Fluctuations in network conditions are a common phenomenon. They arise in the current wired Internet due to changes in demand, and in wireless networks due to changing interference patterns. However, current congestion control design typically does not account for this, and in this sense the majority of congestion controllers proposed so far can be deemed as ``myopic''. The present work deals with the following question: how should network end-users exploit such temporal fluctuations? We introduce a formal framework, in which time diversity is explicitly described by phases in network condition. We propose as bandwidth allocation criterion the solution to an optimization problem, which features both classical (myopic) users and so-called farsighted users. We identify the corresponding farsighted user strategy as that maximizing throughput subject to a social norm related to TCP-friendliness. We establish basic desirable properties of the resulting allocations. We propose adaptive decentralized algorithms for farsighted users to achieve their target allocation. The algorithms do not require either explicit knowledge of dynamics in network conditions, or special feedback from the network.