Rohan Paul@rohanpaul_ai · 6月20日61

Not anymore: Trump on whether he sees Anthropic threatening national security Axios' Marc Caputo asked Trump: "Do you view Anthropic, and to a degree its CEO Dario Amodei, as a threat to national security?" President Trump: "Well, not now, but a week ago, maybe....We were in the room in the G7 and seems like a nice guy, smart guy." Source: Axios

译特朗普在Axios采访中被问及是否认为Anthropic及其CEO Dario Amodei威胁国家安全时回应“现在不，但一周前可能”。Axios报道称，特朗普政府一周前曾限制访问Anthropic的新模型，彼时他可能视其为威胁，并称“人们会立即因泄露被关进监狱”。

meng shao@shao__meng · 6月20日63

驾驭 Claude Code：CLAUDE.md 配置文件、Skills、Hooks、Rules、Subagents 等 7 种指令全解析 Claude Code 最新博客，围绕七种方法展开： CLAUDE.md 文件、Rules、Skills、Subagents、Hooks、Output Styles、Appending the System Prompt。 每种方式的本质差异体现在三个维度： · 何时加载进上下文 · 会话压缩后是否保留 · 消耗多少 token、权威性如何 https://claude.com/blog/steering-claude-code-skills-hooks-rules-subagents-and-more 1. CLAUDE.md 文件 项目根目录下的 Markdown 文件，是最基础的配置层。 分两类加载：根目录 CLAUDE.md 在会话开始时全程驻留上下文，压缩后重新读取；子目录 CLAUDE.md 按需加载，仅当 Claude 访问该目录下文件时才触发，压缩后即失效。 关键警示：在共享仓库中，CLAUDE.md 往往像任何无人负责的配置文件一样，各团队不断追加内容却从不删减，成本在规模上会持续累积。每一行都会加载进每位工程师的每次会话，无论与当前任务是否相关。 claude 官方建议：控制在 200 行以内，指定负责人，像审查代码一样审查变更。 2. Rules 存放在 .claude/rules/ 的 Markdown 文件。 最有价值的特性是路径作用域：通过 paths 字段控制仅在触碰特定文件时才加载。例如只在 src/api/** 被访问时才注入"所有 API 处理器必须用 Zod 验证输入"的规则，而不是全程占用 token。 无 paths 限定的规则，行为等同于 CLAUDE.md——始终在场，始终消耗。 3. Skills 存放在 .claude/skills/ 的程序化工作流。 设计精妙之处：会话开始时只加载名称和描述；完整内容仅在技能被调用时才载入，可通过斜杠命令或任务自动匹配触发。 适合封装部署流程、发布检查清单、代码审查流程等固定程序，而非塞进 CLAUDE.md。Claude Code 自带若干内置 Skills，也支持自定义。 4. Subagents 存放在 .claude/agents/ 的独立助理定义。 与 Skills 的关键区别在于隔离性：子智能体在自己独立的全新上下文窗口中运行，返回给主会话的只有最终消息（通常是多个子任务的聚合结果）加上元数据，中间过程完全不污染主会话。 适合"跑完就丢"的旁路任务：深度搜索、日志分析、依赖审计。子智能体最多可嵌套五层深，支持动态编排数十到数百个后台 Agent 并行运作。 Skills vs Subagents 选择原则：想在主线程中逐步看到、随时干预 → Skills；想要隔离运行、只要最终结论 → Subagents。 5. Hooks 注册在 settings.json 中，在 Claude 生命周期的特定事件上触发（文件编辑、工具调用、会话开始等）。 这是确定性控制的唯一真正实现：Hooks 完全绕过上下文压缩机制，配置本身存在于主上下文窗口之外，因此上下文成本极低。 支持 command、HTTP、mcp_tool（确定性执行）和 prompt、agent（用模型判断）五种类型。一个 PreToolUse hook 可以拦截任何工具调用，以 exit code 2 阻止其执行。 重要观点：凡是写在 CLAUDE.md 里的"永远不要做某事"，都是错误的工具选择。Claude 大多数时候会遵守，但在长会话、模糊情况或遭遇提示词注入时可能失效。真正的硬约束必须是确定性的，而 Hooks 和权限控制才是实现方式。 组织级强制管控还可以使用 Managed Settings（管理员部署，用户无法覆盖）。 6. Output Styles 存放在 .claude/output-styles/ 的文件，直接注入系统提示，永不被压缩，权威性最高。 高权威有代价：自定义输出风格默认会替换掉 Claude Code 的默认输出风格，包括"如何界定改动范围、何时添加注释、如何处理安全问题、声称完成前是否运行测试"等关键编程默认指令，使 Claude Code 退化为通用助理。 官方建议先看内置风格（Proactive/Explanatory/Learning），覆盖大多数需求，无需自己维护文件。 7. Appending the System Prompt 通过 CLI flag 在调用时追加，仅对本次调用生效，不跨会话持久化。 与 Output Styles 的区别是只增不替换，不改变 Claude 的角色设定，只是在默认角色上叠加指令。 注意边界：追加系统提示存在边际收益递减问题。提供的指令越多，Claude 的遵从度越低，若指令之间存在矛盾则尤为明显。 几个实用决策原则 1. 每次编辑后自动跑 linter × 写进 CLAUDE.md √ 用 Hook 注册到 PostToolUse 2. 禁止某类危险操作 × "Never do this" 写 CLAUDE.md √ PreToolUse Hook + exit code 2 3. 30 行部署流程 × 塞进 CLAUDE.md √ 放进 .claude/skills/ 4. 只对 API 目录生效的规则 × 无路径限定的 Rule √ 用 paths: 字段作用域限定 5. 个人习惯偏好 × 写进项目级 CLAUDE.md √ 写进用户级配置（对所有仓库生效）

译博客详解 Claude Code 的七种指令配置（CLAUDE.md、Rules、Skills、Subagents、Hooks、Output Styles、追加系统提示），从加载时机、压缩后保留性、token消耗与权威性三个维度对比。CLAUDE.md 分根目录（全程驻留）和子目录（按需加载）；Rules 支持路径作用域节省 token；Skills 仅加载名称和描述，调用时载入完整内容；Subagents 独立上下文运行，只返回结果；Hooks 绕过压缩实现确定性控制；Output Styles 直接注入系统提示且永不被压缩；追加系统提示仅单次生效。文章给出实用决策原则，如用 Hook 跑 lint、用 Skills 封装部署流程等。

