We study the notion of k-resilient distribution of graph-structured computations supporting agent decisions, over dynamic and physical multi-agent systems. We devise a replication-based self-organizing distributed repair method, namely DRPM[MGM-2], to repair the distribution as to ensure the system still performs collective decisions and remains resilient to upcoming changes. We focus on a particular type of distributed reasoning process to repair, where computations are decision variables and constraints distributed over a set of agents. We experimentally evaluate the performances of our repair method on different topologies of multi-agent systems (uniform or problem-dependent) operating stateless DCOP algorithms (Max-Sum and A-DSA) to solve classical DCOP benchmarks (random graph, graph coloring, Ising model) while agents are disappearing.