AI Builders
Digest

Anthropic Engineering 详述了他们在 claude.ai、Claude Code 和 Claude Cowork 三大产品中控制 agent 爆炸半径的实践。核心思路是环境层优先（沙箱、VM、egress 控制），模型层和行为监督作为补充。人类在环批准很快出现疲劳（用户约 93% 直接通过），因此他们转向确定性边界。Claude Code 用 OS 级沙箱（Seatbelt/bubblewrap）把权限提示减少了 84%。Claude Cowork 用完整本地 VM 隔离工作区。文中复盘了多次真实事故：信任对话框前的钩子执行、用户作为注入向量的钓鱼、以及通过已批准域名的数据外泄。原则是：先设计环境遏制，再匹配用户监督能力，并警惕自定义组件。

Anthropic Engineering details containment strategies for Claude agents across claude.ai, Claude Code, and Claude Cowork. The focus is environment-first defenses (sandboxes, VMs, egress controls) over probabilistic model-layer or human-in-the-loop checks. Approval fatigue was real—users approved ~93% of prompts—so they shipped OS-level sandboxes that cut prompts by 84% in Claude Code, and full local VMs for Claude Cowork. Key incidents included pre-trust-dialog hooks, user-as-injection phishing, and exfiltration via allowlisted domains. Core lessons: contain at the environment layer first, match isolation strength to user oversight capacity, and be wary of custom components.

Anthropic Engineering 公开复盘了过去一个月 Claude Code 感觉变差的三大独立变更。3 月 4 日把默认 reasoning effort 从 high 降到 medium 以降低延迟，结果被用户认为不够智能，4 月 7 日已回滚。3 月 26 日为优化缓存而清理空闲会话旧思考内容的逻辑有 bug，导致后续每轮都清空，模型显得健忘重复，4 月 10 日修复。4 月 16 日系统提示加入严格字数限制以降低冗长，却伤害了编码质量，4 月 20 日回滚。三个问题叠加造成了“整体变差”的印象。API 本身未受影响。Anthropic 已重置所有订阅用户用量限制，并承诺更严格的 prompt 变更评测与渐进发布。

Anthropic Engineering published a postmortem on three separate changes that made Claude Code feel worse for some users. On March 4 they lowered the default reasoning effort from high to medium to cut latency; users preferred higher intelligence and it was reverted April 7. A March 26 caching optimization that should have cleared old thinking only once on idle sessions instead cleared it every turn, causing forgetfulness and repetition; fixed April 10. An April 16 system-prompt length limit hurt coding quality and was rolled back April 20. The combined effect looked like broad degradation even though the API itself was unaffected. Usage limits have been reset for all subscribers.

Anthropic Engineering 介绍了 Claude Managed Agents 的架构演进。早期把 session、harness 和 sandbox 塞进同一个容器，导致“宠物”式运维：容器挂了会话就丢，调试困难，也无法轻松对接客户 VPC。解决方案是把大脑（harness + Claude）与手（sandbox/tools）和解耦，session 作为独立的 append-only 事件日志。容器变成可随时重建的 cattle，harness 崩溃后用 wake(sessionId) 从日志恢复。凭证永远不进入 sandbox，通过代理或初始化时注入。TTFT p50 下降约 60%，p95 下降超过 90%。设计目标是为“尚未想到的程序”留下稳定接口，让未来不同 harness 都能接入。

Anthropic Engineering describes the architecture of Claude Managed Agents. Early designs put session, harness and sandbox in one container—creating “pets” that were hard to recover and debug, and blocked easy VPC access. The fix was decoupling the brain (harness + Claude) from the hands (sandbox/tools) and making the session an independent durable event log. Containers became cattle that can be reprovisioned; harnesses recover via wake(sessionId). Credentials never enter the sandbox. Time-to-first-token dropped ~60% at p50 and >90% at p95. The interfaces are deliberately unopinionated about future harnesses.

