0:01 Engineers are starting to envision actually being able to allow agents to set up and run the entire simulation in a trusted solver is IG appears to be the key to unlock robust agent-driven workflows. Recorded at the CDFAM computational design symposium in Tokyo October 2026. Learn more at CDFAM.com. Hello, my name is Matt Cederberg. I’ll be talking about bridging the CAD, the simulation gap. And 22 years ago, I was graduating from college and my dad who is a computer science researcher and one of the colleagues of George Allen gave me an unusual graduation gift.

0:48 Matt, he said, “I just invented a technology that will change computer Aed design. Let’s start a company together. Never mind that I had studied economics and I knew almost nothing about design. So we named the company T-splines. We wrote plugins called Tespines for Rhino Rhino and Solid Works and sold it to Autodesk where now it’s a core part of Fusion 360. And that T-spines way of modeling along with its cousin subdivision surfaces is now a part of most major CAD tools.

1:23 10 years ago, it happened again. My friend, Dr. Michael Scott, along with Dr. Tom Hughes, who’s perhaps the most influential living FEA researcher, convinced me there was a better way to do finite element analysis. The traditional way to run finite element analysis didn’t make a lot of business sense. Before a CAD model could be analyzed, it needed to be defeatured and meshed. For parts of any complexity like we’ve seen throughout this conference that’s far from trivial and it needs to be repeated for every design iteration.

1:55 So the better way is called isoggeometric analysis where you simulate directly on the CAD geometry. It’s still a finite element method but doesn’t require conformal meshing. This method was introduced by Dr. Tom Hughes and it’s now been vetted by over 7,000 research papers. And as a side note, Corfirm is sponsoring the annual IGA conference which starts this Monday in Tokyo. So if you’re still in town, come on over there and and see us there.

But Mike Scott and I decided to found a company with some other outstanding co-founders and after a decade of collaboration with automotive and defense companies, we partnered with Disso Systems to put this inside of Abacus and that was released earlier this year. And it especially shines in four places. Complex geometry, complex physics, engineers who aren’t meshing experts, and agent-driven workflows. So, just to touch on a couple of those, complex geometry is one of significant interest to to many of us.

2:56 We’ll just hone into this part of this brake caliper. The traditional route to create a a hex mesh on this took our best engineer 20 hours to defit your fully hex mesh and it had you to remove the critical fillets. And with core from IGA, the full CAD was ready to solve in just under a minute, including the critical fillets where the high stress concentration was. The advantage of integrating with Abacus is that we can take full advantage of their trusted Abacus standard solver capabilities and material models but make these accessible to engineers who may not be meshing experts or might not have days to prepare geometry for simulation.

And engineers are starting to envision actually being able to allow agents to set up and run the entire simulation in a trusted solver is IG appears to be the key to unlock robust agent-driven workflows. By removing the most brittle part of the simulation workflow and with no meshing step, agents like Claude driving abacus CE with core from IG for abacus in this example can run the whole loop, change the design, rerun the analysis, read the results with no human fixing the mesh in between.

4:12 So twice now someone brilliant has come with an idea of how to change the world of how engineering works. Now for the third time I mentioned that IG is opening the possibility of more robust use of LLMs for engineering. And inspired by this and frustrated by his own experience working with agents in general, my partner Mike Scott built a new framework of how to more efficiently use multi- aent workflows for high assurance domains like engineering simulation.

4:42 We will call this core for AI and we’re in the early stages of bringing it to market. And today you get a sneak peek and I’m handing it over to Mike Scott. Okay. So that that’s the title of my my talk because that’s the inner monologue that I spend most of my days with working with agents. I’m afraid I can’t let you do that. Thanks. So I work in what Matt called high assurance domain.

5:06 So this is really sophisticated simulation physics-based workflows with companies that are doing very large projects and they have cover a broad range of physics. They cover complex geometry. Their workflows are deeply embedded in the sense that they’ve built mountains of tools on them and they want to use AI and they hear about AI but they’ve spent decades with these workflows primarily trying to reduce risk. So these engineering organizations are building things that people depend on in a variety of ways including safety.

5:48 They usually cost a lot of money if they get them wrong or they can cause other kinds of challenges to society. And so when I speak with high assurance engineering domains and this is I would say like canonical cases would be defense, aerospace, automotive and so on. This is essentially what they tell me. I don’t like the risk is the primary response. And then they say, well, even when I do spend money on AI, it gives me something plausible, but it’s impossible to finish to get it to a finished state.

