AI agents: when should a decision use a model, and when should it use code?

A common AI architecture mistake is sending every step through a large language model. But an agent...