Large language models can generate evaluation benchmarks, write comments, and output comparative ratings such as Elo scores for agent harnesses