6:26 That closing that gap between kind of halfway done and full done is a huge gap in terms of these challenging applications. So what I’m looking at is how to close the gap not just from like an implementation point of view but what is the mathematical foundation for managing the specific attributes of a large language model inside of these kinds of numerical methods or stacks of numerical methods to solve really hard problems.

6:59 So in this talk I’m going to talk about four concepts that are the foundation of doing this. We’re going to talk about isogric analysis which Matt introduced. We’re going to talk about a concept called the data weave which is a way to handle the mountains of data and relationships between data that a large language model can then reason over effectively and with 100% accuracy. We’re going to talk about the importance of supporting non-frontier models.

7:31 So across the board, the customers that I work with, I mean, I don’t think there’s ever a day that’s going to come that they’re going to share their data or their workflows with a Frontier Lab. I mean, maybe the day will come. I don’t know. But it for the most part, they’re interested in smaller models, models that they can control, airgapped models. So the question is can I get the advantages of the behavior the semantic reasoning that comes from large language models and do it in a safe risk-free way which is a very hard problem.

8:09 Okay. So let’s let’s talk about the first topic here isometric analysis is the first of two solvers in the system. So the system we’re building has a physics solver. Although I don’t like the word physics because it’s a huge field. I’m specifically talking about primarily finite element methods, primarily structural mechanics, structural dynamics, highly nonlinear phenomena, highly nonlinear materials. And so the framework has that solver stack based on isogeeometric analysis for tooling robustness, which I’ll talk about in a minute.

8:48 But there’s also another equilibrium problem hiding underneath all of this, which is the mathematics of coordination. So an LLM, when you think about agents working together, you see a lot of ad hoc stuff about, you know, I I got it to do this and I got it to talk to this and I got but can we control that in a fundamental way? Meaning can you repose agent communication in terms of minimization of a of a of a particular type of energy or landscape energy landscape so that the agents have to work together in such a way that they minimize the energy of coordination.

9:31 And the answer is yes. In fact, it does exist and is built on what you would call game theory, not video games, but game theoretic PTEEs that capture the energy of coordination and allow you then to repose agent coordination in the context of this type of a model. And so the approach is is you have a a advanced solver core that actually has two solvers in it.

10:05 One that’s minimizing the energy of coordination or the behavior of agents and one that’s managing the physics of agents. Let me go back here. And then you mix in in a very bounded control way the one thing that LLMs have that deterministic numerical methods or classical numerical methods do not have, which is semantic reasoning. That’s essentially the only property that they have that’s distinct and unique and you want to introduce it at the right time and you want it to be rigorously bounded in terms of its impact on the actual numerical methods in which it’s operating.

10:49 Okay. So this is IGA. If you’ve used funded element tools, you’ve meshed things, you’ve prepared geometry and you know as model complexity grows, it gets worse and worse in terms of the time you spent on the nonvalue ad aspects of it. So here’s what IGA is like for a designer. So this is a satellite module, the structural pieces of a satellite that NASA, this is an open model from NASA from I think 2017.

11:18 Normally, this would be a very this would be a big headache. And it’s not just the meshing that’s a headache. It’s managing all the structural connections, right? So, you have a lot of parts, you have a lot of fasteners, and you have to get the load paths right or the simulation isn’t right. You’ll solve something, but it won’t be right. And so you need a method that can handle that complexity while giving you the right answers and in fact improving upon traditional lower order finite element methods which is what isoggeometric analysis gives you.

11:47 So in this case, we’re just going to do a simple simulation, just a simple gravity check on the body and just to demonstrate load paths are correct, that you have the appropriate deflections, that the stress is captured correctly. And you, if you remember, the meshing flashed on the screen, right? That’s a completely automated process, completely robust, completely automated. Similar what George was saying about about implicit. It’s a very similar idea.

12:13 In fact, the immersion of beast blinds is in some sense a type of implicit representation. So it’s not surprising that it’s robust. Okay. And what are the properties that you get? You might say, well, that’s a shortcut. Does it give good results? In fact, it’s better than FBA. Far better than lower FBA, which is why 7,000 papers have been published on it. In fact, at the peak of IGA, more papers were being published on on IGA than were published at the height of FEA publishing in the 80s and 90s.

12:46 So, it’s a huge field now in terms of the impact, the breadth of impact that it’s had. The cool thing though is that it gives you that robust tooling that you need for multi- aent architecture in these kinds of high assurance domains. Okay, let’s talk about data. Okay, agents are bound they’re essentially pure functions, right? An agent you have a context, but essentially it’s a pure function, right?

