有开发者发现：Claude Code 只有在开启遥测（telemetry）时才会读取 AGENTS.md。这一细节让上周刚被当作「跨工具 Agent 约定」的 AGENTS.md 重新变成争议话题——HN 讨论 302 分、134 条评论，集中在「工程约定是否应该与数据上报绑定」 ^[12]。。值得参考的是，这条是很好的提醒：Agent 工具的默认行为本身就是一种产品立场。团队在把 AGENTS.md 写进仓库前，值得先确认工具的读取条件，否则「约定」会在你没注意时失效。

A developer found that Claude Code only reads AGENTS.md when telemetry is enabled. The detail turned last week's cross-tool agent convention into a controversy again - 302 points and 134 comments on HN, mostly about whether engineering conventions should be tied to data reporting ^[12]. What to take from it: The reminder: an agent tool's defaults are a product stance. Before committing AGENTS.md to a repo, check the conditions under which it is read, or the convention silently stops applying.

两篇论文给出 harness 的两个方向：Agensh 把多 Agent 组织扩展到 1,024 个 Agent，用组织化协作降低复杂任务延迟；《Grow the Harness, Not the Context》则主张把重复性工作沉淀成可复用的「专家 Agent」，而不是不断往上下文里塞内容 ^[1][2]。。值得参考的是，两篇合起来是一个成熟信号：Agent 工程的重心正从「怎么塞更多信息」转向「怎么沉淀能力」。这和软件工程史上从脚本到模块化的路径几乎一致。

Two papers give harness research two directions: Agensh scales multi-agent organisations to 1,024 agents to cut latency on complex work, while “Grow the Harness, Not the Context” argues for distilling repeated work into reusable specialist agents instead of piling more into context ^[1][2]. What to take from it: Together they signal maturity: agent engineering is moving from stuffing more information to accumulating capability - almost exactly software engineering's path from scripts to modules.

CliffCompaction 针对「百万 token 级上下文」的长程编码 Agent，提出更省成本的上下文压缩方案；SWE-Serve 则把评测对象换成生产级推理服务工程任务，检验 Agent 在真实服务代码上的工程能力 ^[3][4]。。值得参考的是，上下文压缩直接对应账单，而生产级基准对应可信度——这两件事比再刷一个 SWE 榜单分数更有商业意义。

CliffCompaction targets long-horizon coding agents that need millions of tokens of context, proposing cheaper compaction; SWE-Serve re-points evaluation at production inference-serving engineering tasks to test agents on real service code ^[3][4]. What to take from it: Compaction maps directly to the bill and production benchmarks map to trust - both more commercially meaningful than another SWE leaderboard point.

论文 A2M 指出：Agent 依赖语义匹配从第三方 MCP 服务器中挑选工具，于是「语义供应链」成了可攻击对象——攻击者可以通过优化调用轨迹来劫持 Agent 的工具选择 ^[5]。。值得参考的是，这与前几天的 MCP 协议之争形成互补：MCP 的价值是真的，但「第三方工具市场」天然带供应链风险。生产环境里，工具白名单与调用审计不应该等到出事才补。

The A2M paper shows that because agents select tools from third-party MCP servers by semantic matching, the semantic supply chain becomes attackable: an attacker can optimise the call trace to hijack which tool the agent picks ^[5]. What to take from it: It complements the MCP debate: the protocol's value is real, but third-party tool markets carry supply-chain risk. In production, allowlists and call auditing should not wait for an incident.

OpenAI 发文阐述「有效第三方评估」的优先级与原则：严格、安全、独立，并计划让外部机构在模型的训练、评估与部署阶段都能进行技术安全评估 ^[32]。。值得参考的是，把外部评估写进流程，意味着「自证安全」的模式在松动。对企业采购方来说，这也是一个可用的信号：以后可以要求供应商提供第三方评估结论，而不是只有自我声明。

OpenAI published priorities and principles for effective third-party assessments - rigorous, secure, independent - and said it plans to let outside groups run technical safety evaluations across training, evaluation and deployment ^[32]. What to take from it: Writing external evaluation into the process loosens self-certification. For enterprise buyers this becomes an askable artefact: third-party findings rather than a self-declaration.

