How Anthropic's Model Hardware Standard Enables Agentic and Physical AI in Manufacturing

How could AI agents change the way manufacturers interact with industrial equipment? Jan Wokittel, Director of Smart Manufacturing at Roche, explores Anthropic's Model Hardware Standard, its relationship with existing industrial standards, and what manufacturers need to build today to prepare for the future of Agentic AI.

Artificial intelligence is rapidly changing how software is developed and how people interact with technology. AI agents can analyze large datasets, interpret documentation, generate applications, and execute increasingly complex workflows. But translating these capabilities into the physical world of industrial manufacturing introduces a different set of challenges.

Manufacturing equipment operates within highly specialized automation environments, supported by industrial communication protocols, proprietary software interfaces, and established control architectures. Accessing machine data, integrating systems, or executing operations often requires considerable engineering expertise.

What happens when AI agents can understand industrial equipment, discover its capabilities, and interact with it through standardized interfaces?

This is the opportunity behind Anthropic's Model Hardware Standard (MHS), and the central theme of my recent conversation with Jan Wokittel, Director of Smart Manufacturing at Roche, on the AI in Manufacturing podcast.

We explored how MHS could provide an important foundation for Agentic and Physical AI in manufacturing, how it fits alongside existing standards such as OPC UA, Module Type Package (MTP), and Asset Administration Shell (AAS), and why the emergence of AI-native interfaces could fundamentally change the relationship between equipment manufacturers, automation vendors, and their customers.

Beyond the technology itself, Jan offered a particularly valuable perspective on how an enterprise manufacturer approaches these developments: balancing innovation with long-term architectural flexibility, operational reliability, and regulatory requirements.

From AI Agents That Analyze Data to AI Agents That Interact with Machines

Jan describes the evolution of AI agents as a progression through several capabilities.

Initially, AI systems became proficient at analyzing large amounts of information. They subsequently gained the ability to interpret documentation, understand context, design applications, and generate executable software.

The next frontier is connecting those capabilities to physical hardware.

As Jan explains, AI agents have become increasingly capable of performing tasks within software environments, but interacting with the physical world remains considerably more complex.

Industrial equipment is typically accessed through specialized interfaces and automation systems. A production machine might expose operational data through OPC UA, while other capabilities are available through proprietary APIs, vendor-specific software, or embedded control systems.

For an AI agent to interact meaningfully with that equipment, it needs more than access to raw data.

It needs to understand what the equipment is, what capabilities it provides, what operations can be performed, and the context in which those operations are appropriate.

This is where Jan sees the significance of MHS.

Drawing parallels with the Model Context Protocol (MCP), which has simplified how AI agents discover and interact with software tools, Jan argues that standardized interfaces for physical equipment could make industrial capabilities more accessible to AI systems.

Rather than requiring custom integrations for every device or application, manufacturers could increasingly expose equipment capabilities in ways that AI agents can discover and use.

The implications extend well beyond robotics or autonomous machines. They could affect how engineers, scientists, production specialists, and other business users develop applications and interact with manufacturing systems.

The Business Opportunity: Accelerating Innovation Cycles

One of the most interesting aspects of our conversation was Jan's explanation of how these capabilities could change innovation processes inside large manufacturing organizations.

Consider a scientist working in a pharmaceutical manufacturing environment who wants to investigate a production problem or build an application using operational data.

Today, that seemingly straightforward requirement can initiate a lengthy process.

The scientist needs to contact an engineering department to explain the business problem, identify the required equipment data, and determine how that information can be accessed.

Engineers must then evaluate the available interfaces, establish connectivity, consider the impact on automation systems, and coordinate with IT teams on network access and security requirements.

According to Jan, simply clarifying the requirements can take one to two weeks. Additional engineering and IT coordination can extend the overall process to several weeks.

The actual business application may be comparatively simple.

Now imagine that an AI agent can discover the available equipment, understand its data and capabilities, and retrieve the information required to build the application.

Instead of coordinating multiple technical teams for every new request, the scientist could describe the intended outcome, with the agent handling much of the underlying technical complexity within an approved environment.

Jan illustrates how a workflow that currently takes three or four weeks might eventually be completed in a day.

The example is illustrative rather than a measured implementation result, but it highlights the potential business impact.

The greatest value of AI-enabled equipment interfaces may not be automation of individual engineering tasks, but a fundamental reduction in the time between identifying a business opportunity and developing a working solution.

For enterprise manufacturers operating across multiple facilities, this could have significant implications for productivity, experimentation, and the speed at which operational improvements are implemented.

Why Existing Industrial Standards Are Not the Entire Answer

Manufacturing already has established technologies for accessing and interacting with equipment.

