Anthropic unveils new standard letting AI agents operate physical hardware

Anthropic has introduced a new technical standard called the Model Hardware Standard (MHS), designed to let AI models interact directly with physical equipment, ranging from laboratory instruments to robotic arms.

The company showcased several examples of the standard in action. In one, a Claude model adjusts a laser, checks the outcome through a separate camera, and repeats the process to automatically calibrate an entire system. In another scenario, MHS would let an AI model focus a microscope, analyze what it observes, determine which area needs closer inspection, and then move the microscope there on its own. Anthropic also released a video showing Claude figuring out how to direct a robotic arm to pick up an aluminum can, despite not having been specifically trained for that task.

How the standard works

According to Anthropic, MHS includes a standardized tagging system that describes a device's real-world physical constraints for models mostly trained in virtual environments. This covers details such as a robotic arm's weight and range of motion, along with its adjustable settings, available measurements, and built-in safety limits. That information can be compiled into a reference file, giving an AI model fast access to key facts about equipment it has never encountered before. MHS also enables models to write and adjust API scripts to sequence actions across multiple instruments, avoiding the need to reason through every step from scratch each time.

Anthropic says it is currently working with an initial group of research labs and manufacturers during a preview phase, including Amazon Web Services, Hugging Face, Raspberry Pi, Automata, and Universal Robots. These partners will help shape safety evaluations and best practices for AI systems that operate physical equipment. Longer term, Anthropic intends for MHS to become an open-source standard that works across different AI systems, not just its own.

The company reports that a year of early testing with scientific partners showed MHS cutting the time needed to integrate new devices, allowing faster iteration across various experimental setups.