Evaluation mode where models can iteratively refine outputs or use tools, rather than generating a single response.