AI Is Stepping Off the Screen and Touching the Physical World: Anthropic's Model Hardware Standard

When we talk about artificial intelligence, we usually encounter the same image: a computer screen, a chat box, and a person typing something into it.

Over the past few years, AI development has largely progressed within this framework. Especially with the emergence of AI agents, the model has stopped being just a chatbot that answers questions. It can access files, use the terminal, open websites, run code, and interact with other software. But there was an important boundary here: How will artificial intelligence escape the computer?

Connecting to a software API is one thing. Controlling a robotic arm, microscope, or liquid handler in a laboratory is quite another. Anthropic's announcement yesterday, August 27, 2026, of the Model Hardware Standard (MHS) focuses precisely on this problem. MHS is not yet a finished product; it is currently in research preview mode and is being tested with scientific research laboratories and advanced manufacturing companies. Anthropic's plan is to open-source the standard after development. And I think the most interesting part of this announcement is far more compelling than the sentence "Claude can now operate robots."

Let's Start with a Small Problem

Imagine you have a laboratory. Inside it you have a robotic arm. You have a liquid handler. You also have another device that analyzes samples. Normally, each could come from a different manufacturer. One has a Python library. Another has a different programming interface. Yet another device can only be controlled through the manufacturer's own software. So from a computer's perspective, there aren't three devices, but three separate worlds. This situation is actually similar to a problem we've known in the software world for years: integration chaos. One company's CRM works differently, the payment system another way, the email service yet another way. When we put a common standard between them, things become easier. On the internet today, we take this very much for granted. For a website to connect to a payment system, a maps service, or an email service, we don't need to learn how each service works internally from scratch. What Anthropic is trying to do with MHS is to take this similar approach and apply it to physical devices.

What is Model Hardware Standard?

In the simplest terms:

MHS is a standard developed so AI agents can communicate with physical devices through a common method. The word "standard" here is important because Anthropic is not creating a new robot. Not creating a new microscope. Not creating a new laboratory device either. It's trying to make different devices available to AI agents in a more understandable and standardized way. Roughly speaking, we can think of it like this:

Instead of agents learning each device's specific details, devices are defined in a common way. For example, for a device:

information can be presented in a standardized way. According to Anthropic's explanation, MHS can work with devices that have a programmable interface and is not dependent on any particular AI model. In other words, it's not designed as a closed system just for Claude. It's also possible for agents to access MHS through standard protocols like Model Context Protocol (MCP).

If You Know MCP, Understanding MHS Is Easier

For those following the AI world recently, there's a familiar concept here: Model Context Protocol (MCP). We can think of MCP in very simple terms as a common connection standard that makes it easier for artificial intelligence to communicate with different software tools. For example, an AI agent can:

MCP helps standardize this interaction. MHS, on the other hand, takes similar thinking and applies it to physical devices. So we can make a conceptual distinction like this:

MCP:

AI ↔ Software Tools

MHS:

AI ↔ Physical Devices

When we bring these two together, something interesting emerges:

AI ↔ Software ↔ Physical World

There's one thing we need to be careful about: it's not accurate to say MHS is officially defined as "MCP for hardware." This is an analogy we can use to explain what MHS does.

But it's a good analogy.

Because the real issue is making devices usable by AI.

So What Can AI Really Do in a Laboratory?

This is where things become even more interesting. Anthropic, in early tests of MHS with Genentech, shows different devices being used together in a laboratory environment. The goal here is not just for an AI agent to operate a single device, but to interact with multiple devices and conduct an experimental process. For example, results from a liquid handler can be used in the next experimental step. This matters because AI's role here doesn't stay at "turn device on and off."

The agent:

1. performs an operation,
2. observes the result,
3. evaluates the result,
4. determines a new parameter for the next operation,
5. tries again.

So the feedback loop we're used to in the computer world is being transferred to the physical world. In Anthropic's Genentech example, Claude optimized the liquid transfer rate by experimenting on water and viscous BSA samples. The system tested different speeds from approximately 140 µL/s down to approximately 10 µL/s levels, investigating suitable operating points. I think what matters here is not these individual numbers. What matters is this:

AI is no longer just a system that says "what should be done" but can now see the result of an operation and determine the next physical action based on that result.

This fundamentally changes the concept of agency.

A More Interesting Example: AI Improves the System Through Experimentation

Anthropic's example with QuEra is, I think, more interesting for understanding why MHS is a broader idea.

The problem here involves a recovery mechanism for relocking a laser system.

Previously, a special script had been developed for this, and the system could solve the problem in about 150 seconds with about 58% success rate.

When Claude accesses the system through MHS, the agent attempts to improve the recovery mechanism by repeatedly experimenting with different error scenarios.