为 Agent 提供网页抓取能力的 Firecrawl 完成 7500 万美元 B 轮（Smash Capital 领投）^[16]；同日包括美国银行与 Capital One 在内的六家大银行警告：agentic commerce（由 Agent 代下单）自主度提高会带来更多诈骗与欺诈风险 ^[46]；面向受监管行业的本地化 AI 厂商 Go.AI 也拿到 8500 万美元 A 轮 ^[17]。。值得参考的是，三条并排看就是这门生意的真实结构：Agent 需要数据（Firecrawl）、需要合规部署（Go.AI），而它的最大阻力来自金融风控（银行）。做 Agent 商业化的团队，早点把「谁承担责任」写清楚，比多做一个功能重要。

Firecrawl, which provides web scraping for agents, raised a $75M Series B led by Smash Capital ^[16]. The same day six major banks including Bank of America and Capital One warned that higher autonomy in agentic commerce will bring more scams and fraud ^[46], while Go.AI, selling on-prem AI to regulated industries, raised an $85M Series A ^[17]. What to take from it: Read together they show the shape of the business: agents need data (Firecrawl) and compliant deployment (Go.AI), while the binding constraint is financial risk control (the banks). For agent commerce teams, settling liability early matters more than another feature.

Ars Technica 报道：丰田要求工人参与训练人形机器人，同时对外表示「人类不会被取代」。这类安排把「工人既是被替代者、又是训练数据的提供者」这一矛盾直接摆到台面上 ^[27]。。值得参考的是，这类新闻在未来一年会越来越多。对做具身智能的团队，如何设计「参与感与收益分配」，可能和技术方案一样决定项目能否推进。

Ars Technica reports that Toyota is ordering workers to help train humanoid robots while saying humans will not be replaced - putting the contradiction on the table: workers are both the ones being automated and the source of training data ^[27]. What to take from it: Expect more of this. For embodied-AI teams, designing participation and benefit-sharing may decide a project's fate as much as the technical approach.

两篇机器人论文：Imperfection for Precision 主张把不完美的数据（低质量遥操作、失败轨迹）升级利用，以降低高精度操作的数据门槛；Beyond End-Task Success 则指出机器人「记住过去经验并在新场景复用」时，只看最终成功率不足以评估，需要审计它到底选对了哪条经验 ^[6][7]。。值得参考的是，两篇都在处理同一个现实：机器人数据永远不够干净，所以「如何利用脏数据」和「如何确认它用对了历史」比追求完美数据集更实际。

Two robotics papers: Imperfection for Precision upcycles imperfect data (low-quality teleoperation, failed trajectories) to lower the data bar for high-precision manipulation, while Beyond End-Task Success argues that when robots reuse past experiences in new scenes, final success rate is not enough - you must audit which experience they selected ^[6][7]. What to take from it: Both deal with reality: robot data is never clean, so using messy data and verifying which history was reused beats chasing perfect datasets.

Counterpoint Research 数据显示，2026 年上半年全球 AI 眼镜出货量同比增长 263%，其中无显示款与 AR/AI 眼镜共同驱动，Meta 与 EssilorLuxottica 是主要推动者 ^[24]。。值得参考的是，第一人称视觉与语音入口正在以消费电子速度铺开——这比机器人本体更快地积累了「真实世界的视觉数据与使用习惯」。对具身智能来说，这是最现实的感知数据来源之一。

Counterpoint Research says global AI glasses shipments grew 263% year over year in H1 2026, driven by display-less and AR/AI glasses alike, with Meta and EssilorLuxottica leading ^[24]. What to take from it: First-person vision and voice are spreading at consumer-electronics speed, accumulating real-world visual data and habits faster than robots themselves. For embodied AI that is one of the most practical perception sources.

TM-APR 提出用解析式在线自适应做「热成像时序记忆定位」，面向弱光、烟尘等可见光失效的场景——这正是消防、地下巡检等场景里机器人定位的痛点 ^[8]。。值得参考的是，这类「不性感但必需」的感知工作，往往决定机器人能不能真的进入现场。相比端到端模型，它的优势是可解释、可验证。

TM-APR proposes analytic online adaptation for thermal temporal-memory localisation, targeting conditions where visible light fails - smoke, darkness, underground inspection - precisely where robot localisation struggles ^[8]. What to take from it: This unglamorous but necessary perception work decides whether robots can actually enter the field, and unlike end-to-end models it is inspectable and verifiable.

YouTube 同日发布一批 AI 功能：用户可以用自然语言构建自己的推荐算法、创作者获得 A/B 测试与动态缩略图、直播配音，以及 AI 驱动的「Ask YouTube」视频购物 ^[47]。。值得参考的是，把推荐权交给用户，本质是把「算法解释成本」转成「用户表达成本」。这类功能能否成功，取决于用户是否愿意为个性化付出描述成本——值得观察。

YouTube shipped a batch of AI features the same day: users can build their own recommendation algorithm in natural language, creators get A/B testing, dynamic thumbnails and live dubbing, and an AI-powered “Ask YouTube” for video shopping ^[47]. What to take from it: Handing feed control to users converts algorithm explainability cost into user expression cost. Whether it works depends on users paying that cost - worth watching.

