§30 秒速览The 30-second version

先给全貌,后面每一章往下挖一层。

The whole shape first; every later section digs one layer down.

1为什么接一台仪器要几周Why wiring up one instrument takes weeks

动机先讲清楚,后面的设计才有参照系。

Get the motivation straight first; the design only makes sense against it.

先别管 AI。假设你是一个实验室的自动化工程师,手上有一台液体处理器(负责精确移液)、一条机械臂(负责搬运 96 孔板)、一台酶标仪(负责读吸光度,也就是样品吸收了多少光)。要让它们配合跑完一个实验,你要解决三件事:

Forget AI for a moment. You're the automation engineer in a lab with a liquid handler (moves precise volumes), a robotic arm (moves 96-well plates around), and a plate reader (measures absorbance — how much light a sample absorbs). To make them run one experiment together you have three problems:

1. 接口各说各话。每台设备有自己的编程接口、自己的数据格式、自己的驱动。厂商之间没有共同语言,所以「让 A 和 B 配合」这件事本身要写一段专门的胶水代码。
2. Every interface is its own language. Each device has its own programming interface, data format and driver. There's no shared vocabulary across vendors, so "make A cooperate with B" is itself a bespoke piece of glue code.
3. 数量是乘法不是加法。四台设备意味着最多六对关系,每一对都可能需要一个翻译器。再加一台,新增的不是一条线而是一排线——这就是图 1 左侧那团意面。
4. The count multiplies, it doesn't add. Four devices means up to six pairings, each possibly needing its own translator. Add a fifth and you don't add one line, you add a fan of them — the spaghetti on the left of Figure 1.
5. 机器的「常识」不在代码里。这条最要命。机械臂自重多少、这台仪器允许的温度上限是多少、这个操作在什么情况下不该做——这些信息大多躺在纸质手册里、某个人的电脑上,或者干脆只存在于老工程师的脑子里(所谓「隐性知识」)。代码里看不到。
6. The machine's "common sense" isn't in the code. This is the killer. How much the arm weighs, the instrument's temperature ceiling, when an operation should simply not be attempted — most of that lives in paper manuals, on someone's laptop, or purely as tacit knowledge in a veteran engineer's head. None of it is visible from the code.

华盛顿大学的博士生 Zihao Song 在公告里给出的具体感受更有说服力:他的仪器来自不同厂商,各有软件、数据格式和驱动;把它们串起来是一个「要花几个月到几年、成本从数千到数百万美元」的集成问题,「大多数实验室根本负担不起」。他此前的自动化尝试是——花几周评估平台、追着厂商要支持、学着写设备之间的胶水代码,最后放弃了。

Zihao Song, a PhD student at the University of Washington, puts it more concretely in the announcement: his instruments come from different vendors, each with its own software, data format and driver. Wiring them together is an integration problem that "takes months to years and can cost anywhere from thousands to millions of dollars," putting it out of reach for most labs. His earlier automation attempt was: weeks evaluating platforms, chasing vendor support, learning to write glue code between instruments — and finally giving up.

再叠上 AI,问题不是变简单而是变复杂

Adding AI doesn't simplify this — it adds a layer

就算你把设备接通了,还有第二层问题:它们没有共同的方式把数据交给 Agent,也没有共同的方式让 Agent 安全地操作它们。传统自动化实验室的做法是上一个中央调度器(scheduler),把流程写死——这在工厂里成立,因为一条产线可能把同一个流程跑一万次。但研究实验室一年跑几十种协议,一半是新的,而且经常要在跑到一半时改(蛋白产量远低于预期、DNA 组装失败)。传统自动化用「固定流程」换效率,恰好牺牲了研究最需要的灵活性。

Even with the devices connected, a second problem remains: there is no common way for them to hand data to an agent, and no common way to let an agent operate them safely. Classic automated labs solve this with a central scheduler and a hard-coded workflow — which works on a factory line that runs the same protocol ten thousand times. A research lab runs dozens of protocols a year, half of them new, and often has to revise one mid-run because protein yield came back far below assumption or a DNA assembly failed. Traditional automation buys throughput by fixing the workflow — sacrificing exactly the flexibility research needs.

2MHS 到底是什么What MHS actually is

四个组成部分,逐个讲透。

Four moving parts, one at a time.