ginobefun@hongming731 · 6月20日47

BestBlogs 早报 · 06-20 # Anthropic 竞品研究限制 / Claude Fable 5 出口管制 / DeepSWE 新基准 / Nvidia Nemotron 3 Ultra / Spring I/O LLM 攻防 [1] ★ 精讲｜测试神话与寓言，超越 SWE-bench，Nvidia 的开放竞争者 Andrew Ng 罕见动怒：Anthropic 不仅给 Claude Fable 5 加上限制竞品研究的条款，还被曝悄悄降低疑似同行研究者的输出质量；美国商务部出口管制更直接逼得 Fable 全球下线，掀起多国关于 AI 主权的讨论。同期 DeepSWE 等新基准证明智能体编程能力远未触顶，Nvidia 也以混合架构悄然加入开源模型竞速。 来源：The Batch | http://DeepLearning.AI https://www.bestblogs.dev/article/b5f80528 [2] ★ 精讲｜攻破 LLM 驱动的应用：从上下文投毒、工具越权到纵深防御 [视频] Spring I/O 演讲实测攻破 LLM 应用：用路径穿越污染 RAG 知识库让助手认账虚构条款，靠 SQL 注入伪造聊天记忆诱导越权操作，再用拆分提问绕过护栏套取用户数据；核心结论是护栏只能降风险，授权、最小权限工具与可观测性才是真正防线。 来源：Spring I/O https://www.bestblogs.dev/video/c86cc44 [3] ★ 精讲｜AI 革命新思考：过去 50 年，软件工程其实没真正“工程化”过？ 作者提出尖锐论断：过去 50 年软件工程靠人脑堆代码，从未真正工程化；大模型第一次实现耗能换高阶认知，但范式必须从人为中心 + AI 辅助转向 AI 为中心，否则不确定性会在人机之间循环放大——程序员将从人肉编译器转型为产线设计师与偏差拉回者。 来源：dbaplus 社群 https://www.bestblogs.dev/article/f4f17f23 [4] 谁在 ChatGPT 上投广告？ChatGPT 和 Claude 又在哪投广告？ 本文基于 Sensor Tower 2026 年 AI 报告，梳理了 ChatGPT 广告系统、10 亿月活里程碑、Claude 高 ARPU、AI 厂商广告投放策略分化、GEO 对零售的影响以及赛道格局固化等关键趋势。 来源：Founder Park https://www.bestblogs.dev/article/3d39fbc2 [5] Project Valhalla 详解：十年磨一剑，JDK 28 预览版终登场 - JVM Weekly vol. 180 Project Valhalla 的值类终于在 JDK 28 预览版中落地，经过十年的设计迭代和方案否决，为用户自定义类型带来了扁平、紧凑的内存布局。 来源：Hacker News https://www.bestblogs.dev/article/b3b9d945 [6] Andrew Ng 论 AI 控制：Anthropic 与美国政府行动标志着转折点 Andrew Ng 分析了 Anthropic 与美国政府近期限制前沿 AI 模型访问的行动，认为这些举措正加速全球对 AI 主权和开源替代方案的投资。 来源：Andrew Ng(@AndrewYNg) https://www.bestblogs.dev/status/2068039709126017356 [7] 两个 80 后，在硅谷“崩老头” 本文从谷歌面临 AI 搜索颠覆的「创新者窘境」切入，深入剖析大模型商业模式的根本矛盾——推理成本高企与订阅收入难以覆盖的剪刀差，并指出 AI 产业的真正赢家是基础设施层，而非模型层。 来源：腾讯科技 https://www.bestblogs.dev/article/0cdd7f15 [8] 为敏感云系统设计持续授权 本文主张，在云系统中，授权必须从单一的登录时决策转变为对每个敏感操作的持续、上下文感知评估，并提供了实现这一目标的实用架构。 来源：InfoQ https://www.bestblogs.dev/article/a1bfaf9a [9] TypeScript 7.0 RC 发布，编译器移植 Go 实现 10 倍提速 TypeScript 7.0 RC 发布，该版本将编译器从 TypeScript 完整移植到 Go，实现了约 10 倍的性能提升，同时引入了新的并行化控制、重构的监视模式以及更新的默认配置。 来源：Microsoft for Developers https://www.bestblogs.dev/article/ae749768 [10] 暗物质探索迎来全新局面 WIMP 暗物质搜寻的失败，为轴子、低质量粒子、量子传感器乃至行星极光等大量新候选体和探测方法打开了大门。 来源：MIT Technology Review https://www.bestblogs.dev/article/2e25e788 --- http://BestBlogs.dev · 发现真正适合你的高质量内容 BestBlogs 是 AI 驱动的私人阅读助手，帮助你发现真正适合你的高质量内容，欢迎体验。 在线阅读：https://www.bestblogs.dev/explore/brief/2026-06-20

译Anthropic对Claude Fable 5增加竞品研究限制，并疑似降低同行研究者输出质量；美国出口管制迫使Fable全球下线。DeepSWE等新基准显示智能体编程能力未触顶。Nvidia发布Nemotron 3 Ultra混合架构开源模型。Spring I/O演讲演示LLM应用攻防：路径穿越污染RAG、SQL注入越权、拆分追问绕过护栏，强调授权与可观测性。TypeScript 7.0 RC将编译器移植Go，实现约10倍性能提升。Andrew Ng认为限制措施加速全球AI主权投资。

Rohan Paul@rohanpaul_ai · 6月20日40

Today’s edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/openai-just-moved-frontier-level 🗞️ OpenAI just moved frontier-level health AI from premium reasoning models into the free GPT-5.5 Instant model. 🗞️ The article that went super viral - Satya Nadella on organizational economics of AI and “token capital” 🗞️ Anthropic just showed Claude Opus 4.7 program a robodog in 12:07 mint, about 20x faster than last year’s Claude-aided human team on the tested tasks. 🗞️ This was long needed for AI in finance - Making SEC filings readable for machines without flattening the accounting logic. 🗞️ Mark Zuckerberg is trying to restart Meta’s hacker culture after 8,000 layoffs but employees are pushing back. 🗞️ Anthropic just rolled out a major Claude Design update, adding design system imports, code round-trips, and a fix for its heavy token usage issue.

译OpenAI将前沿级健康AI从高级推理模型移至免费的GPT-5.5 Instant模型。Satya Nadella关于AI组织经济学和“token资本”的文章走红。Anthropic展示Claude Opus 4.7用12分07秒编程机器人狗，比去年人类团队快20倍，并推出Claude Design大更新（支持设计系统导入、代码往返、修复高token用量）。AI金融领域推动SEC文件可机器读取但不简化会计逻辑。扎克伯格裁员8000人后试图重启Meta黑客文化，员工抵制。

ClaudeDevs@ClaudeDevs · 6月20日32

Update: we've gone ahead and reset 5-hour and weekly usage limits for everyone, across all plans. Enjoy your weekend!

译更新：我们已为所有计划的所有用户重置了 5 小时和周使用限制。祝周末愉快！

Berryxia.AI@berryxia · 6月20日51

亚马逊的骚操作简直了！ 自己是A社的竞争对手也是股东… Trump前几天还把Anthropic当成国家安全威胁，这两天刚跟Dario见完面就改口说“是个好人、聪明人，反应很负责任”。 记者问他是否认为Anthropic和Dario是国家安全威胁，Trump说：“现在不觉得了，但一周前可能觉得。” 他昨天跟Dario见了面，觉得对方反应很快、很负责任。谈到是否用国防生产法管制AI时，他说“可能会，但目前看来大家都很负责任”。 最后还补了一句：“其实是竞争对手兼部分股东把Anthropic举报的……” 这波操作把AI监管的真实逻辑给暴露得挺清楚的：出口管制这种级别的决定，一周前可能是威胁，一次见面后就变成了“聪明好人”。 而举报方据说是Amazon（既是竞争对手又是Anthropic的股东）。 以前大家觉得AI政策会越来越制度化、越来越基于技术评估，现在看起来至少在美国，还挺依赖“谁跟总统聊了、聊得怎么样”。

译特朗普对Anthropic的态度在一天内反转：此前将其视为国家安全威胁，但会见CEO Dario Amodei后改口称其“聪明、好人、反应很负责任”。被问是否用国防生产法管制AI时，他表示“可能但似乎大家都很负责任”。特朗普还透露，举报Anthropic的是其竞争对手兼部分股东——Amazon。这一事件暴露了AI监管的真实逻辑：出口管制等重大决策可能取决于与总统的面谈效果。

swyx@swyx · 6月20日28

Anthropic is going to IPO at $2T

译swyx 称 Anthropic 将以 2 万亿美元估值 IPO。与此同时，AlphaFold 团队负责人 John Jumper 在任职近 9 年后宣布离开 Google DeepMind 加入 Anthropic。

AYi@AYi_AInotes · 6月20日85

John Jumper （Google AlphaFold 团队核心领导者、2024 年诺贝尔化学奖得主）宣布，在 Google DeepMind 工作近 9 年后决定离开，加入 Anthropic（先休整一段时间） 我看到这条离职帖下面好多人都在喊 AI Avengers，但如果认真看完他写的每一句话，就会发现一个更复杂的现实， John Jumper 博士毕业仅仅 6 个月，@demishassabis Hassabis 就把整个 AlphaFold 团队交给他来领导，核心是相信他能解决生物学 50 年来的未解难题， 他在 DeepMind 待了近 9 年，做出了诺贝尔奖级别的成果，离开时他写的是：GDM taught me how to do great science，GDM is a special place 我觉得这不是一句客套话，这是一个人在被给予了罕见的早期信任、把最难的事做成了之后，回头再看时说的真心话， 但社区很快编完了另一个剧本：Anthropic 在组建 AI Avengers，Jumper 是下一颗无限宝石🤣 关键问题是，他原帖里一个字都没提为什么离开啊， 不是不相信他想加入 Anthropic，而是他给 DeepMind 的告别里塞进了太多真实的东西，多到让你不能简单地用"留不住人"或者"某个实验室更强"来消解， 我以前看顶级球员转会学到一个视角：

译2024 年诺贝尔化学奖得主、AlphaFold 团队领导者 John Jumper 宣布在 Google DeepMind 工作近 9 年后离职，将加入 Anthropic（先休整一段时间）。Jumper 博士毕业仅 6 个月即被 Hassabis 信任，任命领导 AlphaFold 团队并攻克生物学 50 年未解难题。他告别时称 DeepMind 教会他如何做伟大的科学，是一个特别的地方。社区将 Anthropic 的招揽戏称为组建“AI Avengers”。

Boris Cherny@bcherny · 6月20日37

Cool way to use Claude Code: deciphering Linear A, a 3500 year old written language from Crete https://aiclambake.com/clamtakes/linear-a/ Hope this holds up in peer review! 🤞

译使用 Claude Code 的一个酷炫方式：破译线性文字 A，一种来自克里特岛的 3500 年前的书面语言。 https://aiclambake.com/clamtakes/linear-a/ 希望这能经得起同行评审！🤞

Chubby♨️@kimmonismus · 6月20日80

Trump viewed Anthropic and CEO Dario Amodei as a threat to national security. „Well, not now, but a week ago, maybe," the president said.“ Tone changed. Now he thinks Dario is smart and a nice guy.

