For years, the AI boom has largely played out on screens: chatbots writing emails, coding assistants building software, and image generators making pictures from a text prompt. Anthropic is now pushing that idea beyond the browser with a new effort designed to let AI agents work with actual laboratory and factory equipment.

The company has opened a research preview of the Model Hardware Standard, or MHS, a shared technical specification that could allow AI systems to safely discover, monitor, and operate programmable devices such as microscopes, liquid handlers, cameras, lasers, plate readers, and robotic arms. The early goal is less about putting a chatbot in charge of a lab and more about fixing a very old, very unglamorous problem: scientific instruments often do not speak the same software language.

In a modern research lab, a single experiment may involve equipment from several vendors, each with its own control software, data formats, and quirks. Connecting a microscope to a robotic arm, then feeding results into a camera or a liquid-handling machine, can require weeks or even months of custom work from specialists. Anthropic says MHS is meant to bring that setup time down to hours or minutes by giving hardware and AI agents a common way to communicate.

That might sound like infrastructure plumbing, but it is potentially one of the more consequential ideas to emerge from the current AI-agent race. If it works as intended, MHS could turn fragmented collections of expensive equipment into something closer to a coordinated, semi-autonomous system – one that can run experiments, watch for errors, react to incoming measurements, and continue working overnight under human supervision.

A common language for machines

Anthropic describes MHS as a standardized driver layer. In plain English, it acts as a translator between AI software and physical devices that were not originally designed to work together.

Instead of making every instrument use a custom interface, MHS relies on simple shared actions such as “read” and “write.” A “read” request could ask a device for its temperature, position, sensor result, or camera feed. A “write” request could tell that same machine to change a temperature setting, move an arm, start a process, or adjust a laser.

The larger idea is that an AI agent should be able to identify available devices on a network, learn what each one can measure or control, understand its limits, and then coordinate those tools as part of a broader workflow. Anthropic says MHS works with devices that already have programmable interfaces and can be accessed through the Model Context Protocol, command-line tools, or code-based APIs. It is also designed to be model-agnostic, meaning the standard is not supposed to be exclusive to Claude.

That distinction matters. Anthropic is not simply announcing a feature for one of its own AI products. It is trying to establish a common framework that equipment makers, robotics platforms, lab-automation firms, and developers could adopt across the industry. The company plans to eventually make MHS open source, but it is starting with a limited research preview involving scientific labs and advanced manufacturers.

Giving AI a sense of physical limits

The difficult part of using AI in a lab is not just giving it access to a button labeled “start.” Physical systems carry consequences that digital systems do not. A software mistake can cause an error message. A mistake involving a robotic arm, fragile sample, high-powered laser, or chemical process can damage equipment, ruin an experiment, or create a safety risk.

MHS is intended to give agents more context about the equipment they are controlling. That can include information such as a robot arm’s weight, range of motion, adjustable settings, measurement capabilities, and safety limits. Anthropic says this kind of knowledge is often buried in paper manuals, scattered across local files, or held by experienced operators rather than captured in a format software can use.

The standard lets users attach natural-language tags to equipment so an agent can better understand those real-world constraints. From there, MHS can generate a reference file describing the machine and the boundaries within which it should operate. The aim is to keep an AI agent from treating a physical device like an abstract software tool with unlimited room for trial and error.

Even with those guardrails, Anthropic is clear that this is not hands-free science. The company says large language models still have limits in spatial and physical reasoning because they learn about the world largely through text and images. In one example, Genentech researchers had to help Claude distinguish foaming in protein samples – a physical problem – from a software fault.

That is a useful reality check. MHS may make machines easier to coordinate, but it does not eliminate the need for scientists, technicians, and engineers who understand what can go wrong in the real world.

From assistants to lab operators

The most interesting possibility is not that an AI agent can switch on a microscope. It is that the system can connect multiple tools into a feedback loop.

Anthropic describes a scenario where an AI adjusts a laser, checks the result through a camera, evaluates whether the beam moved in the intended direction, and repeats the process until it reaches the right calibration. Once it has found a reliable sequence, it can turn that sequence into a deterministic script that runs as a single operation rather than requiring the model to reason through every action again.

The same basic approach could apply to microscopy. An AI could focus the instrument, assess the images, identify an area that needs closer inspection, reposition the microscope, and continue the process. Ars Technica noted that the system could help agents move from simply analyzing data after an experiment to actively shaping the next step while the experiment is underway.

