Frontier AI lab Anthropic on Thursday announced its efforts to work on a standard for connecting AI agents with hardware. The Dario Amodei-led AI company introduced a new interface that will make it easy for AI agents to operate and communicate with machines.

The company has introduced a research preview of what it calls the Model Hardware Standard (MHS). According to Anthropic, it is designed to work with any device that has a programmable interface, including those deployed in scientific research and manufacturing. Simply put, MHS is Anthropic’s attempt to create a common language between AI agents and physical machines.

“MHS enables AI agents to operate multiple lab and manufacturing instruments, such as microscopes, liquid handlers, and robotic arms, in parallel and perform intricate tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer,” Anthropic said in its blog.

Story continues below this ad

At present, a research lab could use a microscope from one company, a robotic arm from another, liquid handling equipment from another, and a camera, all running on entirely different software. However, getting them to work together can take weeks or months of customisation. According to Anthropic, this is where MHS comes in, making it much easier.

Instead of an AI model having to learn a different interface for every machine, MHS endows each device with a standard software interface. A machine can essentially tell the AI what it can do, what it can measure, what settings can be changed, and its safety limits. According to Anthropic, the AI can discover and operate the machine without engineers needing to build a completely new integration system from scratch.

The development is significant, as it does not simply mean AI can control a robot. Such a system shows that AI can potentially coordinate multiple machines at once. For example, in a drug discovery experiment, the AI agent can work with a liquid handler, robotic arm, microscope/reader, and data analysis. The AI could start an experiment, examine results, change parameters, run the next step and potentially recover from certain errors. This moves AI from being a tool that suggests what a user should do to something that can execute a physical workflow.

In the announcement blog, Anthropic cited some interesting use cases of this. At a US-based biotechnology company, Genentech, Claude coordinated a liquid handler, robotic arm and plate reader while optimising how different liquids should be handled. Similarly, at Carnegie Mellon, researchers used MHS to connect several pieces of laboratory equipment. The team said that an integration that would usually take weeks took about eight hours, and the eventual experiment ran roughly three times faster. At the pioneering research centre, HHMI Janelia, MHS was used to connect a complex microscopy setup involving equipment from multiple companies. With MHS, a researcher said that adding a new camera took only minutes rather than requiring days of integration work.

Story continues below this ad

It needs to be noted that this is not a robot scientist that understands the physical world perfectly. Anthropic’s own experiments showcased some of its limitations. According to the blog, when bubbles appeared during liquid handling, Claude initially responded by restarting the process. And this actually made the problem worse. Later, a human had to explain the physical cause of the failure. The AI could then incorporate that knowledge and handle similar situations better. In essence, the challenge is not simply ‘Can AI control machines?’ but Can it safely understand what is happening in the physical world?

However, MHS could make some spaces like laboratories, factories and other physical environments AI-native to start with. It could change the present setup where humans interact with software to operate machines to humans interacting with AI agents that could manage multiple machines. In simple words, a scientist can describe the goal in natural language, while the AI can manage the coordination, monitoring, and repetitive experimentation. Anthropic has also shared its plans to make the standard open source after this research-preview phase.  Also, the system is designed to be model-agnostic rather than limited to Claude.