The argument that code beats structured tool-calling at scale makes sense. We're at maybe 30 tools and already hitting schema management pain. What does your sandbox setup look like? Any security concerns with arbitrary Python execution?