Anthropic says MHS can also support long-running operations and faster device-level actions by chaining commands into code files. That matters because language models are not built to make every millisecond-level control decision themselves. In practice, the AI may act more like a high-level planner and supervisor, while the machines follow faster, more predictable control routines underneath.

This is the difference between an AI agent narrating what should happen and one helping orchestrate what actually happens. It is still a long way from a fully autonomous scientist, but it is a meaningful step toward AI systems that can participate in experimental work rather than just summarize papers about it.

Early tests are promising

Anthropic’s early partners include major names across biotech, robotics, cloud computing, quantum computing, and lab automation. The company says Genentech used MHS in a proof of concept for automating a BCA protein assay, a common method for measuring total protein concentration that required coordination among a liquid handler, robotic arm, and plate reader.

Researchers at Carnegie Mellon University used the framework for serial-dilution dose-response experiments and reportedly completed those experiments about three times faster than before. The setup involved an AI agent coordinating a liquid handler, plate reader, robotic arm, and cameras running across three computers with incompatible interfaces.

At HHMI’s Janelia Research Campus, where the project originated, researchers are using MHS to bring together microscopy systems that previously required seven different vendor programs with no common interface. The roots of the project trace back to a complex brain-imaging setup involving lasers, motorized focusers, and specialized cameras from different vendors.

One of the flashier demonstrations comes from QuEra Computing, which builds neutral-atom quantum computers. Anthropic says an AI agent using MHS developed a controller that recovered a laser’s precise operating “lock” 99.3% of the time without human intervention. Maintaining that laser frequency is essential for manipulating atoms in QuEra’s quantum systems.

These are early, tightly scoped deployments, not evidence that labs are ready to hand over their operations to general-purpose AI. Still, they show where the value may lie: faster integration, more responsive troubleshooting, and less time spent manually stitching together systems that should already be able to share information.

Why the standard matters

AI companies have spent the past two years competing to make agents more capable at computer tasks. They can browse websites, fill in forms, write code, search files, and use business software. MHS takes that agent concept into environments where the output is not a spreadsheet or an email, but a physical action.

That shift makes interoperability especially important. A laboratory cannot realistically rebuild all of its existing workflows around one AI vendor, one robot maker, or one instrument supplier. If MHS gains traction, its biggest advantage may be that it creates a common interface sitting between AI systems and a messy mix of old and new hardware.

Anthropic is working with organizations including AWS, Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan, Universal Robots, Hugging Face, and Raspberry Pi. Some are building MHS support into their platforms or drivers, while others are exploring uses in quality assurance, autonomous laboratories, instrument recovery, real-time analysis, and robotics development.

For manufacturers, the appeal is straightforward. Equipment that is easier to discover, configure, and automate could become more useful to customers without requiring every lab to build a custom software layer from scratch. For researchers, the payoff could be more time spent designing experiments and interpreting results, rather than wrestling with disconnected vendor software.

The safety question

The obvious question is whether AI should be trusted with physical equipment at all. Anthropic’s answer, at least for now, is that it should be tested carefully in narrow settings, with strong constraints and expert oversight.

The company says it will use the research-preview phase to build safety evaluations and develop deployment best practices with its partners. It is also working on a physical-safety roadmap intended to strengthen how it handles misuse risks involving AI-controlled hardware. Findings and guidance from the preview are expected to be released when MHS is eventually open-sourced.

That caution is necessary. An AI model can be surprisingly competent at identifying patterns, following instructions, and adapting a plan. It can also misunderstand ambiguous instructions, make flawed assumptions, or behave unpredictably when a situation falls outside what it has seen before. A typo in a lab configuration file is not the same as an AI agent moving a robot into the wrong place.

For now, Anthropic’s announcement is best understood as a foundation rather than a finished product. MHS will not instantly turn every research facility into a lights-out autonomous lab. Many machines still lack programmable interfaces, and the technology will need wide adoption from hardware vendors before it becomes a genuinely universal standard.

But the direction is clear. AI agents are gradually moving from tools that only manipulate digital information to systems designed to observe, plan, and act in the physical world. Anthropic’s Model Hardware Standard is an attempt to make that transition less chaotic – and, crucially, safer – before autonomous systems start taking a larger role in the labs and factories where real-world discoveries are made.