MHS 的核心是一个标准化驱动。「驱动」这个词在这里就是它在操作系统里的老意思:一段在计算机系统和硬件设备之间做翻译的软件。MHS 做的事是规定这段翻译该长什么样。它有四个组成部分。

At its core MHS is a standardized driver — "driver" in the ordinary operating-system sense: software that translates between a computer's OS and a hardware device. What MHS specifies is what that translation should look like. It has four parts.

① 极简原语:read 和 write

① Minimal primitives: read and write

MHS 驱动用一组非常简单的命令覆盖所有设备:read(例如「读取温度」)和 write(例如「设置温度」)。任何硬件设备都能理解并执行这个层级的指令。这看起来朴素得可疑,但正是它让「统一」成为可能——你不需要为显微镜和机械臂设计两套语义模型,只需要各自声明「我能读什么、能写什么」。

An MHS driver covers every device with a very small command set: read (e.g. "get temperature") and write (e.g. "set temperature"). Any hardware device can understand and act on commands at that level. It looks suspiciously plain, but that plainness is what makes uniformity possible — you don't design two semantic models for a microscope and a robot arm; each just declares what it can read and what it can write.

② 自动发现:设备和 Agent 互相找得到

② Discovery: devices and agents can find each other

驱动让每台设备以标准格式可被发现,于是设备与 Agent 能跨网络互相找到并通信,中间不再需要一个定制的「翻译器」程序。这是图 1 右侧那条横线能存在的前提。

The driver makes each device discoverable in a standard format, so devices and agents can find each other and communicate across networks without a bespoke translator program in between. That's the precondition for the single horizontal bar on the right of Figure 1.

③ 自然语言标签:把「隐性知识」写进驱动

③ Natural-language tags: writing tacit knowledge into the driver

这是 MHS 最值得注意的设计,也是它和纯工程标准分道扬镳的地方。驱动里带标签,允许用户用自然语言直接写下那些「从代码里看不出来」的机器特性——公告举的例子是机械臂的自重,因为知道它多重才知道怎么安全地操纵它。填写方式有两种:自己写,或者让一个 Agent 来「采访」你的硬件配置,边聊边填。

This is the most distinctive piece of MHS, and where it parts ways with purely engineering-oriented standards. The driver carries tags that let a user write, in plain language, the machine characteristics that aren't discernible from code — the example given is the weight of a robot arm, which you need in order to manipulate it safely. You fill them in either yourself, or by chatting with an agent that interviews you about your hardware setup.

④ 参考文件:Agent 拿到的「说明书」

④ The reference file: the agent's manual

有了这些标签,MHS 驱动会自动生成一份参考文件,写清这台设备的一般特性:它能测什么、什么参数可调、会强制执行哪些安全上限。这份文件就是 Agent 操作这台设备所需要知道的全部——包括它从没见过的设备。

From those tags, the MHS driver automatically produces a reference file describing the device's general characteristics: what it can measure, what can be adjusted, and what safety limits will be enforced. That file is everything the agent needs to operate the device — including one it has never seen before.

3怎么用:三条控制通路How you use it: three control paths

MCP、命令行、代码文件——为什么是三条,不是一条。

MCP, CLI, code files — and why it's three rather than one.

设备接好、Agent 也读懂了参考文件之后,还差最后一步:怎么下达控制。MHS 给了三条通路,它们不是三个可选项,而是三种不同节奏的控制方式,配合使用才能跨设备编排。

With devices connected and the reference file understood, one step remains: how control is actually issued. MHS offers three paths — not three alternatives but three different tempos of control, meant to work together for cross-device orchestration.

最有意思的模式:先在线探索,再固化成脚本

The most interesting pattern: explore online, then freeze into a script

公告里有一个很能说明问题的观察:测试中,Claude 与实验和硬件的互动方式「很像一个科学家」——它对激光做一次调整,通过相机观察结果、评估这次调整把光束移到了哪里,然后重复这个过程,试图理解事件的因果序列。然后它做了一件关键的事:把学到的东西打包进代码文件,写出一段确定性脚本,让整个对准过程可以作为一条命令跑完,不再需要每一步都推理。

One observation in the announcement is especially telling: in testing, Claude interacted with experiments and hardware "much as a scientist would" — it adjusted a laser, observed the result through a camera to assess how the adjustment moved the beam, and repeated, trying to understand the sequence of events. Then it did the key thing: it packaged what it learned into code files, writing a deterministic script that aligned the laser as a single command, without reasoning at each step.