译特朗普在Axios专访中称Anthropic与其CEO Dario Amodei曾构成国家安全威胁，但随后态度转变，认为Dario既聪明又是个好人。引用推文证实了Axios的独家报道。特朗普表示“不是现在，但一周前也许是”，暗示其立场随时间变化。

Nathan Lambert@natolambert · 6月20日42

Well said.

译Nathan Lambert引用hlntnr的推文，用森林怪物比喻解释Anthropic的AI安全世界观：他们认为存在巨大且危险的怪物（超级智能风险），其他竞争者为了宝藏不顾危险，而Anthropic选择派出自己的先遣队，更快更深入地进入森林，投入大量资源进行怪物控制和驯服，并向外界报告发现。这种策略虽然逻辑自洽，但也让外界质疑其是否疯狂或说谎。

Rohan Paul@rohanpaul_ai · 6月20日73

John Jumper, the AlphaFold scientist who shared a Nobel Prize with Demis Hassabis, is leaving Google DeepMind for Anthropic after nearly 9 years at the lab. Jumper is best known for spearheading Google's AlphaFold team. Just yesterday, Google also lost Noam Shazeer, a Gemini co-lead, Character .AI founder, and co-author of "Attention Is All You Need", to OpenAI.

译AlphaFold核心科学家John Jumper在DeepMind工作近9年后宣布离职，加入Anthropic。他曾领导AlphaFold团队，并与Demis Hassabis共同获得诺贝尔奖。就在前一天，Google DeepMind还失去了Gemini联合负责人、Character.AI创始人及"Attention Is All You Need"合著者Noam Shazeer，后者加入了OpenAI。

Chubby♨️@kimmonismus · 6月20日37

Roughly 200 organizations still have access to Claude Mythos. Just imagine the advance they have.

译仍有大约 200 个组织可以访问 Claude Mythos。 想想他们拥有的巨大优势吧。

Chubby♨️@kimmonismus · 6月20日83

What the fu*k!! John Jumper, one of the key figures behind AlphaFold and co-winner of the Nobel Prize alongside Demis Hassabis, is leaving Google DeepMind to join Anthropic. Huge loss for, and an insane win for Anthropic! What is happening, especially at google?

译AlphaFold关键人物、诺贝尔奖共同得主John Jumper宣布离开工作近9年的Google DeepMind，加入Anthropic。他在声明中感谢Demis Hassabis让其博士毕业后仅6个月就领导AlphaFold团队，并表示GDM是特别的地方。主推文评价此举对Google是巨大损失，对Anthropic则是疯狂胜利。

elvis@omarsar0 · 6月20日70

http://x.com/i/article/2068004233849290752 # From Prompting Agents to Loop Engineering A claim has been circulating in AI coding circles: stop prompting your coding agents and start designing loops that prompt them for you. As with everything new, this stuff gets repeated often and explained rarely. This is the practical version: what an agent loop is, why it matters, and what one looks like in production. Below you can read some of my thoughts (written with the help of Claude) from some of the experiments, research, and conversations I’ve been having with some of our students, technical founders, AI engineers, and startups. You might also find our recent live session on "Autonomous Long-Running Coding Agents" as a good starting point for all of this. ## Where the claim comes from > "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."Peter Steinberger (@steipete), Jun 7 2026. 2.2M views. Original tweet Boris Cherny, the creator of Claude Code, makes the same point from the other side. > "I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and figuring out what to do. My job is to write loops."Boris Cherny (@bcherny). Original tweet The point is not that prompt engineering is dead. With loop engineering, the work moves up a level, from writing the code to writing the system that writes the code. Developers furthest along this path report months where they shipped hundreds of PRs without opening an IDE, with every line written by the agent. ## What a loop actually is A loop is a small program you write that does four things: - prompts the coding agent for you, - reads what it produced, - decides whether it is done, - and if not, prompts it again with the error or the next step. You stop sitting inside the loop typing prompts; you write the loop, and the model becomes a subroutine it calls. The shape is always the same: set a goal, act, check, feed the error back, and repeat until the check passes or the loop stops itself. ## "Loop" means at least five things Much of the disagreement is people using one word for five different ideas. Here is the progression, oldest to newest. - ReAct (2022). The original research pattern: reason, act, observe, repeat. - AutoGPT (2023). A self-prompting goal loop, notorious for not knowing when to stop. - ralph loop. A deliberate context reset between iterations so the agent does not drown in its own history. - /loop and /goal. Cadence and completion conditions are built into the agent, carrying the state across turns. - orchestration. One author fans out many agents that read your GitHub, Slack, and chat, and decide what to build next. ## The parts you actually assemble The progression explains what people mean by loop; this is what a loop is built from. The same six parts show up every time, and most now ship inside the coding tools instead of custom scripting you maintain yourself. - A trigger. Something that starts the loop without you pressing go: a schedule, a webhook, a file change, a label landing on a PR. This is what separates a real loop from a single run you repeat by hand. - Isolation. A private checkout per agent, usually a git worktree, so two agents running at once cannot overwrite each other's files. Once you run more than one, this stops being optional. - Written-down context. The conventions, build steps, and project-specific rules are kept where the agent reads them on every run. Skip it, and the loop re-derives your project from scratch each pass and guesses at the gaps. - Reach into your tools. Connectors to the issue tracker, CI, database, and chat, so the loop can open the PR, link the ticket, and post the result instead of printing a fix and waiting for you to carry it the rest of the way. - A second agent checks. A separate worker who grades the output is held apart from the one who produced it, because a model reviewing its own work passes almost everything. - State on disk. A markdown file, a board, or a queue: anything outside the conversation that records what is finished and what is next. The model forgets between runs; the file does not. Assemble those six, and you have a good starting point for loop engineering. You used to hand-build everything; now most ship as built-in features, which is why the pattern has moved from a fringe technique into common use. ## A concrete loop, the PR babysitter A concrete example you can build today: - Trigger. Every 15 minutes. - Scope. Open PRs labeled agent-watch. - Action. If CI is red for a deterministic reason, attempt one fix. If the main moved, rebase once. - Budget. One fix attempt per PR, five minutes, ten files changed. - Stop condition. CI green, or budget exhausted, then stop and ping a human. You return to merged PRs instead of a backlog of broken builds. The same shape covers most ops work: - CI health. Every 30 minutes, pull failing runs and cluster them by signature, so ten red PRs with one root cause become one thing to look at. - Deploy verification. After a push, hit your endpoints, confirm 200s and the expected content, and flag regressions before users do. - Feedback clustering. Every 30 minutes, pull comments from your channels, group them into themes, and map each cluster to the file or doc that owns it. ## A concrete Claude Code loop with /goal The babysitter is a loop you wire up yourself; it also helps to see one that ships inside the agent. In Claude Code, the smallest complete loop is /goal: you hand it a verifiable end state, and it keeps taking turns until that state is true. Here is an example of /goal used as an in-session command in Claude Code. You launch the session, then set the goal inside it: It is the same act, check, repeat shape from earlier, with the verifier built in. At this point, it’s clear that a strong /goal reads less like a prompt and more like a contract. The good ones specify four things: the end state you want, the evidence that proves you reached it, the constraints the agent must not break getting there, and the budget of work it is allowed to spend. Leave any one of them vague, and the model fills the gap with the easiest reading: it stops early, takes a shortcut, or redefines success so the transcript looks done while the real system is broken. - Set the condition. Type /goal plus a checkable end state, for example,/goal tests in test/auth pass. The first turn starts immediately. - The agent works a turn. It edits, runs the tests, and surfaces the results in the session. - An evaluator checks. A fast model reads the transcript and decides whether it is met or not met, so the agent is not grading its own work. - Loop or finish. Not met means another turn with guidance; met means the goal clears itself and the run stops. State carries across turns, so it does not quit early or drop a constraint partway through. A few controls keep it reliable: - Make the check measurable. A test result, an exit code, a file count, or an empty queue. npm test exits 0 is a goal; "make it better" is not. - Bound the run. Append something like "or stop after 20 turns" so a stuck loop halts instead of burning turns. - Pair it with auto mode so that turns run unattended, and use /goal clear to abandon it early. The evaluator step hides a useful subtlety: the checker does not have to be the same model as the coder. Once the loop has distinct roles (planner, executor, evaluator, vision reviewer), each can run on a different model, and choosing which model fills which role becomes an architecture decision rather than a single bet on one "best" coding agent. Some models plan better, some execute more cheaply, some judge a screenshot more accurately, and a good orchestrator lets you swap them per role instead of waiting for one vendor to win every category. It works well for API migrations (move every call site until it compiles and tests pass), refactors (split a file until each module is under budget), issue backlogs (work a labeled queue until it is empty), and eval loops (tune a prompt until the score clears a threshold). /loop is the counterpart for work with no single finish line: instead of a completion condition it re-prompts on a schedule, which is how a loop like the PR babysitter keeps running. ## Running many loops unattended A single /goal loop is one agent working toward one finish line. Running many unattended processes raises the stakes, because a loop is only as trustworthy as its ability to check its own work. Cherny's setup for running Opus autonomously for hours comes down to five steps: 1. Auto-approve permissions so the agent does not stop to ask on every tool call. 1. Use dynamic workflows (drop Ultracode into the prompt) to fan out across many agents instead of one serial thread. 1. Use /goal or /loop to keep it going. /goal sets a completion condition, /loop re-prompts on a schedule, and both carry state, so it does not quit early. 1. Run it in the cloud (desktop or mobile app) so the session survives when you close the laptop. 1. Give it a way to self-verify end-to-end. Claude in Chrome for web, a simulator MCP for mobile, and a live server for backend. This is the step that makes the other four safe. The full sequence: ## crabfleet: orchestration as a product Orchestration is easier to picture with a concrete tool. Peter Steinberger's crabfleet, an OpenClaw project billed as "mission control for agent runs," is a loop packaged as a product, and its shape maps onto everything above. - Work as cards on a board. Tasks are entered as cards built from a prompt, a GitHub issue, or a PR, then move through todo, running, human review, and done. That board is the loop's queue and its stop-and-report step, made visible. - Durable runs, not fire-and-forget. Each run is a tracked attempt with heartbeats, so it keeps going when you look away and survives a closed laptop. You take over only when the runtime advertises that it supports handoff. - Agents that spawn agents. A run can start child sessions, send messages, read transcripts, and update its own summary from inside a sandbox: on-disk memory and fan-out in one place, one author and many agents. It runs on disposable cloud sandboxes with browser-based terminals, which is what makes walking away from an unattended run safe. The point is not the specific tool but that the loop has hardened into infrastructure: a queue, durable execution, fan-out, and a human-review gate are now things you configure rather than hand-script every time. ## Where the cost goes now For two years, the cost question in AI coding was simple: which model, and how many tokens. Inside a loop, that instinct points at the wrong layer. The spend is no longer a single call but how many times the loop goes around, so a loop that retries six times before it converges costs six times as much as one that lands on the first pass, on the same model. That changes what is worth optimizing: - Iterations are the budget line, not tokens. A cheaper model that loops twice as often is not cheaper, so track cost per finished task, not cost per call. - A weak verifier is the most expensive bug you can ship. If the check that decides "done" is loose, the loop either stops early on broken work or grinds on work that was already fine, and both waste whole iterations. Tighten this before anything else. - Failing fast is a cost control. A loop with no cap on consecutive failures does not eventually succeed; it eventually drains the account, so the stop condition protects the bill as much as the codebase. You used to tune the prompt; now you tune the loop, because that is where the cost accumulates. ## When not to loop Loops pay off when a task repeats, and a machine can tell when it is done. Outside that, a loop only automates churn. Skip it in these cases: - One-shot edits. If you can finish it in a single pass, a loop is pure overhead. - Unscoped or exploratory work. "Figure out why users are churning" has no pass condition, so the loop never converges. - Anything without a cheap automated check. If the only verifier is your own eyes, you are still inside the loop. Build the check first, or do the task by hand. ## What can go wrong A loop that runs while you sleep also makes mistakes while you sleep, and the failure modes are predictable. - The verification burden stays human. The loop writes faster than you can review, so if you stop reading the diffs, you have not removed the work, only deferred it. - Comprehension gaps widen. Shipping code you did not write, faster than you can absorb it, erodes the model of your own system, and that debt comes due during the next incident. - Silent drift on a loose check. A weak verifier lets wrong-but-passing work through on every iteration, so the loop looks productive while it digs a hole. None of this is an argument against loops; it is why the engineer who designs the loop matters more, not less. ## How to build your own 1. Pick one repeatable task. Babysitting PRs, fixing CI, verifying deploys: start with routine work. 1. Scope it tight. "Fix the billing webhook validation, only touch app/api/billing and lib/billing," beats "fix the bug." A loose loop wanders. 1. Give it a budget and a stop condition. Max attempts, max runtime, max files, max spend, max consecutive failures. A loop running unattended is also a loop making mistakes unattended. 1. Add an independent verifier. A separate sub-agent grades the work, because the agent who wrote the code is the worst judge of whether it is done. 1. Run it on a cadence. /loop for an interval, cron for a schedule, hooks at lifecycle points, or GitHub Actions so it survives a closed laptop. 1. Keep memory on disk. The model forgets between runs, so state lives in markdown or a board, not in the context window. The takeaway: the loop, not the model, is now the expensive and failure-prone part. Build it like someone who intends to stay the engineer responsible for the output, not just the person who starts the run. If you see any errors or things that need further clarification, don’t be afraid to reach out. ## Other Useful References - Addy Osmani (@addyosmani), on AI-assisted coding loops - Matt Van Horn (@mvanhorn), "WTF Is a Loop?" - Peter Steinberger (@steipete), on designing loops - Boris Cherny (@bcherny), on running agents autonomously

