Simulation has long played a part in testing new configurations and new functionality in a diverse range of software. Through such simulations, the boundaries of the system state are explored and the relationship of that state to other applications tested — sometimes to destruction. A critical differentiator between a simulation and a live, deployed application is that simulations are allowed to fail. As truly autonomous applications evolve, this capacity for simulation must be built in from the ground up or the benefits of experience — including the ability to tolerate failure — will be lost. This must be achieved without undermining the global correctness of visible application behaviour. We suggest an engineering approach to enable the introduction of such simulation with minimal or no recoding and we propose a composition architecture to allow for safe dynamic deployment in substantial autonomic systems. We have identified our approach as application Dreaming.