Agent evals should feel like real work