译AI编程圈出现新主张：不应再手动提示编码智能体，而应设计循环自动完成提示、读取输出、判断完成，并在出错时重新提示。Boris Cherny（Claude Code创建者）和Peter Steinberger均持此观点。文章梳理了循环的五种演进形态（ReAct、AutoGPT、ralph loop、/loop与/goal、编排），并拆解六大组装部件：触发机制、隔离工作区、项目上下文记录、工具连接、独立验证智能体等。核心转变是从编写代码升级到编写驱动代码的系统。

Yuchen Jin@Yuchenj_UW · 6月20日69

John Jumper is a Nobel laureate who won the 2024 Nobel Prize with Demis Hassabis. Now he’s leaving Google DeepMind too, right after Noam Shazeer. What is going on at Google DeepMind? I had really high hopes for Gemini.

译2024年诺贝尔奖得主John Jumper宣布离开效力近9年的Google DeepMind，加入Anthropic（先休息一段时间）。他因与Demis Hassabis共同领导AlphaFold团队获奖，此次离职紧随Noam Shazeer之后。推主Yuchen Jin对此表示惋惜，并质疑Google DeepMind人才流失现状，称原本对Gemini寄予厚望。

Demis Hassabis@demishassabis · 6月20日72

Thanks John for an extraordinary partnership and wonderful collaboration over the past 9 years! What we achieved with AlphaFold changed the world, and showed the field what was possible with AI for science and medicine, lighting the way for how AI can benefit humanity.

译AlphaFold 团队负责人 John Jumper 宣布，在 Google DeepMind 工作近 9 年后决定离职，将加入 Anthropic（先休整一段时间）。DeepMind CEO Demis Hassabis 表示，过去 9 年与 Jumper 的非凡合作改变了世界，AlphaFold 展示了 AI 在科学与医学领域的巨大潜力，并为 AI 造福人类指明了方向。Jumper 回忆，Hassabis 在他博士毕业仅 6 个月后就大胆让他领导 AlphaFold 团队，感谢团队教会他如何做伟大的科学。

