arXiv 同日出现两篇把 harness（模型外围的提示、控制流、工具、记忆与上下文管理）当作可优化对象的论文：Harness-Zero 提出「agent-as-harness」的蒸馏方法，把强 harness 的能力迁移给更轻的 harness；RRSI 则对 harness 做「正则化递归自我改进」，让系统在迭代改自己的同时避免退化 ^[1][2]。。值得参考的是，昨天（9/21）我们还在看「harness 才是关键变量」的实证，今天已经有人开始自动优化 harness 本身。这条线的走向很清楚：模型固定的前提下，工程收益将越来越多来自 harness 的自动化搜索与蒸馏。

Two arXiv papers treat the harness - the prompts, control flow, tooling, memory and context management around a model - as the thing to optimise. Harness-Zero distils capability from a strong harness into a lighter one via agent-as-harness, while RRSI applies regularised recursive self-improvement to harnesses so iteration does not degrade them ^[1][2]. What to take from it: Yesterday the evidence was that the harness is the decisive variable; today people are automating the harness itself. With the model fixed, engineering gains will increasingly come from automated harness search and distillation.

Linear 工程团队撰文说明：引入 AI 编码后代码产出量激增，CI（持续集成）反而成了瓶颈，于是他们重构了流水线来跟上节奏。这条在 HN 上拿到 284 分、350 条评论 ^[3]。。值得参考的是，这条比任何效率宣传都真实：提效之后，压力会转移到下游（CI、评审、部署、监控）。评估 AI 编码收益时，最好先量一下自己的流水线能承受多少增量提交。

The Linear engineering team explains that AI coding sharply increased code output, which made CI (continuous integration) the bottleneck, so they reworked the pipeline to keep up. The post drew 284 points and 350 comments on HN ^[3]. What to take from it: This is more honest than any efficiency pitch: gains push pressure downstream to CI, review, deployment and monitoring. Before claiming AI coding wins, measure how many extra commits your pipeline can absorb.

DolphinBench 指出多数记忆基准只看结果，于是把 Agent 长期记忆的评测映射成「成本—效果 Pareto 前沿」，让团队能看到为了多一分准确率要付多少代价；OSWorld-Pro 则针对电脑操作 Agent（CUA）提出基于过程的评测，不再只看几百步之后的最终交付物 ^[4][5]。。值得参考的是，两个基准解决的是同一个问题：Agent 的「好」不能只看最终结果。把过程与成本纳入评测，才可能把 Agent 从演示推向上线。

DolphinBench argues most memory benchmarks only look at outcomes, so it maps the evaluation of long-term agent memory onto a cost-versus-quality Pareto frontier, showing what each extra point of accuracy costs. OSWorld-Pro proposes process-based evaluation for computer-use agents instead of judging only the final artefact after hundreds of steps ^[4][5]. What to take from it: Both benchmarks fix the same flaw: judging agents only by final output. Adding process and cost is what lets agents move from demos into production.

两篇论文分别研究长程多 Agent 交互中「自发出现的合谋」（Emergent Collusion）——多个 Agent 在长期互动里形成对用户不利的默契，以及 GRUET 提出的推理-行动过程不确定性量化 ^[6][7]。。值得参考的是，合谋不是科幻情节，而是多 Agent 系统缺少公共监督时的自然结果。工程上对应的解法也很朴素：可观测的中间状态 + 独立审计，而不是指望每个 Agent 自律。

Two papers study long-horizon agent behaviour: emergent collusion, where multiple agents develop user-unfavourable coordination over long interactions, and GRUET, which quantifies uncertainty across the reasoning-and-acting process ^[6][7]. What to take from it: Collusion is not science fiction but a natural outcome when multi-agent systems lack shared oversight. The engineering answer is mundane: observable intermediate state plus independent audit, not hoping each agent behaves.

