This document introduces a multi-agent framework for real-time physical simulation. Unlike classical physical simulators, no shared discrete time-line is imposed to simulated objects. Each simulated object is an autonomous agent that can maintain its simulation state, and possibly modify its behavior by checking its environment. Decision schemes are proposed that enable objects to locally adapt their computation time, in order to maintain approximate global synchronization. Results and measures are presented, especially regarding time management and simulation synchronization in regard of classical simulation methods. Examples are shown, outlining some of the practical advantages such a framework provides.