How digital twins framework could help ground agentic AI context

The Digital Twin Consortium (DTC) has recently published its first Agent AI Capabilities Periodic Table (AIA-CPT). This builds on a previous version for digital twins to understand some of the new capabilities and requirements when combining digital twins with agentic components. It’s comparable to extending retrieval augmented generation (RAG) and context engineering from documents to interconnected representations of physical systems.

The AIA-CPT comes in the form of an interactive framework that allows stakeholders to map out the requirements for the digital twin aspects required to support agentic AI systems. It can help a group of stakeholders from management, finance, and various technical domains to develop an interactive roadmap and checklist for agentic AI rollouts. It complements rather than replaces the existing Digital Twin CPT.

I spoke about this recent progress with Dan Isaacs, CTO of the DTC and Pieter van Schalkwyk, CEO of XMPro and co-chair of the AI Joint Consortia Working Group across the DTC and the Object Management Group. They said this would help evolve traditional digital twin frameworks to support an intelligent digital twin composed of collaborative autonomous agents. Also, collaboration with the Large Language Model (LLM) efforts will improve governance for LLMs and extensibility for digital twin frameworks.

Van Schalkwyk says:

There is hardly a conversation now where it's not about the LLMs. It's changed the narrative and the expectations people have. The negative part is that everything is an agent now. If you look at a lot of the tool providers like Anthropic, OpenAI, Google, they are all providing agents, but those agents are more automations with LLMs inside that are camouflaged as agents. It can do things on your machine, but it uses an LLM to be generative. It is not a deterministic system at all. At the other extreme, a fully deterministic system needs hard-coded rules to know exactly what to do, and some use cases require that. We see that there is a blend between the two. How can I develop logical thinking use cases using an LLM, which uses a similar reasoning process to humans? But how can I make sure that I ground them using classical and traditional AI and physics-based models to make sure that the information those reasoning modules work on is grounded in physics, especially in the engineering side?

One example of this complementary approach is the balance of code in an agentic code platform that XMPro is building for energy, utilities, and mining industries, with about forty thousand lines of code. Schalkwyk reflects:

We realized that only eight percent of the source code is for integrating the LLM, while ninety-two percent of the code is about the business processes and governance that sit around managing agency and agents like you would have done without LLMs. You still need a lot of the routing, orchestrating, governance, and confidence level scores, which are not really native LLM capabilities. People inside the DTC are also realizing it’s a hybrid of more deterministic elements and then using the generative parts of the agents. These agents can’t see physics. They have to go through a digital twin as the intermediary that can ground them. That is not left to the LLM to make up.

With the AIA-CPT framework in place, the DTC has begun soliciting ideas for new testbeds that could help resolve interoperability issues and could eventually scale intelligent and agentic digital twins across enterprise boundaries. The DTC is also developing the underlying automation, which will also make it easier to scale quickly. Isaac says:

Honestly, when we made the announcement, we had an overwhelming expression of interest. We have had over a hundred proposals, and only a third of those are members. So there is a pent-up demand to bring this forward.

What’s an agent?

Virtually every vendor these days seems to be recasting its existing technology or platform as agentic. Van Schalkwyk says the DTC approach requires clarifying the autonomous aspects and types:

For us, an agent is an autonomous, or semi-autonomous, software entity that uses AI to perceive, make decisions, take actions and achieve goals in the kind of physical environments that they operate, and that's where the digital twin connection comes in. We do include chatbots in those where we see them as semi-autonomous so that they would get direction.

The DTC has also spelled out five agent category types progressing from zero to four.  These categories include:

• Static Automation: Rule-based systems with pre-programmed responses like chatbots, RPA, and decision trees.
• Conversational Agents: Natural language interaction with basic context management, like customer service chatbots and virtual assistants.
• Procedural Workflow Agents: Multi-step task execution with tool integration like workflow orchestration frameworks and collaborative conversation systems.
• Cognitive Autonomous Agents: Self-directed planning with sophisticated reasoning, like research assistants, coding agents, and strategic advisors.
• Multi-Agent Generative Systems (MAGS): Collective intelligence and emergent system behaviors like industrial multi-agent systems, distributed autonomous organizations, and enterprise-scale coordinated intelligence

Van Schalkwyk clarifies that these category levels are different from the level of autonomy commonly used to describe proficiency in domains like self-driving cars:

You could start with a multi-agent system, or you could start with a conversational system. A conversational system may never become a multi-agent system. There's no need for it. That's the difference between that and the levels of autonomy. Because with the levels of autonomy, you kind of work on the premise that as soon as you achieve a certain level, you aspire to the next, and then you aspire to the next, so you build up. The type is the kind of lane that you're in, and you're going to stay in that.  You may improve your technology readiness level (TRL) and maturity levels inside. But they will stay in their lanes because they serve two different purposes inside an organization.

New backbone required

The DTC is also looking at how to strike the appropriate balance between emerging new agentic AI frameworks like Model Context Protocol (MCP) and Agent-to-Agent (A2A). On the one hand, MPC is much easier to integrate AI tools, akin to what USB—C has done for physical connectors. But Van Schalkwyk believes enterprises, particularly in the industrial space, need to proceed with caution, particularly when looking to take advantage of multi-agent systems across facilities and even companies:

The focus in my company is on things like oil and gas, energy, utilities, and mining. They don’t allow those capabilities and tools because the governance mechanisms have not been established. They’re more used to the industrial protocols and control systems protocols, such as MQTT and others. What we are seeing in the DTC, especially in the security and trustworthiness group, is that there's a higher sense of alert at the moment around all these things because the people on that team come from organizations focused on ensuring these things are secure, safe, and verifiable.  MCP and A2A are pretty much in the developer domain inside the organization, so it does not sit within the governance frameworks of the automation systems, which is where a lot of the actuation happens around digital twins in a production environment.

One technical example of this problem is that MCP is currently only implemented using peer-to-peer connections, which could require up to ten thousand separate connections between ten agents. In contrast, broker-based approaches like the data distribution service (DDS) for embedded systems would only need ten connections and are widely used for real-time communications in safety-critical scenarios in defense, factories, and utilities.

Van Schalkwyk observes that a balanced approach will be required to figure out how to scale today's shiny new agentic toys to practical enterprise solutions:

People can easily get blindsided by the agentic hype and the shiny bright toys. That's why we try to create the capabilities periodic table for agents, so that you can have a mature conversation around what it is that we need.  It's easy to do a pilot or a proof of concept, but it's really hard to scale it out. It is really hard to scale it out. If you have 1000 agents coded in a certain way and you have no visibility of what's in there, we cannot allow that on our factory floors, our oil and gas plants, or to drive our autonomous trucks in the mine. So those are the things that we are focused on.

My take

The LLMs driving the agentic AI hype and the digital twin community, focused on building, managing and optimizing physical products, factories, and infrastructure, live in different worlds. The big AI vendors seem comfortable with the move fast and break things mentality, which is probably OK for the reality media production they are serving up to consumers, regulators and investors. But that will not go over well for safety-critical and expensive equipment. Meanwhile, the industrial automation community is struggling to figure out how to get more of these things to talk to each other. Improving efforts to combine these approaches could help improve guardrails for agentic LLMs on the one hand and the flexibility of digital twins approaches on the other. It will be interesting to see how well the DTC testbeds demonstrating various combinations scale up.