Anthropic introduces MHS — a standard for AI agents to control physical equipment

Until now, agentic AI systems have worked mostly inside the computer — with text, images, code and data. Anthropic proposes taking them out into the physical world with the Model Hardware Standard (MHS) — a set of standardized drivers meant to let AI agents connect to arbitrary devices and control them.

As the company envisions it, MHS gives instruments a common interface and a single data-exchange format so they can “talk” to one another over the network without a separate translator program between them. For now the project is positioned first and foremost as a tool for scientists: an experimental setup that used to take weeks and months of painstaking work on custom integrations can, according to Anthropic, be cut down to “hours or minutes.”

The idea, as Anthropic's Alec Kemeny tells it, came from watching neuroscientist Arko Basta at work on the Janelia (HHMI) research campus. Basta had built a common interface to coordinate rotating lasers, microscopes, cameras and a host of other components in a memory-formation experiment. “You could use this idea to have AI run any scientific experiment in the world,” Kemeny recalls thinking.

The lab-assistant robot

The common machine interface on its own does not require AI: MHS devices can be controlled directly through the command line and an API. But when paired with a model via the Model Context Protocol, scientists can work with the instruments in natural language, while the model can “reason” at each step of the experiment, change parameters in real time and, in some cases, recover from hardware failures on its own.

As examples, Anthropic cites a Claude-level model that tunes a laser, checks the result through a separate camera and repeats the process, automatically calibrating the system; and one that focuses a microscope, analyzes the image and moves it to the area of interest. In a demonstration video, Claude also figured out how to make a robotic arm pick up an aluminum can, even though it had not been specifically trained to do so.

The standard also builds in a system of tags describing the real-world constraints of hardware for models trained mostly in virtual environments: the weight and reach of a manipulator, adjustable parameters, available measurements and hard safety limits. This data is gathered into a reference file that quickly brings the model up to speed on an unfamiliar device.

During the preview stage, Anthropic is working with an initial group of research labs and manufacturers — the named partners include AWS (Strands Robots), Hugging Face (LeRobot), Raspberry Pi, Automata and Universal Robots. They will help build out safety evaluations and practices for AI working with physical equipment. Down the line, MHS is meant to become an open, “agent-agnostic” standard.

“If you can test hypotheses faster, you can build general-purpose technologies faster,” Kemeny sums up. “That way a century of progress could fit into a decade.”

Source: Ars Technica