LangChain released a detailed methodology for iteratively improving AI agent harnesses using evals as a hill-climbing signal. The recipe covers sourcing and tagging evals (hand-written, production traces, external datasets), splitting per behavioral category into Optimization and Holdout sets to prevent overfitting, and running autonomous improvement loops. The key insight: agents tend to overfit to specific tasks during autonomous hill-climbing, so holdout sets are critical for ensuring generalizable improvements.