Most agent benchmarks rank speed, accuracy, or cost—but users care about trust, usability, and real results