OpenAI Codex & ChatGPT 团队的 Thibault Sottiaux 表示下周二 DevDay 将带来真正改变工作方式的功能，这是他们最雄心勃勃的冲刺，Astra 让新东西在极短时间内成为可能。他同时展示了 Voice 模式已深度融入日常：打电话给 ChatGPT 处理工作、邮件、编码、日历，并跨整个插件生态（含第三方）无缝工作。

OpenAI Codex & ChatGPT’s Thibault Sottiaux says DevDay next Tuesday will include things that should change how people work. It has been their most ambitious sprint and Astra made new capabilities possible in short time. He also notes Voice is now habitual—calling ChatGPT to handle work, email, coding and calendar across the full plugin ecosystem including third-party ones.

Claude Code 负责人 Boris Cherny 分享了过去几周 claude.ai 与 Desktop 应用大幅提速的技术细节，博客中有大量可复用的工程经验。他还解释了 Claude 如何对棘手状态机或竞态代码建模、寻找反例、复现并修复 bug——并非全量形式化验证，而是针对最棘手部分建模与修复。Thariq 补充他们开始发布带完整 prompt 与技巧的深度复盘帖，方便开发者复现。

Claude Code lead Boris Cherny shared how they made claude.ai and the Desktop app dramatically faster in recent weeks, with concrete techniques in the accompanying blog post. He also detailed how Claude models tricky state machines or race-prone code, finds counter-examples, reproduces bugs and fixes them—not full formal verification, but targeted modeling of the hairiest parts. Thariq noted they are experimenting with in-depth posts that include the actual prompts and techniques so others can replicate.

Vercel CEO Guillermo Rauch 系统拆解了成功 agent 的三大组件：大脑（模型+harness）、手（工具/浏览器/计算机）、文件（记忆/技能/仓库）。本地跑在 Mac Mini 上简单，但云端成本与安全要求把它们解耦。今天他们为 Sandbox 推出了配套的 Drives——可按需挂载的“外置磁盘”，让夜间记忆整合等任务无需启动完整 agent 计算机。Rauch 还提到自己用 Opus 5.5 优化 shell 启动时间，发现了其他模型错过的大量优化点。

Vercel CEO Guillermo Rauch broke down the three key components of successful agents: brain (model + harness), hands (tools/computer/browser), and files (memories/skills/repos). Running everything on a local Mac Mini is easy; cost-efficient and secure cloud agents require decoupling. Today they launched Drives as the companion to Sandbox—an attachable external disk so memory consolidation or similar jobs can run without booting the full agent computer. He also noted Opus 5.5 found many shell-startup optimizations other models missed.

Claude 官方账号宣布 Claude Marketplace 正式可用：可添加 Slack、Notion 等连接器与插件，购买 Cursor、CrowdStrike 等公司的 agents 与产品，并与 Accenture、Deloitte 等服务伙伴扩展规模。开发者也可在此上架自己的工具、agent 或服务。

The official Claude account announced Claude Marketplace is live. Users can add connectors and plugins such as Slack and Notion, buy agents and products from companies like Cursor and CrowdStrike, and scale with service partners including Accenture and Deloitte. Builders can list their own tools, agents or services.

Swyx（smol.ai / Latent Space）实测后确认 5.5 Opus 比 5 Opus 更简洁、更有品味，更少 slop，已将其设为 AINews 的新默认模型。Peter Yang 也给出高度评价：聪明、快、好玩、写作好、限速友好，且不断让人惊讶于其能力边界。他同时指出目前一家实验室有最好模型，另一家有最好 harness。

Swyx (smol.ai / Latent Space) confirmed after side-by-side testing that 5.5 Opus is far more concise and tasteful than 5 Opus with much less slop, and has made it the new default for AINews. Peter Yang called Opus 5.5 smart, fast, fun to talk to, good at writing, friendly on rate limits, and constantly surprising in capability. He noted one lab currently has the best model while another has the best harness.

Y Combinator 总裁兼 CEO Garry Tan 指出当前获取客户的两个并行重要趋势：一是做出软件 agents 真正想要的产品，二是用 agents 去让人们想要软件。他还对 Muse 给出高度评价，并转发了“这还没被定价进去”的观察。

Y Combinator President & CEO Garry Tan highlighted two simultaneous important trends in how startups get customers: make software that agents want, and use agents to make people want software. He also called Muse very impressive and amplified the view that certain shifts are not yet priced in.