4三份实证与数字Three field reports, with numbers

公告里三家合作方的自述——包括他们翻车的地方。

Three partner write-ups from the announcement — including where things went wrong.

案例一 · Genentech:让 Claude 自己调出移液流速

Case 1 · Genentech: letting Claude tune its own pipetting flow rate

Genentech 的研究者把 MHS 部署到三台设备上——液体处理器(精确移液)、机械臂(搬运板)、酶标仪(读吸光度),用 Claude 做协议编排和硬件通信中枢,自动化 BCA 蛋白定量(一种测样品总蛋白浓度的标准流程)。全部实验在标准 96 孔板里进行。

Genentech deployed MHS across three instruments — a liquid handler (precise transfers), a robotic arm (moves labware), a microplate reader (measures absorbance) — using Claude to orchestrate the protocol and act as the hardware communication hub, automating the BCA protein assay (a standard procedure for total protein concentration). All experiments ran in standard 96-well microplates.

难点在流体物理。BCA 涉及物性差异很大的液体:从简单的水溶液到粘稠、易起泡的蛋白样品。他们用已知浓度的牛血清白蛋白(BSA)作标准品。BSA 粘稠,在高流速下会起泡,而起泡直接影响移液准确度。

The difficulty is fluid physics. BCA involves liquids with very different properties, from simple aqueous reagents to viscous, foamy protein samples. They used bovine serum albumin (BSA) at known concentrations as the standard. BSA is viscous and forms bubbles at high flow rates, which directly degrades pipetting accuracy.

1. 基线:先给 Claude 标准 BCA 协议跑一遍。它把步骤执行下来了,但给水溶液和粘稠液选了同一个通用流速,导致粘稠液起泡、转移不准。
2. Baseline: Claude was given the standard BCA protocol. It executed the steps, but picked generic liquid-handling parameters with the same flow rate for both aqueous and viscous solutions, causing bubbles in the viscous one and inaccurate transfers.
3. 自主优化:研究者给出实验设计和专家定义的流速范围,让 Claude 用染色液体做试转移 + 酶标仪读吸光度来找每种液体的最优流速,并给它一份专家在同一块板上做的「ground truth」转移作对照,目标是最小化与专家结果的差距。
4. Autonomous optimization: researchers supplied the experimental design and an expert-defined flow-rate range, and asked Claude to find the optimum for each liquid type by running trial transfers with dyed liquid and taking absorbance readings, with an expert-performed "ground truth" transfer in the same plate as the target to minimize against.
5. 自评:Claude 计算 RMSE(均方根误差)量化自己的准确度——越低越好,0 为满分。公告举例:目标 100 微升、实际 98 微升,那 2 微升的偏差就计入分数。
6. Self-scoring: Claude computed RMSE (root mean square error) to quantify its accuracy — lower is better, zero is perfect. Per the write-up: aiming for 100 µL but dispensing 98 means that 2 µL miss counts against the score.
7. 结果:水收敛到约 140 µL/s(RMSE 0.016),BSA 收敛到约 10 µL/s(RMSE 0.181)——Genentech 的自动化专家确认这两个参数对他们的装置是合理的。而按常规做法,这种优化需要自动化专家为每一组参数手写编程逻辑并反复分析数据。
8. Result: water converged at roughly 140 µL/s (RMSE 0.016) and BSA at 10 µL/s (RMSE 0.181) — parameters their automation experts confirmed were reasonable for the setup. Ordinarily this optimization requires an automation specialist to hand-write programming logic for every single parameter set and iterate on the data.

案例二 · 华盛顿大学 Baker / Pinglay 实验室:把实验室搬到办公室

Case 2 · UW Baker / Pinglay labs: taking the bench remote

背景是从头蛋白设计(de novo protein design)——凭空设计自然界不存在的蛋白。设计一个像 PETase(分解塑料的酶)这样的蛋白,如今可以便宜到 0.01 美元;但在实验台上验证它每个候选要约 100 美元、一周人力,而他们一次要测 1000 个候选。瓶颈完全在湿实验一侧。

The context is de novo protein design — building proteins that don't exist in nature. Designing something like PETase (the plastic-degrading enzyme) can now cost as little as $0.01; testing it at the bench costs around $100 and a week of labour per candidate, and they test 1,000 candidates at a time. The bottleneck is entirely on the wet-lab side.

