Technology and AI
A standard has been proposed to let AI agents drive laboratory and factory machines
By Staff Writer | 31 August 2026
A research preview opened on 27 August of the Model Hardware Standard, a shared specification through which software agents operate physical instruments. It is limited to a first group of research laboratories and advanced manufacturers, and the company says it will be open sourced once safety evaluations have been built with those partners.
The problem it addresses is not intelligence. It is wiring. A laboratory or a production cell typically runs instruments from several manufacturers, each with its own programming interface and none of them able to talk to the others, so a specialist writes a bespoke connection for every pair. Anthropic, which published the standard, puts the setup and integration work at weeks and sometimes months, and says the standard brings it down to hours or minutes.
The mechanism is a standardised driver, which is the software layer that translates between a computer's operating system and a device. The driver reduces control to a small set of commands, read and write, so that any device can act on them, and it makes each device discoverable in a common format so that instruments and agents can find one another across a network without a translator sitting in between.
We built this for science to sort of show the promise of AI, but there's also huge benefits here for enterprise and for industry
Elizabeth Kelly, head of beneficial deployments at Anthropic
What the driver carries beyond the plumbing
The part worth reading twice is the tags. The driver holds a description, written in ordinary language by the operator, of things about the machine that cannot be worked out from its code: the weight of a robot arm, what it can measure, what can be adjusted, and what safety limits will be enforced. From those tags the driver produces a reference file that tells the agent how to use a device it has never met.
Until now that information sat in a paper manual, on somebody's computer, or in nobody's hands at all. A standard that asks an operator to write down what a machine weighs is a standard that has noticed the answer was never recorded anywhere.
Early users describe what it bought them. Researchers at Carnegie Mellon University ran serial dilution dose response experiments about three times faster than before, with an agent directing a liquid handler, a plate reader, a robotic arm and monitoring cameras spread across three computers whose interfaces were incompatible. QuEra Computing, which builds quantum machines, gave an agent control over part of the laser system and reported a controller that recovers the laser's lock, the precise frequency the lasers must hold, 99.3 per cent of the time without a person intervening. Researchers at Genentech used it to automate a standard protein assay across a liquid handler, a robotic arm and a plate reader.
The limits are stated by the vendor
The announcement is unusually plain about what does not work yet. The model learns about the physical world through text and images, so its spatial and physical reasoning still requires expert oversight. The example given is from the protein assay: researchers had to teach it that errors caused by foaming in the samples were physical failures rather than software bugs, and could only be fixed physically. The standard also does nothing for equipment that has no programming interface at all, which on most sites is most of the equipment.
Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots are named as adding or testing support. Hugging Face and Raspberry Pi are named as early adopters for the next phase. The specification is model agnostic, so a plant that adopts it is not tied to one supplier's models, which matters more to a procurement officer than any of the laboratory results.
For anyone who writes or administers a contract for plant and equipment, the question this opens is old and familiar in new clothing. If an agent moves the arm and the arm damages the workpiece, the answer turns on whose specification governed the interface, what the safety limits said, and who wrote the tags. On the evidence of the announcement, that last one is an operator with a keyboard and no obligation to be right.