Anthropic wants AI agents to control lab machines, giving Claude hands to work

Anthropic wants AI agents to do more than just talk. Its new Model Hardware Standard could let Claude operate lab machines, robots and other physical hardware. Here's how it works.Anthropic unveils Model Hardware Standard.The buzz around AI seems to be shifting to AI agents. And they are doing pretty well using computers. They can write code for you, browse the web, analyse documents and even carry out multi-step tasks without you holding their hand. But the AI innovation doesn’t end here. Anthropic now want to take AI agents a step further, and that is, let the AI agents operate the physical machines used in laboratories and advanced manufacturing. That could mean an AI controlling a microscope, moving samples with a liquid handler, adjusting a laser or even working with a robotic arm.In a recent blog post, Anthropic revealed a research preview of its Model Hardware Standard (MHS), a framework designed to give AI agents a common way to interact with physical devices. The company says MHS can allow multiple machines to work together, while giving researchers a way to control what an AI agent can and cannot do with the hardware.The idea sounds a little like giving Claude a pair of hands. Instead of simply telling a scientist what to do, an AI agent could actually operate the equipment, watch what happens and then decide what to do next.How the Model Hardware Standard worksMost of the laboratories and manufacturing machines don’t really speak the same language. To put it more simply, each device generally has its own software and programming interface. So, if one wants to connect several machines together, that would require a specialist to build these custom integrations. And that is what Anthropic seems to solve. Anthropic says something that can take weeks or even months could potentially be reduced to hours or minutes with MHS. Image credit: Anthropic The standard introduces a common software driver that acts as a translator between an AI agent and a physical device. It uses simple commands such as "read" to get information, like temperature, and "write" to change something, such as setting a temperature.More importantly, MHS can also give an AI information about a machine that may not be obvious from its code. Think of things like how much weight a robotic arm can safely handle or what limits a particular instrument has. Researchers can describe these characteristics in natural language, and the system creates a reference file that tells the AI what the device can do and what safety limits it needs to respect. Once the machines are connected, the AI can coordinate them through MCP, a command-line interface and code files. It can sequence different steps, monitor results and change parameters as an experiment progresses. For longer tasks, it can even turn what it has learned into code so the equipment can continue working without the AI having to reason through every individual step.Anthropic says it has already seen Claude use this approach in experiments. In one example, Claude adjusted a laser, watched the result through a camera and repeatedly changed the settings until it understood how the laser responded. It then turned what it learned into a script that could align the laser with a single command.Anthropic wants AI to become a lab workerAnthropic wants MHS to be a part of the physical research process. The company says early tests with labs and hardware manufacturers have shown that MHS can reduce the time needed to connect different devices, help researchers run experiments faster and assist with live machine operation and fault detection.But there is an obvious catch: physical machines can break things in the real world. An AI that makes a bad decision on a screen is annoying. An AI that makes a robotic arm move the wrong way or changes the settings on expensive laboratory equipment is a different problem altogether.That is why Anthropic is not releasing MHS as a finished, open-to-everyone system just yet. The company is sharing an early version with selected partners across science, robotics, electronics and manufacturing so they can develop safety evaluations and best practices before the standard is made open source.- EndsPublished On: Aug 28, 2026 10:10 IST