Zihao Song 用 MHS 做了三件事。其一,远程监控:以前要在实验室里来回走着看每台仪器,现在所有仪器把状态汇报到一个仪表盘,可以在笔记本上看,甚至用手机问 Agent。其二,Agent 监控 qPCR:qPCR 通过反复升降温复制目标 DNA,信号呈 S 形曲线——先平、后陡升、最后因试剂耗尽进入平台期。跑进平台期会扭曲 DNA 文库、拿不到准确定量,所以必须盯着曲线在正确时刻叫停,这过去要人守好几个小时。现在 Agent 实时分析扩增曲线、识别曲线形态,在关键节点问研究者停还是继续;被告知停就中止反应,并把仪器推进到下一步——4 °C 保温,防止 DNA 降解。其三,机械臂交接:用基于 LeRobot 的开源机械臂,由 Claude Code 通过 MHS 协调机械臂与液体处理器,液体处理器分液完成后约 10 秒触发机械臂取板,反复测试中两台设备从未相撞。

Zihao Song did three things with MHS. First, remote monitoring: instead of roving the lab to check each instrument, all of them report status to one dashboard he can read from a laptop, or query through an agent from a phone. Second, agent-supervised qPCR: qPCR amplifies target DNA through repeated heating and cooling, and the signal follows an S-curve — flat, then steep, then a plateau as reagents run out. Running into the plateau distorts the library and ruins quantification, so someone has to watch the curve and halt at the right moment, often for hours. Now the agent analyses the amplification curves in real time, identifies the pattern, and at the right junctures asks the researcher whether to stop or continue; told to stop, it halts the reaction and advances the instrument to a 4 °C hold that keeps the DNA from degrading. Third, plate handoff: using an open-source LeRobot-based arm, Claude Code coordinated arm and liquid handler through MHS — about 10 seconds after dispensing finished, the agent triggered the arm to lift the plate, and across repeated tests the two instruments never collided.

案例三 · CMU:8 小时 vs 数周

Case 3 · CMU: eight hours vs several weeks

卡内基梅隆的团队做的是系列稀释确定剂量-反应曲线:从高浓度开始,每次按同一比例稀释,用上一次的结果配下一次。人工做这套实验要几周,而且容易出错——最高浓度太高会饱和(信号顶到头,曲线在上方压平,高剂量点不再提供信息)、步长太小覆盖范围不够、步长太大又会跳过响应真正变化的过渡区。

The CMU team ran serial dilution to determine dose-response curves: start with a strong solution, dilute by the same ratio each step, using the last dilution to make the next. By hand this takes weeks and is error-prone — too high a maximum concentration saturates the signal (the curve flattens at the top and high doses stop being informative), too small a step size doesn't cover enough range, too large a step skips the transition region entirely.

他们的装置:CyBio FeliX 液体处理器、Varioskan LUX 酶标仪、一条搬 96 孔板的机械臂、若干监控相机——分布在三台电脑上,接口彼此根本不兼容。用 MHS,他们从零为每台仪器写驱动,再写一个编排层,让 Claude Opus 4.8 自主跑完整个协议。

Their setup: a CyBio FeliX liquid handler, a Varioskan LUX plate reader, a robotic arm for 96-well plates, and monitoring cameras — spread across three computers with fundamentally incompatible interfaces. With MHS they wrote drivers from scratch for each instrument plus an orchestration layer, letting a Claude Opus 4.8 agent run the full protocol autonomously.

5和 MCP / SiLA 2 / OPC UA 比vs MCP, SiLA 2 and OPC UA

MHS 不是凭空出现的——实验室与工厂早就有标准,搞清分工才知道它补的是哪一块。

MHS didn't appear in a vacuum — labs and factories already have standards. Knowing the division of labour tells you what it actually adds.