WIRED 报道 AT&T 正在用自动化替代大量岗位，并把这家老牌电信公司的组织变化放进更大的行业背景里讨论 ^[30]。。值得参考的是，与丰田那条形成对照：自动化在不同行业的推进方式不同——制造业强调「人训练机器」，电信业直接削减岗位。判断就业影响要看行业，而不是笼统谈 AI。

WIRED reports that AT&T is automating away jobs and frames the organisational change at the legacy telecom giant against the wider industry ^[30]. What to take from it: Contrast with Toyota: automation proceeds differently per industry - manufacturing frames it as humans training machines, telecom simply cuts roles. Judging employment impact requires industry specifics.

OpenAI 一边庆祝 OpenAI Academy 两周年、扩展学习路径，一边与 Grab 合作推出「GO Forward with AI」，计划在东南亚培训 3 万名合作伙伴掌握实用 AI 技能 ^[33][34]。。值得参考的是，这类项目的真实作用是扩大使用基数：当「会用 AI」成为基础技能，提效的瓶颈会从工具转向组织流程——这也是这一个月日报里反复出现的主题。

OpenAI marked two years of OpenAI Academy with expanded learning paths and partnered with Grab on “GO Forward with AI”, aiming to train 30,000 partners across Southeast Asia in practical AI skills ^[33][34]. What to take from it: These programmes expand the user base. Once AI literacy is baseline, the bottleneck moves from tools to organisational process - a recurring theme this month.

HN 上两篇文章同日出现：《I Don't Want the Details》讨论在抽象层越来越高的今天，工程师主动放弃细节理解的代价（118 分、72 条评论）；另一篇则面向业务方解释「为什么有了 AI，写软件依然很难」（62 分、57 条评论）^[13][14]。。值得参考的是，两篇正好补上「AI 提效」的反面：当产出变快，理解深度与沟通成本成为新的稀缺资源。它们也解释了为什么本栏目一直强调「可验证」。

Two HN essays landed together: “I Don't Want the Details” examines the cost of engineers deliberately giving up detailed understanding as abstractions rise (118 points, 72 comments), while another explains to business stakeholders why software is still hard to build even with AI (62 points, 57 comments) ^[13][14]. What to take from it: They supply the counterweight to AI productivity: when output speeds up, understanding and communication become the scarce resources - and why this digest keeps insisting on verifiability.

论文指出一个被长期忽略的工程事实：贪心解码（greedy decoding）常被当作确定性行为，但同一模型、同一提示在不同数值精度下会产生不同输出，这会影响评测复现与线上稳定性 ^[9]。。值得参考的是，这条对做评测的人尤其重要：如果你的评测与线上推理精度不同，分数差异可能来自数值精度而不是模型能力。把精度写进评测配置，是一个低成本的习惯。

A paper documents a widely ignored engineering fact: greedy decoding is treated as deterministic, yet the same model and prompt produce different outputs at different numerical precisions - which breaks evaluation reproducibility and production stability ^[9]. What to take from it: Critical for evaluators: if your benchmark runs at a different precision from production, score gaps may come from numerics rather than capability. Recording precision in eval configs is a cheap habit.

Anthropic 发布 Claude Opus 5.5：官方称在多数任务上对标 Fable 5.1，但运行成本比 Opus 5 低约 40%，定价 4 美元/百万输入 token；OpenAI 发布 GPT-6 Sol 与 Luna 两个模型，主打不同的能力/成本配比，并同步改进了提示缓存。两者分别登上 HN 第一、第二名（1664 分 / 1651 分），并进入 X 热榜（GPT-6 Sol 9 次快照、Claude Opus 5.5 3 次） ^{[10][11][15][29][31]}。X 上 Anthropic 的发布帖（12,140 赞）强调 Opus 5.5 来自用户反馈、比 Opus 5.0 更便宜且 token 效率更高；另有实践者帖（1,100 赞）展示用 Opus 5.5 从单一提示构建完整 3D 场景 ^[48][49]。。值得参考的是，两家在同一天把「性价比」作为主打，说明前沿竞争的战场已经从「谁最强」转向「谁的单位成本更低」——这对自建 Agent 流水线的团队是直接利好，也解释了本期为什么有这么多成本优化研究。

Anthropic released Claude Opus 5.5, which it says matches Fable 5.1 on most tasks while costing about 40% less to run than Opus 5 at $4 per million input tokens; OpenAI released GPT-6 Sol and Luna with different capability/cost trade-offs and improved prompt caching. They took the top two spots on HN (1,664 and 1,651 points) and both trended on X (GPT-6 Sol in 9 snapshots, Claude Opus 5.5 in 3) ^{[10][11][15][29][31]}. On X, Anthropic's launch post (12,140 likes) stresses that Opus 5.5 comes from user feedback and is cheaper and more token-efficient than Opus 5.0, while a practitioner post (1,100 likes) shows building a full 3D scene from a single prompt with it ^[48][49]. What to take from it: Both leading with price-performance on the same day shows the frontier fight moving from who is strongest to who costs least per unit - good news for anyone running agent pipelines, and why so much of today's research is about cost.

