Anthropic is trying to give AI agents a more practical path from software tasks into the physical world. Its new Model Hardware Standard, or MHS, is designed as a unified interface for equipment such as microscopes, robotic arms, liquid handlers, plate readers, monitoring cameras, and other programmable devices used in labs and factories.
The goal is straightforward: make hardware easier for AI agents to discover, understand, and operate without rebuilding custom connections every time different machines need to work together. Anthropic says the approach can reduce integration work from weeks or months to hours or minutes, though the early results also show why expert oversight remains central.
A common layer for physical hardware
Research labs and factories often rely on equipment from different manufacturers. Those devices can use different APIs, data formats, and control software, which makes coordination slow and fragile. MHS is Anthropic’s attempt to create a shared standard across that messy hardware environment.
The idea follows the same broad direction as Anthropic’s Model Context Protocol, or MCP, which gives AI models a standard route to external data sources and tools. MHS applies that concept to physical machines instead of software systems.
Under the MHS approach, each device gets an MHS driver. That driver describes basic capabilities such as reading data and modifying data, while also making the device discoverable in a common format. It can also capture details that matter in the real world, including a robotic arm’s weight or its safety limits.
Users can supply some of that hardware information in natural language. MHS then converts it into a reference file that an agent can use. In principle, this lets an agent encounter a device it has not previously worked with, understand its available functions and limits, and operate it through one interface.
Why one reusable driver matters
The practical gain is not that hardware magically becomes simple. Someone still has to write a driver for each device. The difference is that once the driver exists, it can be reused instead of recreated for every new pairing of equipment, controller, or workflow.
That matters because physical automation often depends on several devices working in sequence. In a lab, for example, a process may involve moving material, measuring it, tracking the run, and responding when the result is not good enough. If every connection has to be custom-built, the setup burden can slow down experimentation.
Anthropic says MHS is model-agnostic and works with any device that has a programmable interface. Agents can coordinate multiple devices through the standard and then save workflows as conventional scripts. Those scripts can run without a language model once generated, which separates the design of the workflow from its later execution.
- MHS gives devices a common interface for AI agents.
- Drivers make machines discoverable in a shared format.
- Agents can coordinate more than one device in a workflow.
- Generated workflows can be saved as scripts that run without a language model.
What happened in partner tests
Anthropic has tested MHS with several partners, and the examples show both the promise and the current limits of the approach. At biotech company Genentech, Claude coordinated a liquid handler, a robotic arm, and a plate reader to automate a protein assay, a test used to measure protein concentration.
In that run, the agent optimized pipetting parameters for different liquids by itself. But when bubbles formed in a viscous solution and produced errors, Claude treated the problem incorrectly. It repeatedly restarted the process in the same vessel while changing parameters, which made the issue worse. A human had to explain the physical cause before the agent identified a workable fix.
At Carnegie Mellon University, researchers connected a liquid handler, a plate reader, a robotic arm, and several monitoring cameras across three computers with fundamentally incompatible interfaces. The team says the drivers and orchestration layer took about eight hours to build, compared with the several weeks a vendor setup would typically require.
Claude also responded to experimental feedback in that project. After noticing that a first run had not produced a good enough measurement curve, it lowered the maximum concentration of the substance being tested and achieved a much better result on the second try.
At quantum computing company QuEra, Claude used MHS across hundreds of automated runs to develop a control program for restoring a laser to a stable operating state after disruptions. The resulting script succeeded in 695 out of 700 attempts in a blind test, a 99.3 percent success rate, and ran entirely on its own with the language model out of the loop.
Those examples are notable, but they come with an important caveat: all three are partner projects presented by Anthropic and have not been independently verified.
The hard part is still the physical world
MHS addresses the interface problem, but it does not remove the difficulty of reasoning about physical cause and effect. Anthropic says Claude learns about the physical world through text and images, and that leaves limits in spatial and physical reasoning.
The Genentech example makes that limitation concrete. The system could tune parameters and control equipment, but it did not correctly infer that foaming samples were a physical failure rather than a software issue. Human expertise was required to interpret the situation and redirect the agent.
That distinction is important for any AI hardware standard. A shared interface can make devices easier to connect, and an agent can help generate useful automation. But when equipment interacts with materials, motion, sensors, and safety constraints, the consequences are no longer purely digital.
During the research preview, Anthropic plans to build additional safety evaluations with partners and is developing a physical safety roadmap to guard against misuse. The company developed MHS with the HHMI Janelia Research Campus and is releasing it first as a research preview for select labs and manufacturers. An open-source release is planned later.
Where MHS could go next
If the approach holds up beyond early partner projects, MHS could make it easier for AI agents to operate in labs, factories, and other physical systems. The value would come from reducing the repeated integration work that currently makes mixed-device automation difficult.
Several manufacturers and organizations are already building MHS support or testing the spec. The list includes AWS, Doosan Robotics, QIAGEN, Tecan, Universal Robots, Hugging Face, and Raspberry Pi.
The larger story is not that AI agents are ready to run physical systems without people. It is that Anthropic is trying to standardize the connection between agents and machines, while acknowledging that the physical world still requires oversight, safety work, and expert judgment.