内部 agent 完全没有私聊功能，只能在公开频道交流，副作用是促成了同事之间的互相学习。她还感慨，因为自己从未按传统路径学过编程，使用 coding agent 对她来说纯粹是一种创造和自我表达："GitHub 基本上就是我的 Substack。"

Y Combinator CEO Garry Tan endorsed skill files as a portability layer: "Skill files are portable and free you from frontier model dependency. This is a good thing."

Y Combinator CEO Garry Tan 为 skill 文件作为可移植层背书："Skill 文件是可移植的，能把你从对前沿模型的依赖中解放出来。这是件好事。"

Meta Senior Director of AI Madhu Guru proposed vocabulary for the visceral reaction of realizing mid-read that something is AI-written — "semantic nausea," "uncanny prose valley," or "synthetic shudder" — and admitted AI-isms had leaked into his own writing, so he now consciously uses AI mainly in the brainstorming phase and keeps the final writing human.

Meta AI 高级总监 Madhu Guru 为"读到一半突然意识到这是 AI 写的"那种生理反应发明了几个候选词："语义恶心 (semantic nausea)"、"文风恐怖谷 (uncanny prose valley)"、"合成颤栗 (synthetic shudder)"。他也承认 AI 腔曾渗入自己的写作，所以现在有意识地只在头脑风暴阶段用 AI，最终成稿保持人写。

Claude Code can now capture work progress as artifacts: live, shareable visual pages — PR walkthroughs, system explainers, dashboards, incident timelines, release checklists — that update themselves as your session works. Artifacts are built from the full context of a session (codebase, connectors, and the conversation itself), so a single incident page can combine the failing test, the error spike from a connected monitoring tool, and the root-cause reasoning from the session. Pages update in place: every publish is a new version at the same link, with version history and a gallery. The pitch for teams: "With artifacts, team members and stakeholders don't have to 'walk us through what the agent found' because they're all looking at the same view, with the same context." Artifacts are private to the author by default, shareable only within an authenticated org (they cannot be made public), with admin controls, retention policies, and a compliance API. Available in beta to Claude Team and Enterprise orgs from the CLI and desktop app.

Claude Code 现在可以把工作进展沉淀为 artifacts：实时、可分享的可视化页面，包括 PR 讲解、系统架构说明、仪表盘、事故时间线、发布清单等，并随着会话的推进自动更新。Artifact 基于会话的完整上下文构建（代码库、connectors 以及对话本身），因此一个事故页面可以同时汇集失败的测试、来自监控工具的错误飙升曲线，以及会话里的根因推理。页面原地更新：每次发布都是同一链接下的新版本，带版本历史和作品库。面向团队的卖点是："有了 artifacts，团队成员和利益相关方不再需要'让人讲一遍 agent 发现了什么'，因为大家看到的是同一个视图、同一份上下文。"Artifact 默认仅作者可见，只能在经过认证的组织内部分享（无法公开），并提供管理员控制、保留策略和合规 API。目前以 beta 形式向 Claude Team 和 Enterprise 组织开放，可从 CLI 和桌面应用使用。

The Takeaway: Meeting notes were never the prize — the real battle is over what interface everyone uses to work with AI, and today's leads are "easy come, easy go."

Chris Pedregal is the cofounder and CEO of Granola, the AI meeting notes app that reached a $1.5B valuation and grew to roughly 60–70 people. His first surprise: "Startups are like knife fights. And I thought startups were just really hard when they weren't working. Turns out they're really hard even when they're working as well." When Notion, OpenAI, and Zoom all cloned Granola's core feature, growth didn't budge — because in his view, "what people are fighting for today doesn't matter" compared to the much bigger fight over the AI-native work interface.

His most counterintuitive product bet: Granola pre-generates millions of pre-meeting briefs even though only around 10% ever get opened, because the person running two minutes late has a fifteen-second window where the brief is either instantly there or useless. The guiding metaphor is a handrail: invisible until you trip, then it must be load-bearing. Multiple founders sought him out to praise the feature despite modest usage numbers, which convinced him that "used a little but load-bearing when used" is a category metrics don't capture yet.

Strategically, Granola is going all in on two things: being 5x better than anyone at meeting-adjacent jobs, and making its context available to whatever agent you bring, with the API and MCP set to get "a lot better" over the next couple of months. His productization compass is watching what power users wire up themselves — a Hugging Face founder told him he shut down a months-old, hand-built Claude Code pre-meeting-brief workflow because Granola's version was better. That is the model: guinea-pig users invent workflows, then the product absorbs the best ones. About half of Granola's weekly users already use it in genuinely agentic ways — they just wouldn't call it that. And on cost discipline: the team only started tracking token spend last week, after one employee burned a couple thousand dollars in a day.

核心结论：会议纪要从来都不是终极奖品，真正的战场是所有人用什么界面与 AI 协作，而今天的领先优势"来得容易，去得也快"。

Chris Pedregal 是 AI 会议纪要应用 Granola 的联合创始人兼 CEO，公司估值达到 15 亿美元，团队扩张到大约 60 到 70 人。他的第一个意外感悟："创业就像持刀肉搏。我以前以为只有不顺的时候创业才难，结果发现顺风顺水的时候一样难。"当 Notion、OpenAI 和 Zoom 相继克隆了 Granola 的核心功能后，增长曲线纹丝不动。在他看来，与争夺 AI 原生工作界面这场更大的战役相比，"今天大家抢的东西根本不重要"。

他最反直觉的产品押注：Granola 预生成了数以百万计的会前简报，尽管只有大约 10% 会被打开。因为一个迟到两分钟的人只有十五秒的窗口期，简报要么立刻就在那里，要么毫无用处。指导思想是"扶手"这个比喻：平时隐形，可一旦你绊倒，它必须稳稳撑住你。好几位创始人专程找他称赞这个使用率并不高的功能，这让他确信"用得少但用时极其关键"是现有指标体系还无法衡量的一类价值。

战略上，Granola 全力押注两件事：在与会议相关的任务上做到比任何人好 5 倍，以及把 Granola 的上下文开放给你自带的任何 agent，API 和 MCP 会在未来几个月"大幅升级"。他的产品化指南针是观察高阶用户自己搭了什么：一位 Hugging Face 创始人告诉他，自己用 Claude Code 搭建、维护了几个月的会前简报工作流已经关掉了，因为 Granola 的版本更好。这就是模式：小白鼠用户发明工作流，产品吸收其中最好的。Granola 每周约有一半用户已经在以真正 agentic 的方式使用它，只是他们自己不会这么称呼。至于成本纪律：团队上周才开始统计 token 花销，起因是有员工一天烧掉了几千美元。