Anthropic has opened a research preview of its Model Hardware Standard, letting AI agents operate lab and factory hardware.
The specification, called MHS, is going first to a small group of scientific research labs and advanced manufacturers, not to the public. Anthropic says the standard lets an AI agent run several lab instruments in parallel.
Alek Kemeny, of Anthropic’s Beneficial Deployments team, began the MHS project with Arco Bast, a postdoctoral scientist at HHMI Janelia. Bast was running brain-imaging experiments on a rig that combined lasers and motorised focusers, and the setup also used cameras from different vendors with no shared interface.
Bast built a shared memory dictionary that let the instruments exchange data at memory speed. Kemeny then worked with him to connect AI models to that interface, a design that now underpins the MHS standard.
Anthropic’s preview of MHS shows where automation still needs a human check. Genentech researchers, for example, had to guide Claude to see foaming in protein samples as a physical failure and not a software bug, one that needed a physical correction instead of a code fix.
Claude’s spatial and physical reasoning carries limits, Anthropic notes, because the model learns about the physical world through text and images, not through direct sensing. Human oversight remains vital. That holds even in the deployments the company presents as successes.
Replacing one-off integrations with a shared driver
Most lab and factory devices do not talk to each other, so specialists have had to write one-off integrations for each new instrument, a process that can take weeks or months. MHS aims to cut that timeline to hours or minutes by introducing a standardised driver that sits between a computer’s operating system and a hardware device.
The driver runs on a small set of commands, such as “read” for retrieving a value like temperature and “write” for setting one, that any device can act on. Each connected device becomes discoverable in a standard format. Devices and agents can then find each other across a network without a separate translator program for every pairing.
The driver also stores information that a device’s code alone does not reveal. That includes the weight of a robotic arm, which affects how it can be moved without damage. Anthropic notes how that kind of detail has often sat in paper manuals or in the memory of one engineer.
Tags on each driver accept natural language input. A user can enter setup details directly, or answer questions posed by an agent. The driver then generates a reference file on its own. The file lists what a device can measure and also spells out what can be adjusted and the safety limits that apply, giving an agent what it needs to operate equipment that it has not encountered before.
Three routes for an AI agent to reach the hardware
A device becomes reachable through three routes once it is connected and described. Model Context Protocol, (MCP) is the first. A command line interface offers a second route. Code files that function as application programming interfaces provide a third, and according to Anthropic, the three combine to let one line of code direct several instruments at once.
The agent can then pull operating data from each device and sequence steps across instruments and monitor results. It can also adjust parameters as conditions change. Some tasks run long, while others need to happen faster than an agent can reason through step-by-step. The agent can chain driver commands into a code file in those cases, letting the device carry out the entire sequence on its own without further step-by-step reasoning.
Anthropic describes watching Claude work through a laser alignment problem. It resembled a scientist’s approach. The model adjusted a laser and checked the result through a camera, repeating the process several times to work out how the beam’s position had changed. Claude then wrote a deterministic script from what it learned. The script let the alignment run as one command, without further reasoning at each step.
Case studies show speed gains and where oversight still matters
Anthropic shared early versions of MHS with a handful of labs and hardware manufacturers across biotech, robotics, and quantum computing before opening this wider preview. Partners reported faster device integration and quicker iteration during experiments, and some said MHS helped with live machine operation and real-time fault detection.
Researchers at Carnegie Mellon University used MHS to run serial dilution dose-response experiments about three times faster than before. An AI agent orchestrated a liquid handler and a plate reader for the work. It also directed a robotic arm and monitoring cameras, spread across three computers with incompatible interfaces.
QuEra Computing, which builds quantum computers using neutral atoms, gave an AI agent control over part of the laser system inside its machines. The agent developed a controller that recovers the laser’s “lock” (the precise frequency the lasers must hold to interact with atoms) 99.3 percent of the time without human intervention.
Hardware manufacturers are already building MHS support into their own products. Early adopters span robotics frameworks such as AWS Strands Robots and Hugging Face LeRobot, automation platforms such as Automata’s LINQ and Doosan Robotics’ arms, and scientific instrumentation makers including Danaher, QIAGEN, MBF Bioscience, and Raspberry Pi.
AWS will give research preview participants a private, pre-release version of Strands Robots for the duration of the programme. QIAGEN is running a proof-of-concept on its QIAsymphony Connect nucleic acid purification platform, aiming to show whether agents can help laboratories troubleshoot instrument issues faster and guide operators through recovery.
Physical safety work continues before wider release
Anthropic acknowledges that MHS does not yet work with hardware lacking a programmable interface, and the company is working with manufacturers of such equipment to build in drivers. Many developers already use Claude Code with individual pieces of physical equipment. Anthropic wants the next phase of MHS to cover more of the devices developers build on.
The company is developing a physical safety roadmap intended to strengthen its safeguards policy and enforcement against misuse. It plans to use the research preview to build further safety evaluations with launch partners.
Anthropic plans to publish findings from the preview as guidance for deploying the standard safely once MHS becomes open-source.
Learn more about physical AI during the Physical AI Expo held in Amsterdam, London, and North America.
See also: Generalist AI’s GEN-1.5 robot model learns tasks from one demo
Want to learn more about the IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including AI & Big Data Expo and the Cyber Security Expo. Click here for more information.
IoT News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