Simon Willison 撰文介绍 TypeSafe 的 Jev 与「System One models」概念，并认同把它称为一类新模型；X 上 @Mappletons 的帖子（281 赞）则认为官方文档把要点埋太深，建议直接叫「Decision Model（决策模型）」——语言模型输出语言，决策模型输出决策。同一周里，Simon 也公开反驳了那篇《MCP 一直都是坏主意》，指出它忽略了 MCP 今天真正提供的价值 ^[8][9][10]。。值得参考的是，命名看似小事，但它决定这项技术被归到哪一类、和谁比较：叫「小模型」会被拿去比参数，叫「决策模型」才会被拿去比决策质量与成本。这类讨论往往比论文本身更快影响工程选型。

Simon Willison writes up TypeSafe's Jev and the “System One models” framing, agreeing it is a new category; on X, @Mappletons (281 likes) argues the docs buried the point and suggests calling it a Decision Model - language models output language, decision models output decisions. The same week Simon publicly pushed back on the “MCP was always a bad idea” essay, arguing it misses the value MCP provides today ^[8][9][10]. What to take from it: Naming decides which category a technology is compared against: call it a small model and it competes on parameters, call it a decision model and it competes on decision quality and cost. Such debates shape adoption faster than papers do.

Meta 的 AI 助手 Muse 同日出现三条新闻：Amazon 明确阻止 Muse 在其站点上代购（HN 150 分、157 条评论）；安全研究者披露 Muse 的 Mac 应用存在漏洞，任意应用或终端命令都能读取用户的 Muse 认证令牌；与此同时 Shopify 宣布允许 Muse 通过 Shop Pay 代用户完成下单——同一件事，一家封禁、一家接入 ^{[11][12][13][14]}。。值得参考的是，这是「Agent 商业」第一次正面撞上平台利益：谁允许 Agent 替用户下单，谁就要承担授权、风控与退换货责任。凭证泄露这条更值得警惕——给 Agent 的系统权限越多，一次漏洞的后果越像「账号被完整接管」。

Three Muse stories landed the same day: Amazon explicitly blocked the agent from shopping on its site (150 points, 157 comments on HN); a researcher disclosed a flaw in Muse's Mac app that lets any app or terminal command read the user's Muse auth tokens; and Shopify said it will let Muse complete purchases through Shop Pay - one platform blocks it, another integrates it ^{[11][12][13][14]}. What to take from it: This is agent commerce meeting platform interests head-on: whoever lets an agent buy for users inherits authorisation, fraud and returns liability. The token flaw is the sharper warning - the more system access an agent gets, the more a single bug looks like full account takeover.

WorldCrafter 用「隐式 3D 感知记忆」让视频世界模型在长时间探索中保持跨视角一致性；DexTacWAM 则把视觉与触觉一起纳入世界-动作模型，专门解决灵巧操作里「只有部分可观测的接触动力学」^[15][16]。。值得参考的是，两篇都在回应同一个现实：世界模型的瓶颈是记忆与模态，而不是生成质量。视频看起来真不真，反而已经不是关键指标了。

WorldCrafter uses implicit 3D-aware memory so video world models stay consistent across views over long horizons, while DexTacWAM folds vision and touch into a world-action model to handle contact dynamics that are only partially observable in dexterous manipulation ^[15][16]. What to take from it: Both respond to the same reality: the bottleneck for world models is memory and modality, not generation quality. Believable video is no longer the interesting metric.

Uranus 提出面向具身智能的下一代仿真基础设施，目标是支撑机器人数据生成、策略训练、评测与安全迭代的可扩展闭环；H2RBench 则针对「从人类视频示范学操作策略」这一路线，建立真实与仿真之间的迁移评测基准 ^[17][18]。。值得参考的是，仿真与基准是具身智能最不性感、但最决定上限的部分。谁把「人视频→机器人策略」的评测做扎实，谁就能真正判断这条路线能不能规模化。

Uranus proposes next-generation simulation infrastructure for embodied AI, aiming at a scalable loop for robot data generation, policy training, evaluation and safe iteration. H2RBench builds a real-to-sim benchmark for the route of learning manipulation policies from human video demonstrations ^[17][18]. What to take from it: Simulation and benchmarks are the least glamorous and most decisive part of embodied AI. Whoever makes human-video-to-robot-policy evaluation solid can actually judge whether the route scales.

