Eval Driven Development

"Read the transcripts!" So concludes Anthropic's engineering article on agent evals. An earlier section gives the reason: "You won't know if your graders are working well unless you read the transcripts and grades from many trials."

A score reports that something changed, not why. A team that changes its agent and watches only a number can improve that number while making the product worse. What routine lets you change an agent's instructions or tools, know what the change did, and decide what to change next?