Anthropic's Model Hardware Standard Brings AI Agents Into the Physical World

Posted on 2nd Sep 2026 06:04:47 in Artificial Intelligence, Machine Learning

Tagged as: Anthropic, Physical AI, AI Agents

For the past two years, AI agents have been busy conquering the digital world — writing code, answering support tickets, browsing the web, and automating workflows that live entirely inside computers. The physical world, however, has remained largely off-limits. Laboratory microscopes, industrial robots, and factory assembly lines each speak their own proprietary language, and wiring them up to an AI model has typically demanded weeks, if not months, of custom engineering by specialists. On August 27, 2026, Anthropic announced something designed to change that: the Model Hardware Standard (MHS), a research-preview specification that lets AI agents discover, communicate with, and safely operate physical lab and manufacturing equipment.

The announcement marks Anthropic's first formal move into so-called physical AI — the same arena where Nvidia, OpenAI, and Amazon have been spending billions. But instead of building its own robots or devices, Anthropic is betting on the strategy it has used successfully before: create an open standard that everyone can adopt, and let the ecosystem do the rest.

What Exactly Is the Model Hardware Standard?

MHS is best understood as a translation layer between AI models and physical devices. Every MHS-compatible device exposes a standardised driver built around a small set of universal commands — such as "read" (get a temperature reading) and "write" (set a temperature) — that any hardware device can understand and act upon. The driver also makes each device discoverable in a standard format, so agents and instruments can find each other across a network "without needing a bespoke 'translator' program in between," as Anthropic puts it.

The company compares the approach to hardware connectors we already take for granted. Alek Kemeny, a member of the technical staff at Anthropic, described the company's earlier Model Context Protocol (MCP) as "kind of like the USB for AI to software connection" in an interview with Fortune, and CNBC drew a direct parallel to the USB-C cord that standardised how information flows between devices. MHS extends that same philosophy from software to machinery.

Two design choices are worth noting. First, MHS is built directly on top of MCP, the open standard Anthropic debuted in 2024 for connecting AI models to data sources — which means any agent harness that already speaks MCP can now reach physical instruments through the same pathway. Second, MHS is deliberately model-agnostic. It works with Anthropic's Claude, with models from competitors such as OpenAI, and with open-source models, so adopting laboratories are not locked into any single vendor's AI stack.

MHS also includes a standardised tagging system that encodes a device's real-world constraints — things a text-trained model cannot see: the weight and range of a robot arm, its adjustable parameters, measurement options, and enforced safety limits. These natural-language tags are compiled into a reference file that quickly gives an AI model the crucial information it needs about equipment it has never encountered before.

From a Neuroscience Lab in Virginia to a Global Standard

The idea behind MHS was born not in a product meeting but in a working laboratory. Neuroscientist Arco Bast was running a memory-formation experiment at the HHMI Janelia Research Campus in Ashburn, Virginia, on a rig that combined rotating laser beams, microscopes, motorised focusers, and cameras from different vendors — none of which shared a common interface. Bast built a shared-memory "dictionary" so the instruments could coordinate, and Kemeny, watching the setup work, recalls thinking: "This idea could be used to have AI run any science experiment in the world."

Anthropic turned that observation into a demonstration of what MHS enables. In its announcement video, the company shows Claude adjusting a laser, checking the result through a separate camera, assessing how the beam moved, and then repeating the process until the system was automatically calibrated — the same exploratory loop a human scientist would perform. In another demo, Claude figures out how to guide a robotic arm to pick up an aluminium can even though the model had not been specifically trained on those steps.

Just as important is what happens after exploration. Once Claude understands a sequence, it packages what it learned into a deterministic script that can run the entire procedure as a single command — no model reasoning required at each step. The ability to move from exploration to automation within a single session is what Anthropic says makes MHS valuable for persistent, around-the-clock lab workflows. As Kemeny put it: "If you can test hypotheses faster, you could create general technologies faster. This is how a century of progress can condense into a decade."

