The performance of a TCP connection is typically dictated by what the network can provide rather than what the application would like to achieve. In particular, the Additive-Increase Multiplicative-Decrease (AIMD) mechanism employed by TCP hinges on its ability to meet specific throughput requirements since it has to respond to congestion signals promptly by decreasing its sending rate. The level and the timing of congestion signals impose strict limitations on the achievable throughput over short time-scales. To that end, this paper presents a new architecture, whereby a set of TCP connections (we refer to them as the Stunts) sacrifice their performance on behalf of another TCP connection (we refer to it as the Free) by picking up a delegated subset of the congestion signals and reacting to them in lieu of the Free connection. This gives the Free connection just enough freedom to meet specific throughput requirements as requested by the application, without affecting the level of congestion in the network. We evaluate our architecture via extensive simulation experiments.