Anthropic Opens Preview of Standard Letting AI Agents Run Lab Hardware

SAN FRANCISCO —

Anthropic said Tuesday it is opening a research preview of a new system, called the Model Hardware Standard, that lets artificial intelligence agents operate physical lab and manufacturing equipment. The company is offering early access to a first group of research labs and manufacturers, according to its announcement.

The standard is meant to let AI agents run multiple instruments at once, including microscopes, liquid handlers and robotic arms, Anthropic said. It described tasks ranging from routine drug discovery experiments to calibrating lasers on a quantum computer. Anthropic said it built the system in collaboration with HHMI Janelia Research Campus, a biomedical research center.

How the system works

Anthropic said connecting lab devices to each other and to software has traditionally required custom engineering, since most instruments use their own separate programming interfaces. The company said setting up a lab or manufacturing floor to work this way typically takes weeks or months.

The new standard is built around what Anthropic called a standardized driver, which is software that translates between a device and a computer's operating system. The driver uses simple commands such as "read" or "write" so that any compatible device can be controlled the same way, according to the announcement. Anthropic said the driver also stores plain-language information about a device, such as the weight of a robotic arm, so an AI agent can figure out how to operate equipment it has never used before. Anthropic said the standard works with any device that has a programmable interface and can be used by different AI models through common protocols, including one called the Model Context Protocol.

Early testing and open questions

Anthropic said it shared early versions of the system with a handful of labs and manufacturers in biotech, robotics and quantum computing before this wider preview. The company said researchers at Genentech, the biotechnology firm, tested the standard as a way to automate a protein assay, a standard lab test, though Anthropic did not say what the results showed.

Anthropic said its own Claude models, in testing, adjusted a laser, checked the results through a camera and repeated the process the way a scientist might, then wrote a script so future adjustments would not require step-by-step reasoning. The company did not release the system as open source, saying it wants to work with partners on safety evaluations first. Anthropic did not say when it expects the standard to be more widely available or how it plans to prevent an agent from causing damage to hardware it is controlling.