Consortium Releases Checklist for AI Agents

Recently, the Digital Twin Consortium announced it is entering the next phase of its AI Agent Capabilities Periodic Table Framework (AIA CPT), described as a tool “to help organizations design, evaluate, and assess intelligent agent–based systems.” The AIA CPT is a supplement to the previously published Digital Twin Capabilities Periodic Table. The consortium believes, with this table, organizations can move from “we need AI agents” to “here’s exactly what to build.” The table separates capabilities into six columns:

• Perception and knowledge
• Cognition and reasoning
• Learning and adaptation
• Action and execution
• Interaction and collaboration
• Governance and safety

Pieter van Schalkwyk, CEO of XMPro and Chair for the Digital Twin Interoperability Task Group in the Industrial IoT Consortium, expects AI-driven functions to become part of intelligent digital twins. “There is a lot of agent-washing,” he warned, of the software vendors’ tendency to market their products as AI agents. For first-time digital twin adopters, he proposes thinking of the digital twin as a job applicant. In this hypothetical scenario, they might ask, “What is the problem I’m hiring the digital twin to solve? What are its capabilities—its resume?” he explained. 

Van Schalkwyk thinks, armed with the table, digital twin solution buyers will be able to ask more concrete questions, such as, “Do I need collaboration functions? Do I need a ChatGPT-like interface? Do I need reasoning in my digital twin?” he said. 

The chart, he noted, was the result of working with consortium member groups looking to adopt digital twin technologies. “So we were able to validate [these capabilities] with them through their use cases,” he said. “We might still fine-tune the chart based on feedback, but the core capabilities shouldn’t change.”

The important thing, Van Schalkwyk pointed out, is to remember the first basic question: what is the problem you’re trying to solve with the digital twin? The answer to this question may lead you to consult the AIA CPT to look for specific characteristics in the solution. 

Senior Editor Kenneth Wong spoke at length with Van Schalkwyk about the table in a recent podcast.

Kenneth Wong: The Consortium has just recently published the AI Agent Capabilities Periodic Table. Could you tell us what it is and how should people be using it?

Pieter van Schalkwyk: Yes, so it’s actually an extension on what we did we three or four years ago [when] we published the Digital Twin Capabilities Periodic Table, which really looked at what are the capabilities that you need when you design a digital twin, and not start with technology, but start with what is the problem and kind of the resume of the digital twin that was very popular. It’s been downloaded thousands of times. So Dr. Michael Grieves, who is kind of seen as the father of digital twins, introduced this notion of intelligent digital twins. And a lot of that has got AI-driven capabilities in it. So taking what we’ve learned out of the digital twin table, we then created the AI Agent Capabilities Periodic Table that works in conjunction, and it helps you to create these intelligent digital twins. So if I’m going to make my digital twin smarter by using AI, and especially agentic AI, then what would I look for in capabilities before I choose a technology?

Wong: The table has various columns. If a company is marketing a particular suite as a digital twin suite, would you advise people to evaluate these tools against these features? How should a digital twin buyer treat this list? 

Van Schalkwyk: Typically, it starts again with what is the problem I’m trying to solve with this agent? What is the role that I want the agent to have inside this? We have four types of agents, from a very simple chatbot to a very advanced multi-agent system. So identifying what level [you need], and then inside each of those six categories that we have —the knowledge and perception, cognition, learning, action, interaction and collaboration, safety and governance—inside those we’ve got very specific capabilities. So I think there’s 45 in total across the six areas. The first thing that you do is identify, for this type and my business problem, what are the capabilities? Do I need reasoning? Do I need collaboration? Do I need a chat interface? Based on that, I would create a requirement set based on capabilities, and then go to vendors and say, “Don’t show me your technology. Show me how you achieve or deliver these capabilities.” 

Wong: There are some simulation and maybe some PLM vendors who are now starting to market some of their suites as digital twin solutions. If you look at them in general, are there certain categories or functions in this periodic table that they seem to be having a difficulty satisfying? Or in general, are they satisfying one particular bucket really well, but not so well in other buckets?

Van Schalkwyk: I think that’s been one of the key things that we’ve tried to do in the Digital Twin Consortium, is to provide that guidance around what are real digital twins, what are models, and what are simulations? And sometimes it could be the same underlying model, but it could, in one instance, just be a model, and in another use case, it would be a digital twin. 

It’s really important to look at what are the capabilities that [you] need, and then go and match that to what that specific vendor is saying. We have also seen that vendors, because of this, are now getting better at making sure that they answer the question around, “What are the capabilities of my digital twin?” Inside a capability, you don’t need every capability at level five or, you know, whatever the level might be. For some use cases, I just need a simple version of that. We see vendors are starting to map their marketing to what we have prescribed in the capabilities periodic tables.

Wong: Do you imagine in the future vendors marketing their digital twin suites as, for example, Digital Twin Consortium certified because it meets these periodic table requirements?

Van Schalkwyk: I don’t think the DTC will use a formal certification program. But what we are doing is, for members who can map their capabilities to this, we are creating a catalog inside DTC where you can go to a certain capability, and then see which members have capabilities around that. 

Wong: Digital twins are a relatively new, emerging concept, so there are going to be lots of first-time digital twin buyers. For them, is there a particular [capability] that is more important than others? 

Van Schalkwyk: My answer to that is generally to start with the problem you are trying to solve. You must have a clear description of the problem and what capabilities would deliver the answer you need for that problem. Don’t start with data and then say, “What problem can I solve with it?”, which is what we quite often see. 

The inverse of that is, start with a problem. What are the questions I need to answer in order to solve my problem? And then what? What is that analytic? Now that analytic is what the AI could provide. Sometimes it would be reasoning and cognition. Sometimes it’s just something that has to learn, because of my use case. If you go to the GitHub site for the AI periodic table, there’s actually a sequence of prompts that you can run, but it starts with number one, what are your business problems? Then we [can] help you extract the core requirements, and we actually help you generate what capabilities you need. That is all available for free.  DE