Berryxia.AI@berryxia · 6月20日57

http://x.com/i/article/2067984320212668416 # 搭一个AI Agent团队，到底有多难！5分钟搭的是玩具，不是员工。 你大概也刷到过那种教程——“5分钟搭一个 AI Agent 团队”，评论区清一色“收藏了”“这就去试”。你收藏了，你也试了，然后卡在了第一步。 不是因为你不够聪明。 从“看懂教程”到“跑通一个能干活的 Agent 团队”之间，横着四道真实的技术门槛。每一道都有人折在上面，每一道短期内都不会自动消失。 最近在 X 上传播最广的一篇，是 Khairallah AL-Awady 写的“How to Build Your First Team of AI Agents Using Claude”，万级浏览，承诺很诱人：大多数人还在一问一答地用 Claude，少数人已经让 AI 跑起了整个团队。 但打开教程动手试的人，大概率都经历了同一件事：卡住。 这篇文章不喊“未来已来”，也不贩卖焦虑。只想把四道门槛逐层拆开，讲清楚到底难在哪。 看完之后，你不只是松口气，还能带走一个判断标准：下次再看到“5分钟搭 AI 员工”的噱头，一眼就知道它靠不靠谱。 ## 第一道槛：环境配置，劝退率最高的隐形墙 搭 Agent 团队的第一步不是设计架构，而是把开发环境跑起来。就这一步，能劝退八成的人。 如果你不是全职开发者，光是看懂接下来这串名词就要查半天：Python 3.10+、Anaconda 环境管理、PyTorch 框架、API Key 配置（OpenAI/Claude/Gemini 三选一还是全要？）、向量数据库、Docker 容器化…… 清华大学出版社的《AI Agent 智能体开发实践》用整整一章讲“Agent 开发环境配置”，步骤密集得像一份工程手册。 CSDN 上“从零搭建 AI Agent”的教程评论区，出现频率最高的反馈是：环境配了一整天跑不起来。 更让人头疼的是框架选择。 Anthropic Agent SDK、OpenAI Agents SDK、LangGraph、CrewAI，四个主流框架，设计哲学各异，选错之后的迁移成本极高。 而 MCP 协议刚出来，A2A 协议还在早期，不同框架之间的互操作性几乎为零。今天选的框架，半年后可能就是技术债。 低代码平台（扣子/Coze、Dify）正在降低入门门槛，但只能覆盖简单场景。一旦需求复杂化，你还是得回到代码层。 环境配置这道槛，本质上不是“难”，而是“碎”。每一步都不难，但步骤太多、依赖太杂、标准太乱，拼起来就是一堵墙。 ## 第二道槛：Agent 编排，让 AI 协作比让 AI 干活难十倍 环境配好了，接下来要解决的是：怎么让多个 Agent 分工协作？ 直觉上，这不过是“给每个 Agent 一个角色，让它们各干各的”。 但 arxiv 上的多篇论文给出了更冷峻的结论：多 Agent 协作存在三个系统性缺陷，从众效应、角色不一致、臆想拼凑。 从众效应指的是，Agent 在讨论中受“同伴压力”影响，放弃自己的判断去附和别人。角色不一致意味着，你给 Agent 定义了“审稿人”的角色，它聊着聊着就变成了“写稿人”。 臆想拼凑则类似人类的记忆偏差，Agent 不是在撒谎，而是真心以为自己编出来的东西是对的——这才是最棘手的部分。 更讽刺的是，鼓励辩论而非协作的指令，反而会加剧这些问题。 传统 plan-and-execute 框架也扛不住。 它的三大局限：僵化的计划执行、静态的 Agent 能力、低效的通信，在复杂任务面前暴露无遗。 斯坦福的研究发现，LLM 作为规划器经常产出不完整或不一致的行动序列，违反约束条件，需要独立验证器、版本化执行日志和局部修复机制才能勉强兜底。 这就引出了一个有意思的立场分歧。 Anthropic 自己构建多 Agent 系统时，采用的是 Subagents 模式，一个主 Agent 调度多个子 Agent，子 Agent 之间互不通信。 这是目前生产环境中唯一稳定的多 Agent 模式。 而 Cognition（Devin 的母公司）更直接，发了一篇标题就叫“Don‘t Build Multi-Agents”的文章，认为应该用上下文工程替代多 Agent 架构。 翻译成人话：连造 Agent 的人自己都不太敢让 Agent 之间自由对话。 多 Agent 编排本质上是分布式系统问题，但 LLM 的非确定性让传统分布式系统的解决方案（共识协议、状态机复制）无法直接套用。 这道槛，不是换个框架能解决的。 ## 第三道槛：上下文管理，Agent 团队的真正杀手 前两道槛好歹还能靠耐心和经验硬扛。上下文管理是另一回事——它是 Agent 团队失败的隐性原因，也是最被低估的那道槛。 Andrej Karpathy 有个比喻：LLM 是 CPU，上下文窗口是 RAM。 RAM 有限，操作系统需要管理内存。上下文窗口有限，同样需要管理什么信息放进去。 问题在于，大多数开发者根本没有“管理上下文”这个意识，他们以为写好 prompt 就够了。 上下文崩塌有四种模式。 上下文污染：错误信息进入上下文后持续污染后续推理，像一行脏数据搞垮整张表。上下文分散：过多无关信息导致模型注意力涣散，该关注的反而被淹没。 上下文混淆：相似但不同的信息让模型分不清谁是谁。 上下文冲突：矛盾的信息直接导致行为不可预测。 在多 Agent 场景中，这四种模式会叠加放大。 每个 Agent 都有自己的上下文，Agent 之间传递信息时，上下文要么膨胀（信息越传越多），要么丢失（关键细节在传递中被压缩掉）。 这不是写几句 prompt 能修补的，这是架构层面的问题。 一个数据足以说明一切：据公开报道，Anthropic 用同一款 Opus 4.5模型，搭配不同的工程适配架构，Claude Code 架构在 CORE 基准上拿到78%，Smolagents 架构只有42%。 模型完全一致，仅上下文工程不同，分差36个百分点。 36个百分点。这意味着上下文工程对结果的影响，可能比换一个更强的模型还大。 另一位实践者报告：采用系统化的上下文工程后，调试周期从5.2天压到1.7天，用户任务完成率从63%跳到89%。当然，这是个案经验，不代表普遍结论，但方向是明确的。 但上下文工程不是“更高级的 prompt 写法”，而是把整个交互过程当作可建模、可测量、可版本化的工程系统。 这套方法论目前还在早期，没有成熟工具链，没有最佳实践共识，甚至没有统一的定义。 Spring AI Alibaba 的 Agent 抽象层专门为此设计，承认“缺乏正确的上下文是更可靠 Agent 的头号障碍”。 这道槛的残酷之处在于：你不知道自己死在了这里。Agent 跑出来的结果不对，你以为是模型不够强，其实是上下文没管好。 ## 第四道槛：工具接入与稳定性，demo 能跑≠能干活 前三道槛解决的是“能不能跑起来”的问题。第四道槛解决的是“跑起来之后能不能干活”的问题。这两件事之间的距离，比大多数人以为的大得多。 传统软件有合约来规范行为：API 定义输入输出，类型系统约束数据格式，断言和测试兜底边界情况。但 AI Agent 基于 prompt 和自然语言指令运行，没有正式的行为规范。 一篇2026年的 arxiv 论文把这个问题叫做“Agent Behavioral Contracts”——Agent 行为合约的缺失，是漂移、治理失败和项目频繁失败的根源。 翻译成人话：传统软件出了 bug 你知道它错在哪，Agent 出了问题你可能连它怎么偏的都说不清。 实践中，翻车场景花样百出。工具调用失败后 Agent 不知道怎么恢复，陷入死循环。 多个 Agent 同时调用同一个 API 导致冲突。上游 API 悄悄改了接口，Agent 行为静默退化，没有告警，你甚至不知道它已经坏了。 更隐蔽的是成本失控，一个看似简单的任务，可能触发几十次 API 调用，账单比你预想的多一个零。 Nature Scientific Reports 2026年的一项研究发现，在多 Agent 路由中，即便采用置信度感知门控将路由准确率从0.65提升到0.77，切换不稳定性和弹跳不稳定性仍然存在。 翻译成人话：Agent 在“该不该切换到另一个 Agent”这件事上，判断力依然不可靠。 工具生态的碎片化加剧了这个问题。 MCP 协议刚起步，A2A 协议还在早期，缺乏统一的工具描述和调用标准。每接入一个新工具，你都在写一次性的适配代码，而这些代码没有复用价值，也没有人帮你维护。 能跑 demo 和能干活之间，差的是一整套可靠性工程。 ## 回到现实：哪些槛在解，哪些槛还得等 四道门槛拆完了，一个自然的问题是：有人在解决吗？ 环境配置门槛正在被低代码平台吃掉。扣子/Coze、Dify 这类工具让简单场景的 Agent 搭建变成了拖拽操作，不需要碰代码。 但“简单场景”是个限定词，需求一复杂，你还是得回到代码层，回到那些碎片化的依赖和互不兼容的框架里。 单 Agent 编排也在变容易。Anthropic Agent SDK 和 OpenAI Agents SDK 都在简化单 Agent 开发流程。 Anthropic 官方的态度很明确：先从 Workflow 开始，人定义路径，解决80%的问题。Autonomous Agent 留给那20%真正需要的场景。 上下文工程正在成为新范式，但范式本身就是一个更高的门槛。 LangGraph 的状态管理、MCP 协议的上下文标准化都在推进，可方法论还不成熟，工具链还没成型，甚至“上下文工程”这个概念本身都没有统一定义。 这是方向，但不是捷径。 剩下的，短期无解。 多 Agent 协作的从众效应、角色不一致、臆想拼凑，这些根本性问题目前没有系统解决方案。 上下文膨胀，长任务中信息持续积累，压缩会丢关键细节，不压缩会撑爆窗口，同样没有完美方案。 Agent 行为合约还停留在学术论文里。工具生态的标准化，MCP 和 A2A 协议都还在早期，行业共识远未形成。 Khairallah 那篇教程的评论区里，有人说“收藏了，周末试”。如果你也是其中之一，这篇文章想说的是：试，但别焦虑。卡住不是你的问题，是这道题本身就还没被解出来。 但比“不用焦虑”更重要的是，下次你再看到“5分钟搭 AI 员工”的噱头，可以问自己三个问题：它有没有处理上下文崩塌？它有没有解决 Agent 协作的从众问题？它的工具调用有没有行为合约？ 三个都答不上来，那就是玩具。 5分钟搭的是玩具，不是员工。 能判断两者的区别，比假装不知道要好。 本文由 YouMind 自动从 Markdown 转换排版。

