Building an internal agent: Code-driven vs. LLM-driven workflows