13:18 And the problem with that is that even the context is tiny compared to the data produced in a simulation pipeline, right? So what do you do about that? If an agent can only remember what’s in its context, and if it’s not in its context, then you’re relying on agent tooling, that’s essentially, I’m going to write a note here. I’m going to write a note here and maybe remember to look at that to remind me what I’m supposed to do.

13:42 That’s not acceptable in these kinds of applications. Okay. So, what we do here is we generate a relational representation of data across all the different layers. We call this a data weave and it’s not a database. In fact, it’s a clean algebra. It’s a type of f algebra. So it’s a very sophisticated data representation that allows you allows the model to query all of the relational data that exists in every layer of the simulation pipeline where the expense of that of that call is just the tokens required to express the call in a particular agent language which I’ll talk about in a minute.

14:30 Okay. Okay, so in this case, in this little example I did, you’ve got relational information associated with the solve itself up through the joints, the rolls, contact, the assembly, the geometry. Whereas normally these are all distinct, right? They’re like data files that you copy back and forth. All of the indexing isn’t even consistent, right? So if you have more than one agent, they can’t even work on it together because they’ll each be interpreting the data differently.

14:58 Okay? Okay. And here’s a little thing about if you’re if you’re interested in this kind of thing. It’s not a database, right? Because what it does is it manages relationships in a way that has algebraic guarantees so that if the agent asks the wrong question, it’s refused and then taught exactly the appropriate invariant representation that it that that it should be asking. Okay. So when you have that kind of representation, what’s the consequence?

15:35 You can define language over it. Precise language based on relational information. And so here’s a ve a very simple example of a clean algebra that you can define. This is like a you can think of it kind of like as you’re familiar with finite elements. It’s like a a minimization problem but not over the reals. It’s over kind of a semi- ring which is a more general representation.

15:56 It’s kind of a foundation of graph theory of of reachability in graphs. And so in this case, here’s a very simple algebra. And an agent then doesn’t need to know about the data. It needs to know how to speak over the data. That’s it. And so the data can scale to be arbitrarily large. And the agent’s reasoning or power to reason is not diminished by that. And your context window is more than sufficient to understand how to communicate in this language.

16:34 So here’s an example. Okay, so here’s the actual visual representation of a data weave and here are questions, the natural language version of the questions, then the grammar or language repres or the the computational language representation of that same query. And it’s kind of showing you how it traverses the this graph in order to return the result. And what’s cool about it and what is enabling is that all of the layers from the drawing to the report that has the quantities of interest for the engineer to to pass judgment on is all represented in terms of the same core geometric representation.

17:15 Okay. And so these questions then can be fairly general, in fact fully general over all of the data. So you can say, I’m looking at this number in a report. Tell me the BREP entities that had anything to do with this particular number. And it doesn’t need all the data to answer that question. Okay. And then for the type of work that I’m involved in, you can’t really load data into an aging context.

17:48 Like that’s not like a thing that you can do that has any value really. But if you can do it this way, then you have essentially constant cost over size of data. And so this is a really foundational piece of technology for doing these kinds of so you have the data and then you have a language and that language I’m not saying language in an imprecise way I mean a provable deterministic language over that data representation.

18:26 Now the power of this is if you get this right, you don’t need a lot of semantic reasoning to do amazing things which in turn reduces risk for a customer. So here’s that model that we ran that IG model. Here are all of kind of the reasoning steps that these agents went through. Notice that I have, if you’re familiar with claude, it’s similar with OpenAI. You have these different tiers, right?

18:57 You have kind of the frontier, you have kind of a workhorse, and then you have a small model. And so I’m just using the opus or the the claude level or the anthropic agents here. But notice that even a haiku agent could run that problem. That’s what our our customers want. And the mathematics makes it terminate with the correct answer. Notice Haiku, all of these little red circles are where it wanted to wander off into the wilderness, right?

19:32 And the mathematics of the problem are saying, “Nope, I can’t let you do that.” That pull you back on and it eventually is able to find that minimum energy solution that gives that simulation result. So in this case, they all got exactly the same result. Over a lot of steps. This is a lot of steps where an agent is speaking the language. Each one of these little dots is where an agent is speaking that language, right?

