In RelAgent, an LLM agent explores a relational database, writes SQL feature programs, and selects a classical model. The final predictor runs without any LLM calls and is interpretable by construction.