evals

How to tell if your AI agent actually got better

The agent's reply is not the result, LLM judges have blind spots of their own, and a few dozen real failures beat a thousand synthetic tests.