19:58 So, I put in a natural language prompt and then for the rest of the time that it talks, it’s speaking that language, not natural language. All right, let’s do an example. So, I’m after total engine complete engineering organizations. Right. Can can you have a multi- aent organization that does high assurance work in a way that’s provably correct? That’s what the most I would say forward-looking organizations are interested in for obvious reasons, right?

20:32 Either they can do more work. A lot of my clients have so much backlog work that it’s not about laying people off. It’s like they just want to be able to do all the work that they have. Especially in defense right now in the United States. Is in in particular there’s huge backlogs of work. Okay. So what you do is you define a law which is a data file that defines the behavior of a shop and a shop is a representation of multiple agents.

21:01 The law that they operate under or the invariance of their work the data flow patterns the data representations it’s all specified. So, you go to a customer and you say, “Here’s here’s what we’re going to do. We’re going to sit down. You’re going to tell me how your organization works. What are the data patterns? What are the the no- nos or the things that you you want to you want to incentivize and you build essentially a constitution once and that’s the level that the user operates on.

21:38 And once you have that, you can set up organizations of arbitrary size which will, like I’ll show in a minute, converge to the minimum energy of the coordination of that shop. Okay, so here’s just some overview, right? So total history of mankind, markets have been priced by labor, labor markets, time for money, right? Well, we’re entering a phase where that’s no longer the market dynamics that actually make sense.

22:16 In fact, trust is the new labor, right? It they’re trust markets. They’re trust economies. My customers aren’t asking, “Can an LLM do a bunch of stuff?” That’s not and they know that that it can happen fast. They want to know how can I quantify that it’s a trustworthy result. And in fact, you can make that a precise mathematical statement and develop trust market dynamics mathematically. So here’s a little example.

22:52 So you have some functional in terms of some variables, right? Right? And I can’t show everything here for obvious reasons, but the point is you have a functional and you’re you’re essentially determining what are the things you care about and they become like variational derivatives in some sense of the functional like in a traditional variational calculus point of view. But things you can say posted prices that rise the least stake at which honesty is the best reply.

23:25 Another check only while it removes more loss than it costs. Confidence is screened by a menu never asked. Pay is released as the evidence arrives. So in this market, if an agent says, “I’m done.” And then a deterministic check is run and it’s determined that it’s not done. It actually goes back and penalizes how much money that agent has in its bank and tells it exactly what needs to change.

23:56 Right? So there are these gates that an agent cannot get through unless they tell the truth. And if they tell the if they don’t tell the truth, they’re actually assigned a credit rating and eventually they’re going to be kicked out of the shop and replaced with a different. They’re caught in some weird local minima or something. So, I run about 50 agents a day. 50 agents 24 hours a day, 7 days a week on a whole bunch of hardware.

24:22 And this is how I run my business. So, I price I price work. I put bounties on work and I distribute all of the work according to this particular model of a market and then they all go to work and one of the things that’s priced into the model is my attention. That’s a scarce resource. So if they want to interrupt me, they better be sure that they have a lot of evidence that they’re right and they have to prove it to me.

24:58 And so here’s an example where for fun I have engineering agents that are running my own software and identifying places where there’s user friction. They’ll say, “Boy, this was hard to do. I didn’t like doing this.” They’ll report the issue through a liaison to the development desk. That development desk will take those issues, put a bounty on them, give it to a software development agent inside of a software development organization.

25:27 They’ll implement the solution, report back, the engineer tries it again, the evidence mounts, and then that particular work funded work item completes. Right? And this is how I run my business. This is this is the way that I operate every day. Okay, people ask me, well, how on earth do you build software like this at scale? Well, here’s one of the secrets. Every method is a theorem.

25:56 Okay, so everything in my software is theoretically proven in lean. So I have one of the largest lean libraries I think in the world actually right now in the sense that I have almost I think I have over 6,000 theorems underneath the software. So every single result that returns in this software returns with an error bound a particular mathematical certificate that tells you exactly which theorems govern its behavior.

26:26 Okay. And if you haven’t tried lean for software development with agents, I would highly encourage you to do it. It’s far more powerful than developing in any other language in terms of agent behavior and their ability to be correct. Okay. And I’ll stop there. And you know, we’re I’m collaborating with Matt Matt Core Form. I have a new company called Six Gates. And we just want to have conversations with people and understand what people are stuck on and where they’d like to go with their technology.

27:00 We have a little early adopter program that we’re running right now as we get this off the ground. Thank you. Recorded at the CDFAM Computational Design Symposium in Tokyo, October 2026. Learn more at CDFAM.com.