和 MCP 的关系:互补,不是替代

Relative to MCP: complementary, not a replacement

最容易混淆的一组。MCP 是 Anthropic 此前推出的标准,让 AI 模型访问外部数据源与软件工具;MHS 把同一个原则延伸到物理设备。两者的连接点很具体:MHS 声明任何 Agent harness 都可以用标准协议接入,MCP 就是其中之一——也就是说,MHS 通过 MCP 把设备暴露给 Agent,Agent 于是能用自然语言处理指令、决定需要哪些先后动作。

The easiest pair to confuse. MCP is Anthropic's earlier standard for giving AI models access to external data sources and software tools; MHS extends the same principle to physical devices. The join is concrete: MHS states that any agent harness can access it using standard protocols, such as MCP — so MHS exposes devices through MCP, and the agent can then take natural-language instructions and work out the required sequence of actions.

和实验室既有标准的关系

Relative to the standards labs already have

实验室互操作性不是新话题。SiLA 2 是目前最成熟的实验室仪器连接开放标准,基于 HTTP/2 与 Protocol Buffers、走 gRPC 传输格式,用「特性定义语言(FDL)」给出机器可读、带类型的能力描述,并要求 mDNS/DNS-SD 做即插即用发现,同时包含加密、认证、授权与审计追踪。OPC UA 来自工业自动化,是一套平台无关、面向服务的架构,支持发现、读写、方法调用、通知、历史数据访问与审计,能从嵌入式设备一路扩展到企业系统和云。

Lab interoperability is not a new topic. SiLA 2 is the most mature open standard for laboratory instrument connectivity: built on HTTP/2 and Protocol Buffers over the gRPC wire format, with a Feature Definition Language giving machine-readable, typed capability descriptions, mandated mDNS/DNS-SD plug-and-play discovery, and provisions for encryption, authentication, authorization and audit trails. OPC UA comes from industrial automation: a platform-independent, service-oriented architecture supporting discovery, read/write, method execution, notifications, historical data access and auditing, scaling from embedded devices to enterprise systems and cloud.

学术侧也有同类尝试

Academia is trying the same thing

值得知道 MHS 不是唯一一个方向:arXiv 上有一篇 《LAP: An Agent-to-Instrument Protocol for Autonomous Science》(arXiv:2606.03755),同样在做「Agent 到仪器」的协议。⚠️ 该论文本次未做全文阅读,仅作为「这个方向有平行工作」的指路,不对其内容做评价。

Worth knowing MHS isn't the only attempt: arXiv carries "LAP: An Agent-to-Instrument Protocol for Autonomous Science" (arXiv:2606.03755), working on the same agent-to-instrument problem. ⚠️ The paper was not read in full for this document; it's flagged only as evidence of parallel work, with no judgment on its content.

6坑、限制与未解问题Pitfalls, limits, open questions

这一节的材料大部分来自官方公告自己——难得地没有粉饰。

Most of this comes from the announcement itself, which is unusually candid.

坑一:物理直觉缺口(最重要的一个)

Pitfall 1: the physical-intuition gap (the important one)

Genentech 写得很直白:模型擅长通用推理,但仍然在物理、化学、生物约束上吃力,尤其是在需要真实世界物理直觉的故障排查上。他们给的例子是气泡——看起来无害,实则触发一连串问题:

Genentech states it plainly: models excel at general-purpose reasoning but still struggle with physical, chemical and biological constraints, particularly when troubleshooting errors that call for real-world physical intuition. Their example is bubbles — seemingly benign, actually a cascade:

• 要吸 40 µL 试剂但液里有气泡,实际转移的液体量会因为空气占位而偏少;
• aspirating 40 µL of reagent when there are air bubbles means the actual liquid volume transferred is lower, because air takes up the space;
• 液面传感器碰到的是泡沫而不是液体时,会触发硬件错误;
• liquid-level sensors throw hardware errors when the tip meets foam instead of liquid;
• 气泡还会扭曲光学读数,而光学读数正是这个实验的最终结果。
• and bubbles distort the optical readings that are the experiment's final readout.

而 Claude 遇到混合过程中由气泡引发的运行时错误时,默认反应是在同一个孔里换参数重试——这只会把液体搅得更厉害、产生更多气泡。因为它不理解失败背后的物理机制。研究者不得不引导它使用更温和的液体处理参数。

And when Claude hit runtime errors caused by bubbles during mixing, its default instinct was to retry in the same well with different parameters — which only agitated the fluid further and created more bubbles, because it did not understand the physics of the failure. The researchers had to steer it toward gentler handling parameters.

