Three Universities. Three Papers. One Week. The New Race to Automate Engineering Judgment.

Three universities. Three papers. One week. The new race to automate engineering judgment.

Between May 19 and May 22, three academic groups at ETH Zurich, MIT, and the Ulsan National Institute of Science and Technology in South Korea independently published research describing multi-agent LLM systems designed to automate topology optimization — the computational process of finding the most efficient material distribution inside a prescribed design domain. The convergence is too tight to be coincidence and too academic to declare a market inflection.

The most concrete result in this cluster comes from MIT. A team led by Isabella Stewart and Faez Ahmed built a system where a judge agent evaluates visual renderings of each design iteration and recommends parameter revisions. Their TO-Agents pipeline produces preference-aligned structures — designs a practicing engineer would have chosen — in 60 percent of trials. Strip out the visual and historical feedback loops, and that rate drops to roughly 10 percent. Sixfold improvement, on a metric that maps directly to whether a design is any good.

The Korean paper, from Hyunjee Park and Hayoung Chung, introduces TopOptAgents: a system of six LLM-based agents that handle problem formulation, code generation, validation, and quality assessment, cycling through iterative self-refinement until the design converges. The ETH Zurich group, led by Gioele Molinari and Mark Fuge, went furthest toward benchmarking. Their EngiAI framework achieved 96 to 97 percent task completion on a beam-design problem using proprietary model backends, with open-source 4-billion-parameter models reaching 55 to 78 percent.

What unites all three is the same empirical pattern: the multi-agent approach earns its gains most where a single LLM fails. The Korean paper is explicit. Iterative self-refinement, the authors write, is "particularly pronounced for problem classes where the pretrained language model has limited prior exposure, such as formulations whose literature and open-source implementations are comparatively sparse." The MIT results show the same ceiling. Open-source models underperform proprietary ones most dramatically on tasks requiring conditional branching and multi-step instruction following — exactly the reasoning-heavy work that separates real engineering judgment from pattern matching.

That gap is the thing worth watching. EngiAI's benchmarking makes it concrete: the performance difference between proprietary and open-source models is largest precisely where the reasoning demands are highest. The gap between 96 percent and 55 percent is not a calibration problem. It is the shape of what multi-agent reasoning can do that single-agent prompting cannot.

Major CAD vendors have taken notice at a structural level. Ansys, Siemens NX, Altair, and Autodesk all offer generative design modules — but they are single-agent wrappers around mature solvers, not multi-agent decision-making systems. Whether multi-agent LLM orchestration is on their roadmap, or represents a different architectural bet entirely, is a question the vendors are not answering publicly.

Here is what that junior engineer at a manufacturing firm actually does. Topology optimization has been computationally solved for decades. The physics is well-understood, the finite-element solvers are mature, the commercial tools are established. What has resisted automation is the judgment call: which parameter values to try next, whether a converged result is physically plausible, whether the initial problem formulation captures what the engineer actually wanted. That judgment has, until now, required an expert human in the loop. These three papers describe systems that appear to be closing that gap — at least on benchmarks.

The honest version of this story is: three groups of researchers independently built similar systems in the same week, all showing that multiple LLM agents talking to each other outperforms one LLM working alone, especially on problems where training data is thin. That is a genuine result in the AI-agents literature. Whether it is also the beginning of a shift in how engineering firms allocate human attention to computational design problems is a question none of these papers can answer from the bench.

The next chapter of this story writes itself: a manufacturer runs one of these systems on a real structural problem — not a benchmark — and reports back on whether the designs held up. Or a commercial CAD vendor publicly commits to multi-agent LLM orchestration in their product roadmap. Until then, the story is an academic cluster with a plausible industrial implication — worth writing, worth reading, and still open on the question of whether it changes anything.

Sources: TopOptAgents (Park, Chung — UNIST) | TO-Agents (Stewart, Chen, Ahmed — MIT) | EngiAI (Molinari, Felten, Fuge — ETH Zurich) | GetLeo.ai Topology Optimization Landscape