I run Binary Core LLC , a small consultancy that builds OpenUSD and AI tooling for visualization, digital twin, and robotics teams. Over the past several months, we've been building BowerBot , an open source LLM agent that assembles structured OpenUSD scenes from natural language. This post walks through the architectural decisions behind it: why we built it FastAPI-style, how the plugin system works, why validation is core to the agent loop, and the tradeoffs we made along the way. What problem are we solving OpenUSD (Pixar's Universal Scene Description) is becoming the universal 3D scene format. Pixar uses it. Apple ships USDZ to Vision Pro. NVIDIA Omniverse is USD-native. Adobe, Autodesk, Siemens, and Trimble all back it through the Alliance for OpenUSD . But authoring USD scenes correctly is hard.…