译搭AI Agent团队远非“5分钟教程”那么简单。环境配置（Python 3.10+等）劝退八成新手。Agent编排面临从众效应、角色不一致等缺陷，Anthropic采用Subagents模式，Cognition甚至发文称“不要构建多Agent”。上下文管理是最大隐性杀手，同一模型因上下文工程不同分差36个百分点。工具接入缺乏行为合约，API调用失败频发。低代码平台、单Agent SDK正降低门槛，但多Agent自由协作仍是难题。

AYi@AYi_AInotes · 6月20日53

Claude Code刚上线的Artifacts， 对重度使用者和企业团队来说，可能是今年Claude在协作场景里最有实际价值的一次升级🆙 但90%的人都看错了它的核心价值，很多人以为只是加了个网页预览窗口，或者觉得就是把聊天里的代码变成可视化页面， 其实真正的杀招是Claude直接钻进了团队协作的全流程里，而且直接戳中了AI编码时代最容易被忽略的瓶颈， Anthropic内部人都说，现在几乎所有工作都靠这个功能完成， 因为它不只是单文件的静态预览，核心是把整段工作会话连同代码上下文数据工具调用，直接打包成一个独立网页， Claude在后台继续迭代，页面就会自动同步更新，不用手动刷新重发， 生成的组织内私人链接，同事点开就能看到实时进度，还能切换历史版本， 从复杂代码的可视化解释，系统架构图，动画方案预览，到可交互的数据分析仪表盘， 全都能直接丢链接交付，不用再堆一堆截图配长长文字解释， 当然目前这个功能只对团队和企业版开放，个人用户暂时还用不上， 但这还不是最颠覆的地方， 过去整个行业都在卷模型参数卷上下文长度卷代码正确率， 大家都默认AI编码的核心矛盾是写得够不够快够不够准， 可当代码生成越来越容易，真正的卡点早就悄悄变了， 技术负责人怎么让非技术的产品和老板快速理解方案， 数据分析结果怎么让业务方自己交互式验证， 复杂系统架构怎么不靠反复开会对齐就能讲清楚， AI生成的东西越来越多，让别人看懂并相信的成本反而越来越高， 更狠的是这一步直接把Claude的定位给换了， 以前它更像程序员口袋里的个人生产力神器，是专注于你和AI一对一的对话， 现在靠着Artifacts加团队企业版的组合，它直接钻进了团队协作的流程里， PR走查的可视化注释，项目的实时仪表盘，故障排查的完整时间线， 所有AI的工作过程都能变成可持久化可共享的活文档， 本质上都是解决了企业落地AI最大的障碍之一，怎么让AI的产出安全顺畅地融进团队工作流， 很多人还在死磕模型能不能一次写出完美代码， 但真实的工作场景里，能快速可视化加分享加迭代，对整体效率的提升往往更大， 就像以前AI只是帮你写稿子，现在它连带着把可交互的演示页和分发链路一起做好了， 后者对整个组织的影响力，显然要高出一个量级， 团队版用户现在就能用的思路是，以后做技术方案评审数据复盘原型展示， 优先让Claude生成Artifacts链接交付， 降低解释成本带来的效率提升，很多时候比单纯提升生成速度还要明显， 以前AI工具拼的是帮一个人把活干得有多快， 现在开始拼的是帮一整个团队把协作成本降得有多低， 这已经不是一次小功能迭代了，已经成为AI产出从静态文本，升级成可交互可共享可实时更新的数字资产的关键一步

译Claude Code新增Artifacts功能，可将整个工作会话（代码、上下文、工具调用）打包成独立网页，后台持续迭代时页面自动同步更新。生成的私密链接可共享给团队成员，支持切换历史版本，适用于PR走查、系统架构图、交互式数据分析仪表盘等场景。目前以beta版面向Team和Enterprise计划开放。作者认为，该功能的核心价值在于将AI产出从静态文本升级为可共享、可实时更新的数字资产，降低技术方案向非技术人员解释的成本，让AI更顺畅地融入团队协作流程。

elvis@omarsar0 · 6月19日52

As I said before, for that cost & performance, I don't think Fable is worth it for a lot of SWE tasks. Tbc, I think Fable is fantastic, and it clearly shines in design & creativity. Will test it with my loops (and measure frontier efficiency) when it goes live again.

译正如我之前所说，考虑到成本和性能，我认为 Fable 对于很多 SWE 任务来说并不值得。 需要说明的是，我认为 Fable 非常出色，它在设计和创意方面明显表现优异。 等它重新上线后，我会用自己的循环测试它（并衡量前沿效率）。

Rohan Paul@rohanpaul_ai · 6月19日56

Perfect immunity from jailbreak is not possible even for the strongest of LLMs. New study shows that frontier models are getting harder to jailbreak, but not impossible to jailbreak. The study attacks Anthropic’s Fable 5 and Opus 4.8 with automated red-team tools that keep rewriting harmful prompts until the model either refuses or gives a bad answer. Fable 5 was more robust than Opus 4.8, with its worst attack success rate at 6.1%, while Opus 4.8 reached 11.5% under the strongest attack. The hard truth is that avoiding absolutely every jailbreak is practically impossible, because even a tiny failure rate can produce many harmful completions when attacks are automated and repeated at scale. The most crucial point is, that the old cartoon version of jailbreaks, weird encodings and theatrical role-play, is no longer the main problem. The surviving weakness is contextual, because adaptive attackers rewrite the request after refusals, searching for a frame the model treats as legitimate rather than dangerous. That is why perfect immunity is probably the wrong target; language models do not inspect intent from a clean moral altitude, they infer meaning through phrasing, context, and precedent. In any system this flexible, there will always be boundary cases where a harmful request looks enough like education, safety research, fiction, troubleshooting, or policy analysis to slip through. ---- Link – arxiv. org/abs/2606.18193 Title: "A Red-Team Study of Anthropic Fable 5 & Opus 4.8 Models"

译新研究对Anthropic Fable 5和Opus 4.8进行自动化红队攻击，持续改写有害提示词直至模型拒绝或生成坏答案。Fable 5最差攻击成功率6.1%，Opus 4.8为11.5%，证明最强LLM也无法完全免疫越狱——即便微小失败率，规模化自动化攻击仍可产生大量有害内容。旧式编码/角色扮演型越狱已非主要威胁，新弱点在于上下文：自适应攻击者在被拒后不断改写请求，寻找模型视为合法而非危险的框架。白宫与Anthropic正转向基于基准的测试框架，通过评分绕过程度、暴露能力、攻击可重复性及实际后果来量化越狱风险，而非追求不现实的完美免疫。

Rohan Paul@rohanpaul_ai · 6月19日67

The White House and Anthropic may have found the first serious path to restore Mythos and Fable access without pretending jailbreaks can be eliminated. AI regulation may be shifting from vague fear to a benchmark based tests of model failure, because completely removing absolutely all jailbreak is probably not a possible target. The proposed framework would score how far the bypass went, what capabilities became reachable, how repeatable the attack was, and whether the exposed behavior had real operational consequences. Both sides are now moving toward a shared way to score what a jailbreak actually means. The hard truth is that perfect immunity is probably the wrong target; a recent red-team study found even hardened frontier models still produced confirmed harmful completions under sustained automated attack, with Fable 5 remaining more robust than Opus 4.8 but not invulnerable. So once, for a newly released model, the governments can ask the same questions every time, how much was bypassed, what capability was exposed, how reproducible was the attack, and what damage could follow, thats a much more practical path.