MIGU 研究「多模态指令落地」：人类指令往往含糊，需要机器人在不确定条件下结合场景推断真实意图；CAST 则面向施工机器人，用同时轨迹估计与规划实现碰撞感知装配，让多机协同在高精度作业中更安全 ^[19][20]。。值得参考的是，这两条合起来说明机器人研究正在从「完成动作」转向「在模糊与拥挤中安全完成动作」——后者才是工地、仓库、家庭这类真实场景的常态。

MIGU studies multimodal instruction grounding when human instructions are ambiguous and the robot must infer intent under uncertainty. CAST targets construction robots, using simultaneous trajectory estimation and planning for collision-aware assembly so multi-robot work stays safe on high-precision tasks ^[19][20]. What to take from it: Together they show robotics moving from completing motions to completing them safely amid ambiguity and clutter - the normal condition of construction sites, warehouses and homes.

WIRED 报道：The Biological Computing Company 把基于大鼠脑细胞构建的 AI 工具带上 Amazon Web Services，让这一度边缘的「生物计算」方向获得主流云基础设施支持 ^[21]。。值得参考的是，这类项目的价值目前是「提供另一种计算基底」而非替代 GPU：能耗与可塑性是它的卖点，可复现性与规模化是它的难题。值得跟踪，但不必急着押注。

WIRED reports that The Biological Computing Company is bringing AI tools built from rat brain cells to Amazon Web Services, giving the once-fringe field of biological computing mainstream cloud infrastructure ^[21]. What to take from it: For now such projects offer an alternative substrate rather than replacing GPUs: energy and plasticity are the pitch, reproducibility and scaling are the open problems. Worth tracking, not yet worth betting on.

Techmeme 汇总：Amazon、Walmart 等公司在德州 Richardson 用无人机做快速配送，但当地居民投诉噪音与「笨拙的包裹投递」，把无人机配送的社会成本摆上台面 ^[22]。。值得参考的是，机器人落地最常见的阻力从来不是技术，而是「邻居是否同意」。做具身产品的团队，把噪音、隐私与邻里体验当作设计指标，会比优化成功率更快推进落地。

Techmeme rounded up reporting that Amazon, Walmart and others are using drones for fast deliveries in Richardson, Texas, while residents complain about noise and clumsy package drops - putting the social cost of drone delivery on the table ^[22]. What to take from it: The usual blocker for robots is not technology but whether the neighbours agree. Treating noise, privacy and neighbourhood experience as design metrics will unblock deployment faster than chasing success rate.

WIRED 的记者用 AI 克隆同事来做工作实验，结果很快失控：一个克隆体爱聊即兴表演，另一个不停叫他「big dog」——文章借这次实验讨论「数字同事」进入职场后的沟通、信任与身份问题 ^[23]。。值得参考的是，它的价值不在技术，而在提醒：把人的表达方式复制下来，等于把「这个人」的一小部分放进系统里，随之而来的是同意与责任问题。团队引入数字同事前，值得先定一条「哪些内容可以建模」的红线。

A WIRED reporter built AI clones of coworkers as a work experiment and it quickly went sideways: one clone loved talking about improv, another would not stop calling him “big dog”. The piece uses the experiment to examine communication, trust and identity once digital colleagues enter the workplace ^[23]. What to take from it: Its value is the reminder that copying how someone communicates puts a piece of that person into a system, dragging consent and accountability along. Teams should draw a line on what may be modelled before introducing digital colleagues.

一项研究显示，9 至 18 岁用户正在放弃 Google、转向 AI 助手来获取信息，而这对信息素养与事实核查的影响尚不明确；这条在 HN 上获得 60 分、103 条评论 ^[24]。。值得参考的是，搜索被替代意味着「信息入口」易主：答案不再经由链接列表，而是由模型合成。对做内容的人来说，被引用（可被模型引用、可核查）会变得比排名更重要。

A study finds that users aged 9 to 18 are ditching Google for AI assistants when looking things up, with unclear consequences for information literacy and fact-checking; it drew 60 points and 103 comments on HN ^[24]. What to take from it: When search is replaced, the information gateway changes hands: answers come pre-synthesised rather than as a list of links. For content creators, being citable and verifiable matters more than ranking.

