Original title: What Do CAE Simulation Agents Really Need Beyond a Generic Harness?
Jiasheng Shi, Tianhan Zhang · September 2026
Setting up engineering simulations usually takes deep human expertise. Recently, researchers built complex AI systems with multiple specialized agents to automate this task. This paper tests whether all that extra machinery is actually necessary. The authors compared a complex multi-agent system against a standard single-agent harness that simply runs code and inspects the errors. When given the same resources, the simple single-agent setup matched or beat the complex systems, scoring 96.4% on FoamBench compared to 88.2%. The researchers found that scripted reflection does not help at all. Instead, two basic things drive success: fixing mistakes using execution feedback, which raised scores from 71.8% to 96.4%, and providing solver tutorials as domain knowledge, which drove the score from 80.9% to 96.4%.
Imagine you're following a recipe to bake a cake for the first time. Normally, you'd read all the steps first, then start mixing ingredients. But if you have a helper who watches and tells you right away if you're adding too much flour, you might not need to memorize every detail upfront. In this paper, the AI agent is like that helper—it gets feedback during simulation setup and fixes errors as it goes, which works better than having multiple specialists with detailed plans. The tutorials are like having a quick guide that helps it understand the cooking tools better.
Engineers spend significant time and expertise setting up computer simulations for designs. This work shows that a simpler AI approach can handle this just as well, which could speed up product development and reduce the barrier for using advanced engineering tools.
The abstract states that these results hold with information access and repair budget held fixed, meaning it assumes equal resources for all systems. It was tested specifically on FoamBench, which is a benchmark for OpenFOAM simulations, so the findings might not apply to all CAE tools or scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
What Do CAE Simulation Agents Really Need Beyond a Generic Harness? — Jiasheng Shi, Tianhan Zhang
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.