In this paper, we consider a network where an access point (AP) wirelessly transfers energy to user equipments (UEs) and UEs use the harvested energy to communicate with the AP. Due to different radio propagation losses, the UEs far from the AP harvest little but consume much, thus suffering high transmission delay, while the UEs close to the AP has redundant energy. To reduce the average delay, we propose an adaptive harvest-then-cooperate (AHC) protocol, where every UE sends data to the AP directly or via other UEs acting as relays in a time-division multiplexing manner. Based on the load and energy distributions, the AP controls the transmit power and routing topology to minimize the average delay. To avoid data overflow in UE relays, we propose a method generating scheduling schemes given the transmit power and the routing topology. The average delay minimization is approximately modeled as a Markov decision process, and a policy iteration algorithm is employed to obtain the optimal combination of the power and the routing. Simulation results demonstrate the effectiveness of the AHC protocol in the delay minimization and reveal that the protocol can relieve the unbalance of the energy distribution.