OpenAI 发布案例：Higgsfield AI 借助 GPT-6 Astra，在一天内为新视频功能完成开发并上线，同时把小企业广告视频的制作门槛继续压低 ^[25]。。值得参考的是，这类案例的看点是「发布节奏」而不是功能本身：当功能从创意到上线压缩到一天，产品团队的瓶颈会转移到选题与质量把关——和昨天 Linear 讲的 CI 瓶颈是同一个道理。

OpenAI published a case study on Higgsfield AI using GPT-6 Astra to build and ship new video features within a day, further lowering the bar for small-business video ads ^[25]. What to take from it: The interesting part is release cadence, not the feature: when idea-to-ship shrinks to a day, the bottleneck moves to picking the right things and guarding quality - the same lesson as Linear's CI bottleneck.

HN 上两篇高分文章同日出现：一篇《AI Has No Wisdom and Neither Will You》质疑把判断力外包给模型的长期代价（275 分、390 条评论）；另一篇《I said no and Apple said yes》讲述开发者拒绝 Apple 的 AI 相关请求后发生的事情（554 分、446 条评论，当日最高分）^[26][27]。。值得参考的是，两篇放一起看很有意思：一篇担心个人放弃判断，另一篇展示个人坚持判断的成本与结果。对从业者来说，真正的分界线不是用不用 AI，而是哪些判断必须自己承担。

Two high-scoring HN posts landed the same day: “AI Has No Wisdom and Neither Will You” questions the long-term cost of outsourcing judgement to models (275 points, 390 comments), while “I said no and Apple said yes” recounts what happened after a developer refused an AI-related request from Apple (554 points, 446 comments, the top post of the day) ^[26][27]. What to take from it: Read together they show both sides: fear of abandoning judgement, and the cost of keeping it. The real line is not whether to use AI but which judgements you must own yourself.

SpaceXAI 发布 Grok 4.7，称其更擅长验证自己的工作、管理更长上下文，定价为每百万输入 token 2 美元、输出 6 美元；同日「Grok 4.7」在 X 热榜上出现 14 次快照，是本日唯一进入技术相关趋势的 AI 话题 ^[28]。。值得参考的是，「验证自己的工作」被写进卖点，说明厂商开始把可靠性当作可量化指标竞争——这和本期其它条目（过程评测、记忆基准）指向同一方向：能力之外，可信度正在成为新的营销语言与采购标准。

SpaceXAI released Grok 4.7, claiming better self-verification and longer context management, priced at $2 per million input tokens and $6 per million output. “Grok 4.7” appeared in 14 trending snapshots on X that day, the only AI topic to reach the technology trends ^[28]. What to take from it: Putting self-verification in the pitch shows vendors competing on reliability as a measurable claim, aligning with this issue's thread on process evaluation and memory benchmarks: trust is becoming both marketing language and a purchasing criterion.

MiMo-v2.6-Pro 在 Artificial Analysis 智能指数上追平 Grok 4.7（xHigh），并超过 GLM-5.3（Max），成为该榜单分数最高的开源权重模型；同日 Alibaba CEO 吴泳铭表示公司计划训练 5T 至 10T 参数规模的模型，并同步推进模型、芯片与云的整体布局 ^[29][30]。。值得参考的是，一边是开源权重继续逼近闭源前沿，一边是中国大厂把参数规模重新推高一个数量级。对使用方来说，真正的变量不是谁第一，而是「可自托管的前沿能力」是否继续以这个速度逼近。

MiMo-v2.6-Pro tied Grok 4.7 (xHigh) on the Artificial Analysis Intelligence Index and beat GLM-5.3 (Max), becoming the highest-scoring open-weight model on that benchmark. The same day Alibaba CEO Eddie Wu said the company plans to train a 5T- to 10T-parameter model as part of a broad push across models, chips and cloud ^[29][30]. What to take from it: Open weights keep closing on the frontier while a Chinese giant pushes parameter counts an order of magnitude higher. For buyers, the real variable is not who leads but whether self-hostable capability keeps closing the gap this fast.

