AI Builders
Digest

Anthropic Engineering 分享了过去两年在 claude.ai、Claude Code 和 Claude Cowork 上部署 agent 时的安全实践。核心观点是：风险由失败概率和理论爆炸半径组成，模型能力提升后爆炸半径只会变大，因此必须先在环境层做硬边界，再在模型层做行为引导。他们总结了三种隔离模式——ephemeral container（claude.ai）、human-in-the-loop sandbox（Claude Code）和 sealed VM（Claude Cowork）——并坦承多次事故都发生在自定义组件而非成熟的 hypervisor/gVisor 上。关键教训包括：用户批准疲劳会让人机循环失效、允许名单应视为能力授予而非目的地过滤、以及“你自己写的软件往往是最薄弱的一环”。文章强调匹配隔离强度与用户监督能力，并呼吁行业在 agent 身份、共享基准和披露规范上共同投入。

Anthropic Engineering shares two years of lessons containing agents across claude.ai, Claude Code, and Claude Cowork. Risk is failure probability times theoretical blast radius; as models improve the latter only grows, so hard environmental boundaries must come first. Three isolation patterns emerged: ephemeral containers for chat, HITL sandboxes for developers, and sealed VMs for knowledge workers. Most painful incidents stemmed from custom proxies and pre-trust config loading, not battle-tested hypervisors. Key principles: treat allowlists as capability grants, match isolation strength to user oversight capacity, and never rely solely on model-layer defenses.

Anthropic 调查了过去一个月用户反馈 Claude Code 变笨、健忘、啰嗦的问题，定位到三个独立变更：3 月 4 日把默认 reasoning effort 从 high 降到 medium（已于 4 月 7 日回滚）、3 月 26 日清理空闲会话旧 thinking 的 bug 导致后续每轮都丢上下文（4 月 10 日修复）、以及 4 月 16 日为降低 verbosity 加入的系统提示词意外伤害了编码质量（4 月 20 日回滚）。API 本身未受影响。团队已重置所有订阅用户的用量限额，并承诺更严格的系统提示词评估、更广的内部公测以及改进的 Code Review 工具。

Anthropic traced reports of degraded Claude Code quality to three separate changes: defaulting reasoning effort to medium (reverted), a bug that repeatedly cleared prior thinking after idle sessions (fixed), and a verbosity-limiting system-prompt line that hurt coding quality (reverted). The API layer was unaffected. All subscribers received usage-limit resets, and the team is tightening prompt-change controls and expanding evals.

Box CEO Aaron Levie 指出，结合 agent swarm、更好的 computer use、新一轮 API/MCP 以及 Muse/Instinct 等新形态，企业和个人将把 agent 投向远比单次会话提示多得多的任务——7×24 招聘、客户信号扫描、全量代码安全审查、暴力测试等。他认为我们目前只处于这些 agent 形态的 1%，部署位置、管理方式和预算分配都将发生根本变化。他同时强调，agent 对系统的使用量将是人类的 100 倍，企业数据保护与治理将是 21 世纪最复杂的安全挑战之一，安全与生产力必须同步设计。

Box CEO Aaron Levie argues that agent swarms, improved computer use, new APIs/MCPs and form factors like Muse will drive agents into 100× more background tasks than anyone currently imagines—continuous talent scouting, customer-signal mining, full-code security review and more. We are only 1% into the shape of these systems. He also flags that protecting enterprise data when agents use systems 100× more than people will be one of the century’s hardest governance problems.

OpenAI Codex & ChatGPT 团队的 Thibault Sottiaux 表示，本周将出现原本预期只在 DevDay 2025 才会出现的级别的产品发布，并公开征集“应该从 Codex 移除的无用功能”，引发大量社区反馈。

OpenAI Codex & ChatGPT lead Thibault Sottiaux said this week will deliver a level of product shipping previously expected only for DevDay 2025, while also crowdsourcing which Codex features are no longer useful.

FPV Ventures 合伙人 Nikunj Kothari 宣布领投 Piston 的 Series A 并加入董事会。美国每年数百亿美元燃油支出大多通过各类加油卡完成，但卡常与司机脱钩，易被盗刷且对账困难。Piston 直接砍掉卡，用自有支付通道让油站收款、车队后付，每笔交易绑定司机、时间、地点和油品，即将与车辆本身集成。支付量已增长 8 倍，商户网络增长 40 倍，留存率超过 98%，覆盖 2000 个油站和 48 个州。车队零费用即可获得防欺诈，油站获得直接车队关系。

FPV Ventures partner Nikunj Kothari is leading Piston’s Series A. The company eliminates fuel cards entirely, settling directly between stations and fleets with every transaction tied to a specific driver, time, location and fuel type. Payment volume is up 8×, merchant network 40×, retention above 98%, live at 2,000 stations across 48 states.