Meta AI 高级总监 Madhu Guru 明确不同意 Ben Thompson 的观点。他指出消费者与“做事”的关系并非单一：逛街可以是娱乐，但雇屋顶工对大多数人是痛苦流程。早期互联网也有人质疑网购衣服或电视，技术最终提供了更好方式。对需要大量摩擦的任务，agent 的潜在需求巨大。

Meta AI Sr Director Madhu Guru respectfully disagreed with Ben Thompson. Consumers do not have one relationship with “doing things”—browsing clothes can be entertainment while hiring a roofer is miserable for most. Early-2000s arguments against online shopping sound similar. Technology keeps giving better ways; the latent demand for agents that remove friction on high-pain consumer tasks is immense.

前 Cursor / Notion / Stripe 设计师 Ryo Lu 写了一篇长文反思：我们把效率、生产力、速度奉为圭臬，却忘了如果没有时间深度思考，无穷生产有何意义。AI 本可成为新画布，让人做出奇怪、个人、不可能的东西，却正在变成 slop 工厂。真正的前沿或许不是更多速度，而是辨别力——知道什么不该做、何时停下、何时最有生产力的事情是睡觉、读历史、打电话给朋友或发呆。工具应给人更多生命，而不是把生命变成工具。

Former Cursor/Notion/Stripe designer Ryo Lu argued that worshipping efficiency, productivity and speed misses the point if there is no time left to think deeply. AI could be a new canvas for strange, personal, impossible things, yet much of it is becoming slop factories. The danger is not that AI makes us lazy but that it makes us endlessly busy—giving infinite production before intention, taste or stillness. The real frontier may be discernment: knowing what not to make and when to stop. Tools should give us more life, not turn life into a tool.

Box CEO Aaron Levie 转发并认同一段关于创意产业未来的愿景：随着门槛与成本下降，电影会更多而非更少，工作室能承担更多风险，全新叙事形式会出现。历史每次技术跃迁（同步声、彩色、计算机动画）都扩大了媒介边界。底层对创意技能与品味的需求不会消失，只是更多人能有效运用。

Box CEO Aaron Levie amplified a vision for the creative industry: as barriers and costs fall, more films will be made, studios will take more risks, and entirely new forms of storytelling will emerge. Historical tech shifts—from synchronized sound to color to computer animation—have expanded the medium. Underlying need for creative skill and taste does not go away; more people can leverage those skills effectively.

SPC 普通合伙人、前 Facebook / Dropbox Aditya Agarwal 指出，频繁跳槽和低人员流动率都说明很多问题。一起工作 3 年以上的团队吞吐量和韧性显著更高。投资者和员工常犯的错误是只问团队规模而不问 longevity。团队的 AUC 才是正确指标。

SPC General Partner and former Facebook/Dropbox Aditya Agarwal noted that high career mobility and low company turnover both reveal a lot. A team that has worked together 3+ years has significantly higher throughput and resilience than a newly formed one. A common mistake is asking only about team size rather than longevity. AUC for teams is the right metric.

Google Labs VP Josh Woodward 预告更多热门应用即将接入 Gemini。Replit CEO Amjad Masad 强调年轻人需要打破陈规、质疑深层信念并尝试不可能。Boris Cherny 在计算机历史博物馆见到 Donald Knuth。Peter Yang 展示了一个能生成高质量非企业风幻灯片的 prompt。Peter Steinberger 称赞 Astra 表现出色。Dan Shipper 宣布 @every agent 策划了九月线下聚会。

Google Labs VP Josh Woodward teased more favorite apps coming to Gemini. Replit CEO Amjad Masad stressed that young people are needed to break molds and attempt the impossible. Boris Cherny met Donald Knuth at the Computer History Museum. Peter Yang shared a prompt that produced high-quality non-corporate slides. Peter Steinberger said Astra is crushing it. Dan Shipper noted the @every agent planned their September meetup.