OPC UA, for example, provides a standardized framework for industrial communication and supports information models, structured data access, and the invocation of methods on industrial systems.

So what does an AI-oriented hardware interface add?

Jan's argument centers on the difference between exposing a technical capability and providing the broader context an AI agent needs to use that capability intelligently.

Consider a pump that exposes methods to start or stop its operation.

An industrial interface may allow an authorized client to invoke these methods. But understanding whether the pump should be started requires additional information.

What process does the pump support? What conditions must be satisfied before it can operate? What happens downstream when it starts? Which other systems or equipment could be affected?

That knowledge may exist across engineering documentation, standard operating procedures, process models, control logic, and the expertise of human engineers.

It is not necessarily available through a single machine interface.

Jan sees an opportunity for AI agents to bring these different sources of context together, allowing them to understand not only the available actions but also their operational implications.

This distinction is important.

OPC UA itself supports rich information modeling, and its companion specifications can provide standardized semantic descriptions of industrial equipment. The challenge is that a complete understanding of operational intent, dependencies, and consequences often extends beyond what a particular OPC UA implementation exposes.

An AI-oriented interface could help bridge that gap by making equipment capabilities accessible within a broader reasoning and orchestration environment.

The objective is not simply to connect an AI model to a machine.

It is to enable agents to understand equipment capabilities in the context of the manufacturing processes they support.

Will MHS Replace OPC UA?

Despite the potential disruption, Jan does not expect MHS to replace OPC UA in the foreseeable future.

Manufacturing operations have stringent requirements for reliability, determinism, safety, and operational continuity. Established industrial communication and control technologies have evolved to address these requirements.

Manufacturers cannot simply replace proven automation infrastructure whenever a new technology emerges.

Instead, Jan anticipates a period in which AI-oriented interfaces and existing industrial standards coexist.

OPC UA could continue to support industrial connectivity, structured information exchange, and established automation functions, while MHS-style capabilities provide another interface through which AI agents discover and interact with equipment.

This raises an interesting possibility for the industrial automation ecosystem.

Rather than competing directly with emerging AI standards, organizations responsible for established industrial technologies could make their existing information models and capabilities more accessible to AI agents.

For example, the OPC UA ecosystem has invested significant effort in developing companion specifications that describe machines, equipment, and industrial processes.

These standardized information models could become valuable sources of context for AI-driven applications.

The opportunity is therefore not necessarily to replace existing standards, but to make decades of industrial standardization more useful and accessible in an AI-driven environment.

MHS and Module Type Package: Two Approaches to Modularity

Another interesting comparison is between MHS and Module Type Package (MTP).

MTP emerged from efforts to make industrial production systems more modular and flexible, particularly in process industries.

Its underlying principle is often described as plug and produce.

Instead of manually engineering every integration between a production module and its supervisory automation environment, MTP provides a standardized description of the module's services and interfaces.

This makes it easier to integrate new equipment, reconfigure production systems, and adapt manufacturing capabilities.

Jan sees some conceptual overlap between this approach and the capabilities discussed around MHS.

Both aim to reduce the complexity involved in interacting with industrial equipment. Both seek to make equipment capabilities more accessible through standardized descriptions.

However, their purposes and operating environments are not identical.

MTP addresses modular automation integration, while MHS introduces the prospect of making physical equipment capabilities accessible to AI-driven workflows.

For manufacturers already investing in MTP-based architectures, adopting a new interface should not mean discarding existing investments.

Jan suggests that an especially valuable outcome would be for AI agents to understand and use capabilities already described through MTP.

Rather than choosing one standard over the other, manufacturers could benefit from complementary technologies that support different layers of the architecture.

Ultimately, the appropriate approach depends on the manufacturing environment, existing technology investments, and the business requirements being addressed.

Asset Administration Shell: The Semantic Foundation for AI

The discussion becomes particularly interesting when we consider Asset Administration Shell (AAS).

Unlike MTP, which focuses on modular automation integration, AAS provides a standardized framework for representing and exchanging information about industrial assets.

An Asset Administration Shell can contain structured information about equipment characteristics, technical documentation, engineering properties, operational data, and lifecycle information.

For example, an equipment manufacturer could provide standardized digital nameplate information, technical specifications, or digital calibration certificates.

This information can then be consumed by engineering, maintenance, asset management, and other enterprise applications.

Jan highlights the significance of this capability in the life sciences industry, where manufacturers must maintain detailed information about physical assets throughout their operational lifecycles.

Consider the challenge of managing thousands of production assets across multiple facilities.

Each asset may have technical characteristics, calibration requirements, maintenance documentation, and other information needed for regulatory compliance.