除了发布 Opus 5.5，Anthropic 还宣布把 Pro、Max 与 Team 套餐的五小时使用额度提高 20%；并与 OpenEvidence 合作，为约 100 个中低收入国家的医生提供专门的医疗 AI 搜索工具 ^[18]。。值得参考的是，「提高额度」是应对竞争最直接的手段，而医疗合作则是把合规与专业场景当作护城河。两条动作针对不同人群，但都在争夺同一件事：日常使用习惯。

Beyond Opus 5.5, Anthropic raised five-hour usage limits by 20% on Pro, Max and Team plans, and partnered with OpenEvidence to give physicians in about 100 low- and middle-income countries a specialised medical AI search tool ^[18]. What to take from it: Raising limits is the most direct competitive lever, while the medical partnership turns compliance and specialist workflows into a moat. Different audiences, same goal: daily habit.

英国竞争与市场管理局（CMA）提议要求 Google 与 Android 厂商在首次使用和每年提醒时提供搜索选择屏，并对 AI 相关功能加规则；美国方面，一位 Facebook 联合创始人提议设立「AI 监督办公室」，把监管团队嵌入前沿实验室、任期三年；同时有报道称财政部长 Scott Bessent 成为「AI 沙皇」热门人选 ^[22][23][21]。。值得参考的是，三条指向同一个问题：AI 的监管到底该「在外面管」还是「进到里面看」。嵌入式监管如果成真，会改变实验室与政府的信息关系，也改变外界能看到的证据类型。

The UK's CMA proposed requiring Google and Android OEMs to show search choice screens at first use and annually, with new rules for AI features. In the US, a Facebook co-founder proposed an “Office of AI Supervision” embedding regulators inside frontier labs on three-year terms, while Treasury Secretary Scott Bessent emerged as a frontrunner for the AI czar role ^[22][23][21]. What to take from it: All three ask whether AI should be regulated from outside or supervised from inside. Embedded supervision would reshape what labs tell governments - and what evidence outsiders ever see.

据报，由前 Anthropic 研究员创立、目标是构建自我改进 AI 的 Mirendil 正洽谈约 10 亿美元融资（Kleiner Perkins 领投）；AI 训练数据公司 Micro1 融资超 1 亿美元，估值 40 亿美元（一年前为 5 亿）；此外法律 AI 公司 Noxtua 被德国出版社 C.H.BECK 收购多数股权 ^[19][20][26]。。值得参考的是，一边是「让 AI 改进 AI」拿到大额资金，一边是训练数据公司估值一年涨 8 倍——资本在下注最上游的两件事。对从业者来说，这意味着数据与评测能力会更值钱。

Mirendil, founded by former Anthropic researchers to build self-improving AI, is reportedly in talks for a ~$1B round led by Kleiner Perkins. AI training-data startup Micro1 raised over $100M at a $4B valuation, up from $500M a year earlier, while legal-AI company Noxtua sold a majority stake to German publisher C.H.BECK ^[19][20][26]. What to take from it: Money is flowing to making AI improve AI and to training data, with the latter's valuation up eightfold in a year. Upstream data and evaluation capability will keep getting more valuable.

彭博报道，五角大楼表示对 AI 的过度依赖是导致伊朗一所学校遭导弹袭击的因素之一（HN 777 分、398 条评论）；Anthropic 披露阻断了一起针对宗教领袖与海外侨民的「亲北京」影响力行动；微软则捣毁了一个借助 AI 的平台，该平台入侵了 12,000 个账号 ^[45][25][28]。。值得参考的是，三条共同说明：AI 的风险已经不以「模型会不会失控」的形式出现，而是以「人在流程里如何信任 AI 输出」的形式出现。这也让「可核查的证据链」从工程要求变成安全要求。

Bloomberg reports the Pentagon said overreliance on AI contributed to a missile strike on a school in Iran (777 points, 398 comments on HN). Anthropic disclosed disrupting a Beijing-aligned influence operation targeting religious leaders and diaspora communities, and Microsoft disrupted an AI-assisted platform that compromised 12,000 accounts ^[45][25][28]. What to take from it: Together they show AI risk arriving not as runaway models but as humans trusting AI output inside a process - which turns verifiable evidence chains from an engineering preference into a safety requirement.