Sounds interesting, but I'm not quite getting the relevance for people writing code with an agent. Should I be doing evals?