When building AI agents, it doesn't end with just "making the LLM think." You eventually want to execute some kind of program ...