Today, much of this information is maintained across engineering tools, equipment documentation, and other enterprise systems.

Even when information is available digitally, inconsistencies in naming conventions, units of measurement, and data descriptions can make automated interpretation difficult.

AAS helps address this problem through standardized information structures and semantic models.

For AI agents, these models could provide a valuable foundation for understanding the physical equipment they interact with.

Jan sees AAS and MHS as particularly complementary.

AAS can provide standardized knowledge about an industrial asset, while an AI-oriented equipment interface can make capabilities of that asset accessible to agents.

The combination could enable AI systems to work with richer equipment context without requiring every manufacturer to build proprietary semantic models from scratch.

There is also a broader opportunity.

AI itself could help reduce the engineering effort required to implement and maintain these standards by assisting with documentation, model generation, metadata enrichment, and compliance with predefined information structures.

Standards that have historically been difficult to implement manually could become more accessible through AI-assisted engineering workflows.

MQTT and Unified Namespace: The Real-Time Event Backbone for Industrial AI Agents

Beyond individual equipment interfaces, manufacturers also need to consider how AI agents fit into the wider operational architecture.

This brings us to MQTT and the Unified Namespace (UNS).

Jan sees the Unified Namespace as a powerful foundation for future AI capabilities because it provides a common environment through which systems can exchange operational information.

Rather than connecting every application directly to every piece of equipment, the Unified Namespace enables producers and consumers to publish and subscribe to information through a shared messaging infrastructure.

An AI agent can participate in this architecture as another consumer or producer of operational data.

For example, an agent responsible for production monitoring could subscribe to equipment events, process measurements, or quality information published through the namespace.

Using that information, it could investigate deviations, generate insights, or trigger downstream workflows.

In our discussion, I emphasized another important architectural role for the Unified Namespace: event-driven orchestration of AI agents and industrial systems.

Consider a manufacturing process where a quality inspection event is published to the UNS.

An AI agent responsible for laboratory investigations could subscribe to that event, evaluate the inspection result, and initiate an investigation.

Depending on the required workflow and its authorized capabilities, the agent might interact with other applications, retrieve additional operational data, or request information from laboratory equipment.

Once the investigation is completed, its outcome could be published back into the Unified Namespace for other systems to consume.

This example illustrates how multiple agents and applications could coordinate through a shared operational event backbone without requiring tightly coupled point-to-point integrations.

It also highlights the complementary roles of the different technologies.

MQTT provides a lightweight messaging protocol. The Unified Namespace provides a common information and event-sharing architecture. Industrial information models provide semantic context, while AI-oriented interfaces can expose equipment capabilities to agents.

Together, these capabilities could form an important part of the architecture required to operationalize agentic AI across manufacturing environments.

Building AI-Ready Manufacturing Foundations: Lessons from Roche

One of the most valuable aspects of Jan's perspective is his focus on long-term technology investments.

Large manufacturing projects operate on very different timelines from software development.

Jan explains that major facility investments can take three to five years before production begins.

During that period, the technology landscape can change dramatically.

Manufacturers therefore face a difficult architectural question: How do you make technology decisions today that will remain relevant when a facility becomes operational several years from now?

Jan's approach emphasizes flexibility and modularity.

Rather than trying to predict exactly which AI technologies will dominate in the future, he argues that manufacturers should establish foundational capabilities that make it easier to adopt new technologies as they emerge.

These include standardized equipment information, interoperable interfaces, consistent naming structures, and modular system architectures.

He illustrates this thinking with an example from Roche.

As part of the company's transition to SAP HANA, Roche also undertook work to standardize naming conventions and metadata associated with physical equipment.

The objective was to improve consistency and traceability between physical assets and their corresponding information within enterprise systems.

Although this is not an example of deploying MHS, it demonstrates the type of foundational work that Jan considers essential for future AI applications.

AI agents become considerably more useful when they can reliably identify equipment, understand its characteristics, and relate operational information to physical assets.

Without those foundations, each new AI application risks becoming another isolated integration project.

For Jan, the longer-term objective is to create an environment where emerging technologies can be introduced without repeatedly redesigning the underlying architecture.

That is an important distinction between building individual AI use cases and building an enterprise architecture capable of supporting AI at scale.

What Does This Mean for Industrial Automation Vendors?

The emergence of AI-oriented equipment interfaces also raises important questions about the future of industrial technology vendors.

Historically, equipment manufacturers have often supplemented their hardware offerings with proprietary software platforms, analytics applications, visualization tools, and specialized integration capabilities.

These solutions can create value, but they can also introduce dependencies that make it difficult for manufacturers to adopt alternative software.

