A new AI agent doesn't just generate CAD code from a text prompt - it runs the code, critiques its own results, and rewrites until the geometry actually matches the request.
Researchers have built RA-CAD, a "ReAct" agent that cycles through a generate-execute-critique-rewrite loop when turning natural-language design instructions into parametric CAD code. After each attempt, the model executes its own code, looks at the outcome, and writes an explicit critique - either signing off on the result or flagging what to fix in the next pass. The system trains in two stages: a supervised fine-tuning step called CAD Code Bootstrapping teaches basic CAD syntax, and a second stage, Feedback-Driven Agent Optimization, uses reinforcement learning to reward the entire generate-critique-rewrite trajectory based on how closely the final shape matches the target, measured by F1 score and Chamfer Distance. Tested against the CADFusion and Text2CAD benchmarks, RA-CAD reportedly beat both existing text-to-CAD methods and larger proprietary language models on execution validity and geometric accuracy.
Most text-to-CAD tools treat critique as an afterthought - a fixed rule, a hand-written prompt, or a separate model bolted on after generation. RA-CAD instead trains the critique step itself as part of the reward signal, so the model learns which kind of self-feedback actually leads to a better rewrite rather than just sounding plausible. That distinction matters for any workflow that wants AI-generated CAD models a working engineer could trust without re-drawing them by hand.
The paper is a revision of an earlier arXiv submission, and as usual with agent loops, it says nothing about how many execute-critique cycles a real design takes or what that costs in compute and time - the kind of detail that decides whether this is a lab benchmark win or something an actual CAD user would sit through.