Building AI agents with raw LLM SDKs works fine for prototypes until you need structured outputs, testable code, and production reliability. The gap shows up in a predictable way. Your notebook code w