> "Instead of improving an agent against a fixed test, we let the evaluation evolve alongside the agent"

This quote should have been highlighted earlier in the article.