坑二:自主恢复能力是真的,但有边界

Pitfall 2: autonomous recovery is real, but bounded

同一场实验里,Claude 确实自主从取头失败、液面检测错误中恢复了——公告指出这是当前科研仪器大多不具备的能力。所以「能不能自主恢复」不是一个是非题:能恢复的是「已知错误码 → 已知处置」这一类,不能恢复的是需要理解物理机制才知道怎么办的那一类。做系统设计时,这条界线比「Agent 能不能自主」更有用。

In the same run Claude did recover on its own from tip pickup failures and fluid detection errors — a capability the announcement notes current scientific instruments mostly lack. So "can it self-recover" isn't a yes/no: it recovers from "known error code → known remedy" cases, and fails on cases where you need to understand the underlying physics to know what to do. When designing a system, that line is more useful than asking whether the agent is "autonomous."

坑三:长时监控的算力成本

Pitfall 3: the compute cost of long monitoring windows

华盛顿大学那份自述里有一句容易被跳过但很关键的话:「让 Agent 在长监控窗口里持续运行,产生的算力成本需要与省下的研究者时间相权衡。」——这不是抽象顾虑。qPCR 盯曲线可能要几小时,如果 Agent 每个循环都在线推理一次,账单是实打实的。这也正好解释了为什么 MHS 要提供「代码文件」这条确定性通路。

One easily skipped but important line in the UW write-up: "Running an agent continuously over long monitoring windows also has compute costs that need to be weighed against the researcher time saved." That's not an abstract worry. Watching a qPCR curve can take hours; if the agent reasons online every cycle, the bill is real. It also explains exactly why MHS offers the deterministic code-file path.

坑四:阶段限制

Pitfall 4: stage limitations

• 研究预览,需申请。目前面向首批科研实验室与先进制造厂商,通过 modelhardwarestandard.com 申请;尚未开源,Anthropic 表示预览结束后计划开源。
• Research preview, by application. Currently open to a first group of scientific labs and advanced manufacturers via modelhardwarestandard.com; not yet open-source, though Anthropic says it plans to open-source it after the preview.
• 安全评测仍在建设中。Anthropic 明确说共享早期版本的目的之一,就是与合作方一起建立安全评测和最佳实践——换句话说,这套东西的安全评估体系目前还不完整。
• Safety evaluations are still being built. Anthropic explicitly says one reason for sharing an early version is to collaborate on building safety evaluations and best practices — meaning the evaluation regime for this is not yet complete.
• 只是概念验证。华盛顿大学的原话:「这些演示仍然只是概念验证。更复杂的实验协议需要大量优化才能可靠工作,还需要整合更广泛、更复杂的物理操作。」
• Still proofs of concept. UW's own words: "These demonstrations are still just proofs of concept. More complicated experimental protocols will require significant optimization to work reliably, as well as the integration of broader and more complex physical manipulations."
• 人的专业判断仍然必需。据 Techzine 转述,Anthropic 指出 AI 模型在空间与物理推理上仍有局限,人类专业知识依然必要。
• Human expertise remains necessary. Per Techzine, Anthropic notes that AI models still have limits in spatial and physical reasoning, so human expertise is still required.

7对做 Agent 的人意味着什么What it means if you build agents

就算你这辈子不碰移液器,这份公告里有四条可以直接搬走的东西。

Even if you never touch a pipette, four things here transfer directly.