The Early Partners and What They Are Testing

Anthropic developed MHS in partnership with HHMI Janelia, and a first group of research labs and manufacturers across biotech, robotics, and quantum computing received early access. The partner list announced so far includes:

• Genentech — ran a drug discovery experiment with real-time error handling on a standard BCA protein assay used to measure protein concentration
• AWS — supporting MHS through Strands Robots, its library for connecting AI agents to physical devices
• Universal Robots — adding MHS support to its collaborative robotics platform
• Doosan Robotics — testing MHS for automated quality assurance and multi-robot coordination
• QIAGEN — built a proof-of-concept on its nucleic acid purification platform to show how agents can troubleshoot instruments faster
• Tecan — adding MHS support for its Fluent liquid handling platforms
• Hugging Face — integrating MHS into LeRobot, its open-source robotics library
• Raspberry Pi — enabling MHS integration across a number of its products
• Carnegie Mellon University, quantum computing company QuEra, Danaher, and Doosan Robotics — among the other early adopters named by Fortune

The Genentech trial also surfaced one of the most instructive limitations of the current preview. When a sample began foaming during the assay, Claude treated the problem as a software bug until researchers explained that foaming was a physical failure. It is a telling illustration of the gap between digital reasoning and physical intuition — and a reminder that, for now, MHS deployments still need expert human oversight.

Why This Matters for Science, Manufacturing, and the AI Race

For scientists, the pain point MHS attacks is real. "Scientific equipment suffers from proprietary solutions that are very brittle and often don't meet the need of scientists," said Jonah Cool, head of partnerships and deployment of science at Anthropic. "We want to avoid vendor lock-in for scientists." A standardised, programmable interface means researchers can buy an MHS-enabled instrument, connect it to whatever model they prefer, and have it working out of the box — a process that previously took weeks or months and specialist engineers, now reduced to "hours or minutes."

For industry, the benefits may be even larger. "We built this for science to sort of show the promise of AI, but there's also huge benefits here for enterprise and for industry," Elizabeth Kelly, head of beneficial deployments at Anthropic, told CNBC. Anthropic frames MHS as the enabling layer for autonomous, round-the-clock experiments and workflows on factory floors and production lines.

The competitive context is equally significant. OpenAI and Amazon have poured billions into AI-native devices and manufacturing tools, while Nvidia CEO Jensen Huang predicted in March that "every industrial company will become a robotics company." Anthropic itself is moving deeper into hardware — it is building a silicon team to design custom chips and recently hired hardware executive Caitlin Kalinowski, formerly of OpenAI, Meta, and Apple. Tellingly, the MHS announcement landed on the same day Hugging Face debuted its first physical AI product, a robotic duck built on LeRobot, and amid Nvidia's roughly $13 billion agreement to acquire Hugging Face.

Anthropic's play, however, is different: rather than owning the robots, it wants to own the interface. MHS is available now as a research preview to a select group, and the company plans to open-source the standard after collaborating with partners on safety evaluations and best practices — the same playbook it followed with MCP, which has since been adopted across the industry.

The Limits — and What Comes Next

MHS is not yet a universal answer. Hardware without a programming interface cannot connect out of the box, and Anthropic is working with device manufacturers to add the necessary interfaces to new products and retrofit existing ones. Claude's physical reasoning, learned through text and images, still requires expert supervision — the Genentech foaming episode being the clearest example. And the standard's safety story is still being written: the research preview exists precisely so the first partner cohort can help "build safety evaluations and develop best practices for AI systems operating physical equipment."

If the open-sourcing plan proceeds as promised, the long-term implications are significant. A common, agent-agnostic interface between AI and machinery would lower the cost of automating labs and factories the same way USB lowered the cost of connecting peripherals to computers. The AI agents that spent 2025 and 2026 mastering software may spend the next few years learning to turn knobs, move arms, and run experiments — one standardised command at a time.

Sources

• Fortune — Anthropic makes first move into physical AI with new way for scientists, manufacturers to bring equipment to life
• CNBC — Anthropic pushes into physical world with new standard to help AI agents operate machines
• Ars Technica — Anthropic's new hardware standard lets AI agents control the physical world
• VKTR — Anthropic's First Physical AI Standard Lets Agents Operate Real-World Equipment