An AI Agent Ran 187 FEM/CFD Solves Over MCP. Here's How Many Matched the Textbook
An AI agent, Claude Code, drove the open-source NovaSolver Studio application to perform 187 engineering simulations, including structural, thermal, and flow problems, between November 6 and 12, 2026. Logs show 136 of these solves matched external references or another mesh of the same model, with a median difference of 0.13%.
Claude Code, acting as an AI agent, did not use a conversational interface but instead generated Python scripts to control NovaSolver Studio's MCP server. These scripts were responsible for creating geometry, computing reference values, and comparing results across various engineering problem types. This approach allowed the agent to drive the simulation process directly, rather than through graphical user interface interactions.
A notable aspect of the agent's operation was its capacity for iterative problem-solving, particularly during the `check_convergence` step. This process assesses solution stability across different mesh refinements. Initially, some cases, such as chime rods and spinning disks, failed due to insufficient element growth. The agent responded by adding face refinement and re-solving these cases, successfully meeting the convergence criteria and demonstrating an ability to adapt and correct simulation parameters.
The overall process, however, involved managing various complexities and failures. Eight initial DC current cases failed validation due to unaddressed Joule heating, and 13 out of 23 problem topics required at least one redo, design change, or abandonment. The agent also encountered and managed resource constraints, leading to the abandonment of one case where the finest mesh exceeded available memory, highlighting the practical challenges in autonomous simulation.
Share this article
Related reading
6 stories
An AI agent with no step limit stops when its context window is full. On GPT-5.6 that's $160 for one run.
An AI agent running many solves can get expensive; we looked at the cost of a single GPT-5.6 run.

Vietnamese Travellers Can Now Scan NETS QR Codes to Pay in Singapore

Weixin Pay Sees 84% Rise in Overseas E-Wallet Transactions

BPC Pairs SmartVista with Huawei for Payment Modernisation