1. 把安全边界放进工具层,不要放进 prompt。MHS 的参考文件里写的是「会被强制执行的安全上限」。模型会忘、会被说服、会在长上下文里丢失约束;工具层不会。凡是「越界后果不可逆」的操作,约束都应该下沉一层。
2. Put safety limits in the tool layer, not the prompt. The MHS reference file carries "safety limits that will be enforced". Models forget, can be argued around, and lose constraints in long contexts; the tool layer doesn't. Any operation whose overrun is irreversible should have its constraint pushed one layer down.
3. 给工具补一份「人话说明书」。MHS 最反常规的设计是自然语言标签——把「代码里看不出来」的隐性知识写进驱动。你的 MCP server 或工具定义同样可以这么干:除了参数 schema,写清「这个工具在什么情况下不该调、失败通常意味着什么、有哪些不写在类型里的前提」。Genentech 的气泡事故本质就是这份说明书缺了一页。
4. Give your tools a plain-language manual. The most unusual thing in MHS is the natural-language tags — writing tacit, code-invisible knowledge into the driver. Your MCP server or tool definitions can do the same: beyond the parameter schema, say when the tool should not be called, what a failure usually means, and which preconditions aren't expressible in the types. The Genentech bubble incident is, at bottom, a missing page in that manual.
5. 「在线推理」和「确定性脚本」要有明确分工。MHS 把这条做成了架构:探索期用 MCP 一步一想,稳定后固化进代码文件让设备自己跑。物理世界给了一个硬判据(长时任务、超过推理速度的操作),软件世界的判据其实一样——只是账单来得没那么疼,所以容易拖着不做。
6. Draw a clear line between online reasoning and deterministic scripts. MHS builds this into the architecture: reason step by step through MCP while exploring, then freeze into code files the devices run themselves. The physical world supplies a hard criterion (long-running tasks; operations faster than reasoning). The criterion is the same in software — the bill just hurts less, so it's easy to keep putting off.
7. 把一次性纠正沉淀成 skill。研究者纠正 Claude 的气泡处理后,把结论写成了可复用的液体处理 skill。这是 Agent 工程里回报率最高的动作之一:每一次「我又得跟它解释一遍 X」都是一个应该被写下来的 skill。
8. Turn one-off corrections into skills. After correcting Claude's bubble handling, the researchers codified the takeaways into a reusable liquid-handling skill. This is one of the highest-return moves in agent engineering: every "I had to explain X to it again" is a skill that should have been written down.

§来源清单Sources

调研时间 2026-08-31。标签区分一手官方与第三方。

Researched 2026-08-31. Tags separate first-party from third-party.

• 官方一手Anthropic《Previewing the Model Hardware Standard》(2026-08-27,含 Genentech、华盛顿大学 Baker/Pinglay 实验室、卡内基梅隆大学三份合作方自述)—— anthropic.com/news/model-hardware-standard-research-preview
• First-partyAnthropic, "Previewing the Model Hardware Standard" (2026-08-27; includes partner write-ups from Genentech, the UW Baker/Pinglay labs and Carnegie Mellon) — anthropic.com/news/model-hardware-standard-research-preview
• 官方一手MHS 研究预览申请入口 —— modelhardwarestandard.com(本次未 fetch 站点内容,仅按官方公告指引列出)
• First-partyMHS research-preview application page — modelhardwarestandard.com (not fetched for this document; listed as pointed to by the official announcement)
• 第三方Techzine Global《New Anthropic standard MHS connects AI to machines》(2026-08-28,转述 Ars Technica 关于「MHS 不依赖 AI」的观察,并列出厂商试验清单)—— techzine.eu
• Third-partyTechzine Global, "New Anthropic standard MHS connects AI to machines" (2026-08-28; relays Ars Technica's point that MHS is not AI-dependent, and lists vendors experimenting with it) — techzine.eu
• 第三方EVOBYTE《SiLA 2, OPC UA, and the Modern Lab Stack》(2026-05,SiLA 2 与 OPC UA 技术特征的综述,引用 SiLA 联盟与 OPC 基金会官方材料)—— evo-byte.com
• Third-partyEVOBYTE, "SiLA 2, OPC UA, and the Modern Lab Stack" (2026-05; overview of SiLA 2 and OPC UA characteristics, citing SiLA Consortium and OPC Foundation material) — evo-byte.com
• 官方一手SiLA 联盟规范与 FAQ —— sila-standard.com/faq;OPC 基金会 UA 概览 —— opcfoundation.org
• First-partySiLA Consortium spec and FAQ — sila-standard.com/faq; OPC Foundation UA overview — opcfoundation.org
• 第三方《LAP: An Agent-to-Instrument Protocol for Autonomous Science》,arXiv:2606.03755 —— arxiv.org/pdf/2606.03755(仅列为平行工作,本次未读全文)
• Third-party"LAP: An Agent-to-Instrument Protocol for Autonomous Science", arXiv:2606.03755 — arxiv.org/pdf/2606.03755 (listed as parallel work; not read in full)
• 官方一手Anthropic 关于 MCP 的介绍(MHS 与 MCP 的衔接说明引自此)—— anthropic.com/news/model-context-protocol
• First-partyAnthropic on MCP (the MHS↔MCP relationship is cited from here) — anthropic.com/news/model-context-protocol