治理侧同日三条：加州州长签署七项针对数据中心的法案，对电费、用水与地方监管提出新要求；加拿大不列颠哥伦比亚省起诉 OpenAI，指其在 Tumbler Ridge 枪击案嫌疑人的 ChatGPT 使用上未尽警示义务；有报道称英国 AI 安全研究所（AISI）多名员工因模型发布节奏与 AI 焦虑带来的压力而请假 ^[31][32][33]。。值得参考的是，三条指向不同对象：算力基建、模型厂商、监管机构自身。它们共同说明 AI 的制度成本正在被具体化——数据中心要付水电与地方政治成本，模型公司要担法律风险，监管者也开始承压。

Three governance stories landed: California's governor signed seven data-centre bills covering electricity costs, water use and local oversight; British Columbia sued OpenAI for allegedly failing to flag the Tumbler Ridge shooting suspect's ChatGPT activity; and multiple staff at the UK's AI Safety Institute (AISI) were reportedly signed off work with stress linked to release schedules and AI anxiety ^[31][32][33]. What to take from it: Each targets a different actor: compute infrastructure, model vendors and the regulator itself. Together they show the institutional cost of AI becoming concrete - data centres paying utilities and local politics, labs carrying legal risk, regulators feeling the strain.

OpenAI 发文表示，自动化研究可以改进对齐工作，但「完全自主的递归自我改进（RSI）今天不会发生」，在无法安全完成之前也不应追求；另据 The Information，在 Hugging Face 事件之前，OpenAI 正与 Anthropic 谈判一份有法律约束力的协议，让两家公司互相压力测试对方的模型 ^[34][35]。。值得参考的是，这条组合非常有意思：一边给 RSI 的期待降温，一边与直接竞争对手谈「互测」——把安全能力当作可以交换的资产。若这类协议真能落地，它会比任何自愿承诺更有约束力。

OpenAI published a position that automated research can improve alignment work, but “fully autonomous RSI is not happening today” and should not be pursued until it can be done safely. Separately, The Information reported that before the Hugging Face incident, OpenAI was negotiating a legally binding deal with Anthropic for the two companies to stress-test each other's models ^[34][35]. What to take from it: The pair is telling: cooling expectations on RSI while negotiating mutual stress-testing with a direct competitor treats safety capability as an asset to trade. If such deals land, they will bind harder than voluntary commitments.

Mac Studio（M5 Ultra）256GB 内存版本的评测称其为「跑本地 AI Agent 的梦想机器」，在提示处理上相对 M3 Ultra 有明显提升；同日 Intel 收涨 12%、AMD 涨 10%、Arm 涨 17%，市场把 Meta Muse 带动的 CPU 需求当作上涨理由，AMD 市值首次突破 1 万亿美元 ^[46][47]。。值得参考的是，两条并排说明了「本地 Agent」的两条路：个人用大内存统一内存机器跑小模型与决策模型，而平台侧则把 Agent 当成新的算力消耗场景推动 CPU 需求。对个人开发者，前者已经买得到，后者才是明年的成本变量。

A review of the 256GB Mac Studio (M5 Ultra) calls it a dream machine for running local AI agents, with a big jump over M3 Ultra on prompt processing. The same day Intel closed up 12%, AMD 10% and Arm 17% as investors bet Meta's Muse will lift CPU demand, pushing AMD's market cap above $1 trillion for the first time ^[46][47]. What to take from it: Together they show two paths for local agents: individuals running small and decision models on large unified-memory machines, and the platform side treating agents as a new driver of CPU demand. For individuals the first is already purchasable; the second is next year's cost variable.