We model the notion of k-resilient distribution of computation graphs supporting agent decisions, over dynamic physical multi-agent systems (e.g. IoT systems). We devise a self-organizing repair method, namely DRPM[DMCM], based on distributed optimization to repair the distribution as to ensure the system still performs collective decisions and remains resilient to upcoming changes. Resilience is based on the concept of replicas of computations so that those hosted by disappearing agents can activate on other agents. We focus on a particular type of reasoning process to repair: distributed constraint optimization (DCOP), where computations are decision variables and constraints distributed over a set of agents. We provide a full stack of mechanisms to install resilience in operating stateless DCOP algorithms, which results in an robust approach using MGM-2 to repair any stateless DCOP algorithm at runtime. We experimentally evaluate the performances of our methods on different topologies (uniform or problem-dependent) operating DCOP algorithms (A-MaxSum and A-DSA) to solve classical benchmarks (random graph, graph coloring) while agents are disappearing.