Anthropic Wants Its AI Agents To Operate Physical Devices, Equipment

Anthropic has opened a research preview of its Model Hardware Standard (MHS), a proposed shared specification that allows AI agents to operate programmable physical equipment across scientific laboratories and manufacturing environments. 

The company said the standard can connect devices including microscopes, liquid handlers, and robotic arms, allowing agents to coordinate multiple instruments and run workflows such as drug-discovery experiments and laser calibration for quantum computers.

MHS was developed initially through a collaboration between Anthropic and HHMI Janelia Research Campus. 

Anthropic said integrating equipment in a laboratory or manufacturing facility can typically take weeks or months because individual devices often use different programming interfaces and require bespoke integrations. 

MHS is designed to reduce that integration work to “hours or minutes”, according to the company.

The standard introduces a common hardware driver that translates between an operating system and a physical device. It uses basic commands such as “read” and “write” to allow software to retrieve information or change device settings. 

Devices can also be described in a standard format, allowing AI agents to discover available equipment and interact with it without a separate translator for every device.

MHS also gives agents information about equipment's physical characteristics and operating limits. Users can provide this information through natural-language tags, which the system converts into a reference file describing what a device can measure, what can be adjusted and which safety limits apply.

Once connected, devices can be controlled through the Model Context Protocol (MCP), a command-line interface, or code files. 

“MHS works with any device that has a programmable interface,” Anthropic said in a release. The standard is “model-agnostic”, allowing different AI agents and harnesses to access supported equipment through standard protocols.

Anthropic said agents can sequence operations across several instruments, monitor results, and adjust parameters in real time. 

For longer or repetitive tasks, agents can turn what they learn into scripts, allowing the equipment to run the process without the AI making every decision.

In one test, Anthropic said Claude adjusted a laser, used a camera to observe the resulting beam movement and repeated the process to understand the relationship between its actions and the hardware. Claude then converted what it learned into a script that could align the laser through a single command.

Anthropic is initially making MHS available to a group of scientific research labs and advanced manufacturers across biotech, robotics, quantum computing, and other fields. Companies including Amazon Web Services, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, and Universal Robots are working on or testing MHS integrations.

AWS, for example, plans to support MHS through Strands Robots, while Tecan is adding support for its Fluent liquid-handling platforms. Universal Robots has received early access, and Doosan Robotics is testing the standard with its robotic arms.

Anthropic will use the preview to develop safety evaluations and deployment practices before making MHS open source. 

However, the company acknowledged that Claude’s physical and spatial reasoning remains limited and that expert oversight is still required. In one example, researchers from Genentech had to guide Claude to recognise that foaming in protein samples represented a physical failure rather than a software problem.