Hill-Climbing Agent Quality with Agents CLI Evaluation

Google's Agents CLI eval namespace runs a hill-climbing loop for agents: eval generate, grade, compare, analyze and optimize. In a walkthrough on the LangChain template in Agents CLI v1.8.0, a baseline scored 4.0/5 with gemini-3.8-flash as judge, and a wrong chef system prompt dropped it to 2.75.