As a result, a solution in about 6 seconds is obtained with about 96% success rate in tests. Later, success rate reaches 99.3% in a test with 700 attempts.

The truly interesting detail here is that the result is not a continuously running AI agent.

What the agent does is, in a sense:

"Let me experiment with the system, break it, run it again, look at the results, and find a better solution."

Subsequently, the solution that emerges becomes a more deterministic and verifiable script.

So AI doesn't have to be the final product here.

AI becomes a research tool used in developing better physical systems or software.

I think this distinction is quite important.

Why Is It Incomplete to Say "AI Controls Robots"?

Because what MHS brings is not just robot control.

A more general problem is being addressed:

Converting physical devices into standard tools that AI agents can use.

The robot arm is just one example of this.

The same approach can be used in:

Anthropic also notes that MHS has been tested with different devices in the fields of science, robotics, electronics, and manufacturing.

So we shouldn't get too hung up on the word "robot" here.

The real subject is programmable physical systems.

But There's a Small Problem Here: The Real World Is Not as Orderly as Computers

There's a serious difference between an AI agent clicking a button on a website and transferring a liquid to another container in a laboratory.

If there's an error on a webpage, you can usually try again.

In the physical world, however:

More importantly, expecting AI to understand all the consequences of the physical world perfectly from text and sensor data is not realistic today.

The limitations of this are evident even in Anthropic's own examples. The system can recover from some hardware errors, but it can't always correctly understand the causes of physical events. For example, it initially failed to correctly diagnose the physical cause of bubbles forming in a liquid transfer experiment.

For this reason, reading MHS as:

"Now AI manages the laboratory on its own."

would be a bit too optimistic.

A more accurate statement is:

Infrastructure is being built so AI agents can work with physical systems in a more controlled and standardized manner.

This is a much more modest statement.

But it's actually more important.

Because AI's Future Isn't Just About Better Models

When we talk about AI, we usually focus all our attention on the model itself.

A new model is released.

Better reasoning.

Longer context.

Better coding.

Faster agents.

All of these matter.

But for an AI agent to be useful in the real world, being "intelligent" alone isn't enough.

It needs access to tools.

Those tools need to be able to communicate with each other.

The safe boundaries of these tools need to be defined.

And physical devices need to be represented in a common way that AI can understand.

MHS touches this infrastructure layer directly.

For this reason, I think it's more correct to read Model Hardware Standard not as a new "AI feature" but as the expansion of AI agent infrastructure.

Can Physical Standards Today Do What MCP Does Tomorrow?

We can speculate a bit here.

Today, a software developer doesn't need to learn the internal architecture of every service to integrate with hundreds of different services.

Thanks to standards, we can use a common language.

If similar standardization occurs for physical devices, in the future it might be possible for an AI agent to access not just files or APIs on computers, but physical devices from different manufacturers.

A researcher could theoretically give a command like:

"Evaluate this experiment's results and plan the next experiment based on the results."

In the background, the agent would:

all together.

At this point, the AI agent is no longer just software that processes information.

It becomes a kind of digital research operator.

And perhaps AI's real transformation in the physical world will begin here, before humanoid robots.

Is This Usable Technology Today?

No, at least not at this point.

MHS is currently in research preview mode.

Anthropic is testing it first with scientific research laboratories and advanced manufacturing companies. The goal is not just to develop the standard; it's also to develop security assessments and best practices for AI systems working with physical equipment.

Anthropic says it plans to open-source the standard after this early phase.

So today we can't go and connect Claude to our home coffee maker and say "make me coffee at 8 AM."

Not yet.

But from a technical perspective, something more important is happening:

Work is beginning on the protocol layer between AI agents and the physical world.

Conclusion: The Real Change Might Not Be in the Model, but in the World the Model Can Reach

Model Hardware Standard's current form will likely change in the years to come.

The name might change.

The protocol might change.

The standard might be developed by other companies.

Different models might be used.

Even the approach Anthropic suggests could transform into something completely different over time.

These are all possible.

But the fundamental idea, I believe, is lasting:

We need common infrastructure for AI agents to use not just digital tools, but physical devices.

Just as MCP simplifies the relationship between AI and software tools, MHS could be a starting point for making physical devices usable by AI.

For this reason, I don't see Model Hardware Standard as a small Anthropic announcement.

I see it as one of the early signs of a broader transformation.

AI's first great era taught us to communicate with machines on screens.

The agent era enables machines to do some things on our behalf.

In the next phase, we expect machines to do things not just on computers, but in the real world.

And for that to happen, they first need to speak the same language.

Model Hardware Standard's real story might be exactly that.