This study focuses on self-adapting multi-agent systems modelled using the Sonar framework. Our key focus is on forecasting the costs and benefits of adaptation during execution within the Sonar MAPE loop. Analysing these adaptation processes is complex due to Sonar enabling second-order activities, such as structural adaptation involving agent interaction protocols or the organisational network itself. We forecast these dynamic processes using a digital twin and abstractions. For instance, the agent’s decision-making logic is represented with a stochastic model. Since Sonar is conceptualised with Hornets (a Nets-within-nets formalism), we need a “probabilistic” extension of Hornets. To illustrate our approach’s effectiveness, we showcase a small case study of a self-modifying MAS organisation and provide an analysis of adaptation dynamics.