Dynamic program optimization is the only recourse for optimizing compilers when machine and program parameters necessary for applying an optimization technique are unknown until runtime. With the movement toward portable parallel programs, facilitated by language standards such as OpenMP, many of the optimizations developed for high-performance machines can no longer be applied prior to runtime without potential performance degradation. As an alternative, we propose dynamically adaptive programs, programs that adapt themselves to their runtime environment. We discuss the key issues in successfully applying this approach and show examples of its application. Experimental results are given for dynamically adaptive programs that seek to eliminate redundant runtime data dependence tests, to select the optimal tile size for tiled loops and to serialize loops that do not profit from parallelism.