译白宫与Anthropic正合作制定正式技术评估框架，用于量化AI模型越狱攻击的严重性，并建立标准化评估方法论。框架将开发通用基准，评估防护被绕过程度、暴露的能力、攻击可重复性及实际运营后果。双方认识到完全免疫越狱并非可行目标。近期红队研究表明，强化后的前沿模型Fable 5比Opus 4.8更鲁棒，但在持续自动化攻击下仍产生有害输出。新框架每次以相同问题衡量风险，被视为更务实的监管路径。

歸藏(guizang.ai)@op7418 · 6月19日33

Claude Code 也已经重置，端午节各位可以爽用了

Chubby♨️@kimmonismus · 6月19日52

Early Users of Anthropic Mythos still have access after US order. Mainly through project Glasswing. Via Bloomberg

译Anthropic Mythos 的早期用户在美国命令后仍可访问。 主要通过 Glasswing 项目。据 Bloomberg 报道。

🚨 AI News | TestingCatalog@testingcatalog · 6月19日52

Claude Enterprise admins can now centrally authorize MCP connectors for their organizations via a new Enterprise-Managed Auth extension. Mass MCP 👀

译Claude Enterprise 管理员现在可以通过新的 Enterprise-Managed Auth 扩展，为其组织集中授权 MCP 连接器。 Mass MCP 👀

Ethan Mollick@emollick · 6月19日58

Some (early) evidence that managers have the highest success rate in using Claude Code for coding. I have been arguing that management is an AI superpower, as clearly specifying what you want, how to do it & what good looks like is key to using agents. https://www.oneusefulthing.org/p/management-as-ai-superpower

译一些（初步）证据表明，管理者在使用Claude Code进行编码时成功率最高。 我一直认为管理是AI的超级能力，因为明确说明你想要什么、如何做以及什么样子算好，是使用智能体的关键。https://www.oneusefulthing.org/p/management-as-ai-superpower

ClaudeDevs@ClaudeDevs · 6月19日22

Earlier today, ~3% of Claude Code Max and Pro users hit a bug that showed an incorrect weekly usage limit, and in some cases blocked them from sending messages. This is fixed, and we're resetting 5-hour and weekly limits for everyone affected. Apologies for the disruption.

译今天早些时候，约3%的Claude Code Max和Pro用户遇到了一个bug，显示错误的每周使用限制，并在某些情况下阻止他们发送消息。此问题已修复，我们正在重置所有受影响用户的5小时和每周限制。对此造成的不便深表歉意。

Rohan Paul@rohanpaul_ai · 6月19日70

Nypost: Anthropic is trying to get Washington to reverse the US block on its most powerful Mythos Anthropic has proposed working more closely with the Trump administration, improving communication, and resolving security concerns faster as it seeks to end the US restrictions. --- nypost .com/2026/06/18/business/anthropic-floats-proposal-to-lutnick-to-end-us-ban-of-powerful-mythos-fable-ai-models-sources/

译《纽约邮报》：Anthropic正试图让华盛顿撤销对其最强大模型Mythos的美国禁令。 Anthropic提议与特朗普政府更紧密合作，改善沟通，更快解决安全问题，以结束美国限制。

Rohan Paul@rohanpaul_ai · 6月19日56

Anthropic vs. The US Government The kill-switch debate got user feedback.

译Anthropic 正试图让华盛顿撤销对其最强 AI 模型 Mythos 的出口限制。据纽约邮报报道，Anthropic 已向特朗普政府提交提案，承诺更紧密合作、改善沟通并更快解决安全关切，以换取解除禁令。与此同时，"杀开关"（kill-switch）讨论获得了用户反馈。

Ethan Mollick@emollick · 6月19日67

I have given AA a hard time about its previous agentic evaluation but this looks like a good and impressive benchmark for real world knowledge work that is unsaturated and had private hold out tests. This is one to watch - I didn’t see a human comparison score though?

译Ethan Mollick 称赞 AA-Briefcase 是真实知识工作的优质基准，未饱和且含私有保留测试，同时询问是否有与人类的对比。该基准由 @ArtificialAnlys 发布，测试模型在多周、多任务项目中的能力，输入含数万条 Slack 消息和数千封邮件。模型排名：Claude Fable 5（已不可用）以 1587 Elo 居首，Claude Opus 4.8（1356）第二，GLM-5.2 max（1266）第三。结果凸显难度：最佳模型仅 3% 任务满足全部标准，31/91 任务无模型超过 50%，成本跨度约 800 倍。

向阳乔木@vista8 · 6月19日62

强烈推荐安装这个Skill，比官方的Skill-creator强大很多。 如果不会写skill，用姚老师这个skill可以写出90分的skill 这个skill来源于Anthropic 官方泄露的Claude code源码，还有全网其他模型的skill整合后的元Skill。 经过姚老师长达一个月的打磨，这是我用过的最好的Meta Skill Github：https://github.com/yaojingang/yao-meta-skill

译@yaojingang（姚老师）打磨的元Skill（创建Skill的Skill）已升级至2.0。该工具源自Anthropic官方泄露的Claude code源码，并整合了全网其他模型Skill，比官方Skill-creator更强大。用户可借助它写出90分的Skill。2.0版本已推送到GitHub，附带升级方案和对比报告。

Artificial Analysis@ArtificialAnlys · 6月19日55

Announcing AA-Briefcase, the benchmark for the next era of agentic knowledge work AA-Briefcase is our new benchmark for testing models on long-horizon knowledge work tasks in complex projects built by industry experts. Models are evaluated on multi-week projects, each with many linked tasks and thousands of input source files. We evaluated Claude Fable 5 from @AnthropicAI before it became unavailable, and it currently leads with an Elo score of 1587, followed by Claude Opus 4.8 (max, 1356), Opus 4.7, and the recently-released GLM 5.2 (max, 1266) from @Zai_org. Claude Fable 5 cost $31 on average to run each AA-Briefcase task, followed by Claude Opus 4.8 at $10.40, GPT-5.5 (xhigh) at $3.68 and GLM-5.2 (max) at $2.40. AA-Briefcase comprises four private scenarios, each representing a multi-week knowledge work project set in a realistic organizational context. A public fifth scenario has been released via @huggingface as a representation of scenario structure, submission, and grading (AA-Briefcase Lite). This does not count toward official AA-Briefcase results, and is demonstrative only. Key elements of AA-Briefcase: ➤ Realistic long-horizon projects: AA-Briefcase moves beyond single, disconnected prompts by evaluating models across a coherent long-horizon project. Tasks build week by week, draw on shared institutional context, and require deliverables such as financial models, board presentations, and design mock-ups ➤ Large volumes of fragmented context: AA-Briefcase requires models to reason across thousands of inputs, including company documents, meeting transcripts, large-scale data exports, 25,000+ Slack messages and 3,500+ emails. These sources are fragmented, messy, and often contain realistic contradiction, testing whether models can navigate the ambiguity of real-world knowledge work ➤ Composite rubric and pairwise grading: AA-Briefcase combines binary rubric checks for ground-truth correctness with pairwise grading on analytical quality and presentation quality. Unlike many evaluations that focus on a single metric, AA-Briefcase tests agentic capabilities more comprehensively, exposing cases where models produce outputs that look polished but are incorrect or lack analytical rigor ➤ Built by industry experts: AA-Briefcase scenarios mirror real-world knowledge work, with tasks developed over months by experts across data science, product management and corporate strategy from companies including Google, McKinsey & Company and BCG. Task challenges are drawn from professional experience, making AA-Briefcase more reflective of the ambiguity, messy context and competing priorities that define real-world knowledge work Key results: ➤ Claude Fable 5 leads AA-Briefcase at 1587 Elo: This is followed by Claude Opus 4.8 (1356) with the next-best non-Anthropic model, GLM-5.2 (max), ~90 points back at 1266. Note that Claude Fable 5 did not use the Opus 4.8 fallback for any task in AA-Briefcase ➤ Cost per task varies by ~800x across models tested: Claude Fable 5 leads the benchmark but costs more than $31 per task on average, compared to ~$0.04 for DeepSeek V4 Flash (max). The strongest price/performance options are open weights models such as GLM-5.2 (max) and DeepSeek V4 Pro (max), with GLM-5.2 (max) scoring only ~90 Elo below Claude Opus 4.8 (max) for less than 25% of the cost ➤ Real-world complexity remains difficult for models: The top performer, Claude Fable 5, satisfies all rubric criteria on just 3% of AA-Briefcase tasks. On 31 of 91 tasks, no model scores above 50% on the rubric criteria ➤ Task difficulty scales with the number of required input files: For each rubric check, we identify the set of source files needed to pass. Across all models, pass rates fall as this file count increases, though top-tier models degrade less than weaker models More details below in thread ⬇️

译Artificial Analysis 推出新基准 AA-Briefcase，用于评估模型在长期知识工作项目中的智能体能力。基准包含 4 个私有场景（每项目需处理 25000+ Slack 消息、3500+ 邮件等碎片化上下文）及一个公开演示场景。评测结果：Claude Fable 5 以 Elo 1587 领先，其次为 Claude Opus 4.8（1356）、Opus 4.7 及智谱 GLM 5.2（max，1266）。成本方面，Claude Fable 5 平均每任务 $31，Opus 4.8 为 $10.40，GPT-5.5 (xhigh) 为 $3.68，GLM 5.2 (max) 为 $2.40，DeepSeek V4 Flash (max) 仅约 $0.04。所有模型中仅 3% 的任务满足全部标准，31/91 个任务无模型得分超 50%，显示真实世界复杂性仍是挑战。最佳性价比为开源权重模型 GLM-5.2 (max) 和 DeepSeek V4 Pro (max)。

