The stability and the predictability of a computer network algorithm's performance are as important as the main functional purpose of networking software. However, asserting or deriving such properties from the finite state machine implementations of protocols is hard and, except for singular cases like TCP, is not done today. In this paper, we propose to design and study run-time environments for networking protocols which inherently enforce desirable, predictable global dynamics. To this end we merge two complementary design approaches: (i) A design-time and bottom up approach that enables us to engineer algorithms based on an analyzable (reaction) flow model. (ii) A run-time and top-down approach based on an autonomous stack composition framework, which switches among implementation alternatives to find optimal operation configurations. We demonstrate the feasibility of our self-optimizing system in both simulations and real-world Internet setups.