Jan challenges this model from the perspective of an enterprise customer.

When purchasing industrial equipment, he wants access to the equipment's capabilities and information without being forced into a particular vendor's software ecosystem.

For example, an equipment supplier might offer its own predictive maintenance application. But a global manufacturer may already operate an enterprise-wide predictive maintenance platform that needs to work across equipment from multiple vendors.

In that situation, the customer may place greater value on standardized access to equipment data and capabilities than on another proprietary analytics application.

As AI makes software development more accessible, this distinction could become even more important.

Manufacturers may increasingly use AI-assisted development to build applications tailored to their operational requirements rather than relying exclusively on software supplied by equipment vendors.

Jan expects this to create pressure on traditional industrial technology business models.

The competitive advantage could progressively shift away from controlling proprietary interfaces and toward providing differentiated equipment capabilities, reliable information models, and interoperable products.

For vendors, the question becomes less about how to keep customers within a proprietary technology environment and more about how easily customers can use their equipment within increasingly open, AI-enabled architectures.

Agentic AI in Regulated Manufacturing: From Human-in-the-Loop to Human-on-the-Loop

The final part of our conversation focused on the practical challenges of introducing AI agents into regulated manufacturing environments.

In the life sciences industry, new technology deployments are subject to stringent qualification, validation, and compliance requirements.

Manufacturers must demonstrate not only that individual systems perform as intended, but also that the overall processes supported by those systems remain compliant.

This becomes particularly challenging when AI agents begin generating applications, making recommendations, or executing increasingly complex workflows.

Jan highlights an important scalability problem with the traditional human-in-the-loop approach.

In many AI applications, a human reviews and approves the output of an AI system before it is used.

While this provides an additional level of oversight, it also creates a potential bottleneck.

AI systems can generate outputs at a speed and volume that human reviewers may struggle to match.

If every action requires manual approval, much of the productivity advantage of AI automation may be lost.

Jan therefore raises the possibility of moving toward greater human-on-the-loop oversight.

Rather than manually approving every individual output, humans would increasingly supervise the overall process, supported by appropriate controls, validation mechanisms, and monitoring.

This is not a solved problem, particularly in regulated manufacturing.

Determining how to validate AI-generated outputs, test agent behavior, govern execution permissions, and maintain accountability remains a significant challenge.

For applications that interact with physical equipment, these considerations are especially important.

The ability to understand and invoke a machine capability does not eliminate the need for safety controls, deterministic automation, or appropriate operational authorization.

Nevertheless, Jan believes that manufacturers will need to develop new governance approaches if they want to realize the full potential of AI at enterprise scale.

The Bigger Picture: Preparing Manufacturing Architectures for an AI-Driven Future

Our conversation with Jan reinforced an important architectural principle.

The future of agentic AI in manufacturing is unlikely to depend on a single standard or protocol.

Instead, manufacturers will need a combination of capabilities that allow AI agents to understand industrial environments, access operational information, interact with equipment, and participate in business processes.

Existing industrial technologies already address important parts of this challenge.

OPC UA provides industrial connectivity and information modeling. MTP enables modular automation integration. AAS provides structured asset information and semantic descriptions. MQTT and the Unified Namespace support real-time information exchange and event-driven integration.

MHS introduces another important dimension: the possibility of making physical equipment capabilities directly accessible to AI agents through standardized interfaces.

How these technologies evolve and integrate remains uncertain.

But for enterprise manufacturers, the immediate priority is arguably clearer.

Rather than building architectures around a specific generation of AI tools, manufacturers should focus on creating interoperable, modular, and semantically consistent foundations that allow emerging technologies to be adopted with minimal disruption.

As Jan emphasizes, the goal is to have the right foundations in place so that when new capabilities emerge, manufacturers can simply integrate them and begin creating value.

The real opportunity is not just making machines accessible to AI agents. It is building industrial environments in which new ideas can move from concept to implementation without repeatedly confronting the same integration barriers.

And that could fundamentally change the pace of innovation in manufacturing.

Watch the Full Conversation

In this episode of AI in Manufacturing, Jan Wokittel, Director of Smart Manufacturing at Roche, joins Kudzai Manditereza to explore the Model Hardware Standard, the future of industrial interoperability, and what agentic AI means for manufacturing technology architectures.

Watch the full episode: [Insert YouTube episode link]

About Jan Wokittel

Jan Wokittel is the Director of Smart Manufacturing at Roche, where he works on the technology strategies and architectures supporting the next generation of manufacturing facilities. His interests include industrial interoperability, modular automation, digitalization, and emerging technologies that enable more flexible and intelligent manufacturing operations.

‍