We present a framework for the parallelization of depth-first combinatorial search algorithms on a network of computers. Our architecture is intended for a distributed setting and uses a work stealing strategy coupled with a small number of primitives for the processors (which we call workers) to obtain new work and to communicate to other workers. These primitives are a minimal imposition and integrate easily with constraint programming systems. The main contribution is an adaptive architecture which allows workers to incrementally join and leave and has good scaling properties as the number of workers increases. Our empirical results illustrate that near-linear speedup for backtrack search is achieved for up to 61 workers. It suggests that near-linear speedup is possible with even more workers. The experiments also demonstrate where departures from linearity can occur for small problems, and also for problems where the parallelism can itself affect the search as in branch and bound.