LLM-as-Judge

When correctness is subjective, grade agent output with a second, structured LLM call instead of eyeballing every run.