In late August 2026 Anthropic announced the Model Hardware Standard (MHS) — a shared standard letting AI agents operate physical devices in labs and manufacturing. The news read as technical, but underneath it sits a pattern that concerns companies who will never see a laboratory.
What was announced
The problem MHS solves will sound familiar to anyone who has ever connected two systems: devices do not talk to each other. Each has its own program, its own way of communicating and its own vendor. Connecting them means someone writes the bridge by hand — which, in Anthropic's description, takes weeks or months.
MHS introduces a shared language built on two very simple commands: read ("what is the temperature") and write ("set the temperature"). Alongside them come natural-language tags documenting what until now lived in paper manuals or only in an experienced operator's head. Integration, they say, drops to hours or minutes.
The standard is currently a research preview, open to selected labs and manufacturers through a waitlist. Anthropic says it will open-source it later, without giving a date.
The results they published
The announcement cites concrete figures from testing, and they are more convincing than the headline:
- Genentech — automated a protein assay coordinating a robotic arm, a liquid handler and a plate reader. Claude set the flow rates itself: about 140 µL/s for water and 10 µL/s for viscous samples, matching what an expert would specify.
- Carnegie Mellon — drug-discovery experiments roughly three times faster than before, across three mutually incompatible interfaces.
- University of Washington — six instruments connected in under a week, instead of the usual months.
- QuEra Computing — laser-locking success raised from 58% to 99.3%, with time per attempt cut from 150 seconds to six.
- Tetsuwan Scientific — across 9,143 individual liquid transfers, the model predicted precision about 12% more accurately than the manufacturer's own specification.
Vendors adding support include AWS, QIAGEN, Tecan, Universal Robots, Doosan Robotics and Raspberry Pi.
The least-quoted part, and the most valuable
The same announcement states a limitation without dressing it up: Claude learns about the physical world from text and images, so its spatial and physical reasoning is limited and still needs expert oversight.
The example beats any explanation. At Genentech, researchers had to guide Claude to understand that errors caused by foaming in the samples were a physical failure, not a software bug. The model looked at the numbers and hunted for a fault in the code, because foam does not exist for it — it has never seen any, only read about it.
Our view
The news itself is not for small companies. None of our clients will be wiring a microscope to an AI agent. But the pattern is worth taking.
First: the problem MHS solves is not a laboratory problem, it is a general one. Every company running a few systems lives the same friction — the warehouse scale does not talk to the program, the label printer wants its own format, the till and the website keep separate stock figures, and accounting gets a spreadsheet somebody retyped. When a standard appears that cuts that bridge from months to hours, it eventually reaches ordinary business software too. Worth remembering when you choose a system: closed products that cannot talk to anything else are a riskier bet than they were a year ago.
Second: the direction of travel is from advice to action. Until now the model wrote text and a person carried it out. Here the model moves a robotic arm. The same logic is already entering business software — from "here is a suggested purchase order" to "the order has been placed". Good news for productivity, bad news for companies whose data is untidy, because an automated mistake multiplies faster than a manual one.
Third, and most important: the foam story is the most useful part of the whole announcement. The model was confident, well-reasoned and wrong — it looked for a bug in the code while the problem was in the vessel. That is what an AI error looks like in practice: not obvious nonsense, but a convincing answer inside the wrong frame. It is why everything we build has to be able to say "I do not know", with a person left standing wherever the decision costs something. That is not distrust of the technology — it is the only way to use it seriously.
One sober note: this is a research preview, not a product. The figures come from labs working with the standard's author, using equipment worth more than a small company's annual turnover. It is interesting and it matters, but it is not something you buy tomorrow.
The takeaway
Not a race for the newest thing, but two calm decisions: keep your data in order, because every future automation runs through it, and choose systems that can connect. Companies with both will use whatever tool comes next. Companies without will buy tools that have nothing to work with.
Source: Anthropic — Previewing the Model Hardware Standard.
We do AI marketing, chatbots and business software — with a clear line between where AI helps and where a person decides.
Related services and pages
More from the blog
- What does a digital marketing agency do? (clear and concrete)
- Digital marketing for small businesses: where to start
- How to choose a digital marketing agency (question checklist)
- How much does Google Ads cost in Serbia?
Need a partner, not another PDF?
Rezident works as an external marketing team for SMEs — Google Ads, Meta, SEO, web and automation.
Book a call