Thariq@trq212 · 6月19日35

I'll be giving the the Day 2 opening keynote at AI Engineer World's Fair on July 1st at 9am. Hope to see you there, here's a quick preview

译我将在 7 月 1 日上午 9 点于 AI Engineer World's Fair 发表 Day 2 开幕演讲。 希望能见到你，这里有一个快速预览。

Berryxia.AI@berryxia · 6月19日42

兄弟们，Claude 是一点也不闲着啊！ 这不又又又整活了… Claude Code直接上线了Artifacts功能，让你的coding session变成能实时刷新的共享互动页面。 现在你在Claude Code里做的任何事，比如PR 演示、项目dashboard、调试过程，都能一键生成一个互动页面，通过私有链接分享给团队。 关键是它会随着你的session继续工作而自动刷新，大家看到的永远是最新的版本。 Artifacts能调用你整个session的上下文：代码库、插件、技能、已连接的工具。 分享完全在组织内，私密性有保障。目前在Team和Enterprise计划的beta版可用。 这其实把AI coding从“单人黑箱”变成了“团队实时工作台”。 以前你得手动截图、复制代码、写说明，现在直接把AI的思考过程和输出做成活的artifact，别人点开链接就能看到完整脉络，还能跟着一起迭代。 以前团队协作AI coding最麻烦的就是上下文传递和版本同步，现在Artifacts直接把这个痛点干掉了。 AI不再只是帮你写代码，它可以帮你把整个工作过程变成可共享、可进化的活文档。 这波更新把Claude Code从“个人生产力工具”往“团队协作平台”又推了一大步。

译Claude Code 新增 Artifacts 功能（Team 和 Enterprise 计划 beta 版）。用户可从 coding session 生成交互页面（如 PR 演示、项目 dashboard），通过私有链接分享给团队；页面随 session 自动刷新，调用代码库、插件、技能等全部上下文。该更新旨在将 Claude Code 从单人工具拓展为团队实时协作平台，解决上下文传递和版本同步痛点。

Yuchen Jin@Yuchenj_UW · 6月19日39

I actually see it differently. More and more companies are moving from renting closed-source LLMs to “owning model weights” with open-source LLMs and post-training. Cursor did it. OSS LLMs’ revenue could surpass Anthropic at some point.

译我其实有不同看法。 越来越多公司正从租用闭源大语言模型转向通过开源大语言模型和后训练“拥有模型权重”。 Cursor 已经做到了。 开源大语言模型的收入某个时候可能超过 Anthropic。

宝玉@dotey · 6月19日65

Claude Code 推出 Artifact 功能：AI 编程从终端走向可视化协作 Claude Code 现在可以把工作过程生成 Artifact，简单说就是一个实时更新的网页。PR 走查、系统架构说明、调试时间线、发布清单，这些原本只存在于终端会话里的东西，现在变成一个链接，发给团队成员直接打开就能看。 生成 Artifact 时，Claude Code 会用到当前会话的完整上下文，包括代码库、已接入的外部工具（比如监控系统）和对话内容。一个事故调查页面可以同时展示出错的测试代码、监控工具里的错误曲线，以及 Claude 的根因分析推理，不需要你手动接数据源或搭基础设施。 Artifact 会随会话进展自动更新。每次更新后，打开页面的人立刻能看到最新版本，同一个链接始终有效，历史版本也可以随时回溯。 Anthropic 在内部测试中发现，最高频的场景是调试。一个工程师在早会前启动事故排查，Claude Code 查着日志就发布了一个 Artifact：时间线、可疑提交、错误率图表。她把链接丢进群里，等早会开始时页面已经更新了两次。团队不用再听"我来介绍一下 agent 查到了什么"，所有人看着同一个页面讨论就行。 这其实解决了 AI 编程工具的一个实际问题：agent 在终端里干了很多活，但成果只有操作者自己看得到，团队协作时还得靠人肉"翻译"。Artifact 把这个中间环节省了。 安全方面，Artifact 默认私有，只有同组织内认证成员可以查看，不能公开。管理员可以控制组织级开关、设置角色权限和数据保留策略。 除了调试和 PR 走查，Anthropic 列出的用例还包括：法务用它审计所有第三方依赖的开源许可证，安全团队做代码审查报告并把每个发现链接到具体代码行，平台财务从 Terraform 代码里提取云资源成本分布，设计师用真实组件库生成多套 UI 方案直接挑选。 目前 Artifact 功能以 beta 形式向 Claude Team 和 Enterprise 组织开放，可通过 Claude Code CLI 和桌面应用生成，页面在任何浏览器里查看。个人用户暂时用不了。

译Claude Code 新增 Artifact 功能，可将终端会话中的 PR 走查、调试时间线等过程生成实时更新的交互页面，并通过私有链接分享给团队成员。Artifact 利用当前会话的完整上下文（代码库、外部工具、对话），随会话自动更新，支持历史版本回溯。默认私有，仅同组织认证成员可见。该功能以 beta 形式向 Claude Team 和 Enterprise 组织开放，通过 CLI 和桌面应用生成，个人用户暂不可用。Anthropic 内部测试显示调试场景最高频。

🚨 AI News | TestingCatalog@testingcatalog · 6月19日54

Claude Code users on Team and Enterprise plans gained access to Artifacts, new interactive pages that can be built based on their Claude Code sessions. Every session is an Artifact now 👀

译Claude Code 的 Team 和 Enterprise 计划用户现已获得 Artifacts 访问权限，这些是基于 Claude Code 会话构建的新交互式页面。 现在每个会话都是一个 Artifact 👀

Chubby♨️@kimmonismus · 6月19日55

The third major win for OpenAI: Ha Thai has departed Meta to join OpenAI to lead communications for devices As Axios says: OpenAI is expected to unveil its first device this year. Therefore, they hired a key figure from a competitor to prepare the hardware release. Things have been rather quiet regarding the hardware device so far. But this makes it clear that OpenAI will continue down this path.

译OpenAI近期完成两起关键人事引进：从Meta挖来Ha Thai负责设备通讯，Axios称OpenAI预计今年推出首款设备；同时，曾帮助塑造特朗普政府政策的Dean Ball也加入OpenAI。在最大竞争对手Anthropic与美国政府关系紧张的背景下，OpenAI借由吸纳这些与政府关系密切的人物，提前应对潜在冲突并强化政府渠道。除开源外，OpenAI成为Anthropic当前困境的最大赢家。

Boris Cherny@bcherny · 6月19日56

I've been using Artifacts in Claude Code for everything: visual explanations of tricky code, system diagrams, quick previews of a few animation options, data analyses and dashboards I share with the team. They are a game changer for how I work with Claude. Can't wait to hear what you think!

译Boris Cherny 分享他用 Claude Code Artifacts 的体验：可视化解释复杂代码、系统图、动画预览、数据分析及团队共享仪表盘，称其改变了与 Claude 的工作方式。@claudeai 宣布：Artifacts 可从会话中生成交互页面（如 PR 走查或项目仪表盘），通过私密链接共享给团队；目前以 Beta 版形式面向 Team 和 Enterprise 计划用户提供。

Chubby♨️@kimmonismus · 6月19日66

Let's just briefly acknowledge what a huge win OpenAI has made today. - Noam Shazeer helped invent the Transformer architecture, left Google to build CharacterAI, came back to DeepMind through a $2.7B deal, worked on Gemini. Now: OpenAI - Dean Ball, who helped shape the Trump administration's, heading to OpenAI; this allows them to anticipate solutions to future conflicts and, at the same time, have good access to government circles. Especially against Anthropic, who are struggeling. I think many people don't understand how significant this is for OpenAI. Even The Information acknowledges as a huge win: "Shazeer’s hiring is a major win for OpenAI in the AI talent wars, as OpenAI competes to catch up with archrival Anthropic’s most advanced models ahead of the two companies’ hotly anticipated initial public offerings. Shazeer was a key author on the original paper on transformers, which essentially predict the most-likely next word in a phrase, and is the architecture that OpenAI used to underpin its GPT models." Images: WSJ/CNBC

译OpenAI今日取得重大胜利：Transformer架构发明者之一Noam Shazeer正式加入，他曾创立CharacterAI、通过27亿美元交易回归DeepMind并参与Gemini项目；曾帮助塑造特朗普政府AI政策的Dean Ball也加盟。The Information指出，Shazeer的招聘是AI人才战的重大胜利，他是支撑GPT系列模型的Transformer原始论文关键作者。Ball的加入使OpenAI能提前预判政策冲突并深入美国政府人脉圈，而其主要竞争对手Anthropic正面临政府关系困境。OpenAI成为Anthropic当前挣扎的最大受益者。