Abstract. Design of autonomous systems is facilitated by automatic synthesis of correct-by-construction controllers from formal models and specifications. We focus on stochastic games, which can model the inter- action with an adverse environment, as well as probabilistic behaviour arising from uncertainties. We propose a synchronising parallel composition for stochastic games that enables a compositional approach to controller synthesis. We leverage rules for compositional assume-guarantee verification of probabilistic automata to synthesise controllers for games with multi-objective quantitative winning conditions. By composing winning strategies synthesised for the individual components, we can thus obtain a winning strategy for the composed game, achieving better scalability and efficiency at a cost of restricting the class of controllers.