The Takeaway：真正的 AI 写作价值不在于一次生成完美文稿，而在于把每一次反馈永久写进系统，让下一次输出自动变好。

Every 员工作家 Katie Parrott 从两年前用 ChatGPT 当职业教练，到如今用 Codex 项目维护自己的 OKR、看板和优先级，再到发布 Compound Writing 插件。她把 Kieran Klaassen 的 Compound Engineering 思路移植到写作：头脑风暴、大纲、起草、实质审校、行文审校、最终定稿形成固定流程，并把 Kurt Vonnegut、Hitchcock、Sorkin、Sedaris 等作家的框架做成可调用的 skill。她强调“最后一公里”必须由人类补上——模型知识截止后的真实数据、独特洞察和个人经验才是让内容不商品化的关键。AI 把她从“组装句子的苦力”解放出来，重新找回探索和发现的乐趣。她的 AI 论点是：教育与获取机会将比以往任何时候都更重要，否则 AI 的复利只会集中在少数早期用户身上。

Katie Parrott, staff writer at Every, evolved from using ChatGPT as a career coach two years ago to a full Codex-based “career harness” that maintains OKRs and a Kanban board, and has now released the Compound Writing plugin. Inspired by Compound Engineering, it turns writing into a repeatable loop of brainstorm → outline → draft → substantive edit → line edit → polish, with reusable skills distilled from Vonnegut, Hitchcock and others. She stresses that humans must close the “last mile” with fresh data and lived experience the model cannot yet know. AI restored her love of writing by freeing energy for the hard questions instead of sentence assembly. Her thesis: education and access will matter more than ever, or the compounding gains of AI will stay with the early few.

Anthropic Engineering 介绍 Managed Agents：一个托管服务，通过稳定的 session、harness、sandbox 接口运行长时程 agent。早期把一切塞进同一容器导致“宠物”问题——容器挂了会话就丢、无法安全调试。解耦后 harness 以工具调用方式使用 sandbox，session 成为可查询的外部事件日志，TTFT 的 p50 下降约 60%、p95 下降超过 90%。安全上凭证永不进入 sandbox，通过代理或初始化时注入。Session 不再等于 Claude 的上下文窗口，而是可切片、可回放的持久对象，为未来不可预测的上下文工程留出空间。

Anthropic’s Managed Agents virtualize session, harness and sandbox so each can fail or be replaced independently. Moving the harness out of the container turned sandboxes into cattle, cut p50 TTFT ~60% and p95 >90%, and kept credentials unreachable from generated code. The session log becomes a durable, queryable context object outside the model’s window.

Claude Blog 宣布 Claude Code 可将工作进度捕获为 artifact——可实时更新的可分享网页，包括 PR walkthrough、系统解释器、仪表盘和发布清单。页面基于会话完整上下文（代码库、连接器、对话）自动生成，发布后同链接自动刷新并保留版本历史。默认仅作者可见，可分享给组织内成员，管理员通过合规 API 管控。已在 Team 与 Enterprise 组织 beta 开放。

Claude Code can now turn session work into live, shareable artifact pages—PR walkthroughs, dashboards, release checklists—that update in place from the full session context. Private by default, shareable inside the org, available in beta for Team and Enterprise.

Anthropic Claude Code 团队的 Boris Cherny 宣布 Claude Mods 开始落地，社区已有人做出在 Claude 里玩的 Tetris mod，并指向最新社区更新与技术细节。

Claude Code lead Boris Cherny announced Claude Mods are shipping; someone has already built a Tetris-in-Claude mod, with more demos and technical details in the linked issue.

Vercel CEO Guillermo Rauch 强调 agents 的上限取决于你提供的 proof-checker、编译器、类型系统和 linter。shadcn/lint 帮助 agents 遵守设计系统规则；他称“验证器 + skills 就是新的框架”。同时欢迎 Google Cloud Run 创始人 Steren 加入 Vercel 领导 Fluid 计算产品线，认为 agents 需要专为它们设计的新计算原语。他还分享了 fx 工具的自动升级与更长会话的性能提升。

Vercel CEO Guillermo Rauch argues agents are only as good as the proof-checkers, compilers, type systems and linters you give them; verifiers plus skills are the new frameworks. He also welcomed Steren (creator of Google Cloud Run) to lead Fluid compute products designed for the agent era.

Google Labs / Gemini 的 Josh Woodward 感谢两月前启动的 Gemini 重度用户群已测试 20+ 功能，并宣布新一批用户将提前体验 Daily Brief 与 Personal Intelligence 的后续能力，持续开放加入。

Google Labs VP Josh Woodward thanked the Gemini power-user group that tested 20+ early features and announced a new cohort getting early access to what’s next for Daily Brief and Personal Intelligence.

Anthropic Claude Code 团队的 Thariq 录制了 Latent Space 播客，技术向讨论模型能力狂奔下如何保持同步，以及他们怀念 AI 之前软件工程的哪些部分；另有一期与 Sid & Robert 聊 Claude Code 从零到一的变化。

Claude Code engineer Thariq recorded a technical Latent Space episode on keeping up with model capabilities and what the team misses about pre-AI software engineering, plus an earlier conversation with Sid and Robert on building Claude Code.

AI 教程作者 Peter Yang 表示语音交互让他在自然环境中散步时比盯着屏幕更高效，并转发了相关讨论。

AI educator Peter Yang notes that voice has made walks in nature more productive than staring at a screen.

FirstMark VC Matt Turck 提醒，尽管周末热闹，AI 进步不会放缓——玩家太多、经济激励太强、国内外都存在囚犯困境。他也称赞 Ramp 把账单做成百老汇音乐剧的品牌营销。

FirstMark’s Matt Turck notes AI progress will not slow: too many players, economic incentives and prisoner’s-dilemma dynamics both domestically and globally.

OpenClaw 与 OpenAI 相关的 Peter Steinberger 转发并表示某个功能非常有用，正在推动把它加入 Codex。

Peter Steinberger highlighted a useful feature and is pushing to add it to Codex as well.

SPC 普通合伙人 Aditya Agarwal 为所投公司的 demo 感到骄傲，称“AI 可以像人一样，这是终极目标”，并表示还有更多要建。

SPC GP Aditya Agarwal celebrated an early portfolio demo with the line “AI can be human. That is the ultimate goal.”