Take the case of parallel, derivative-free optimization. A set of agents independently traversing a solution space need some kind of shared information to base

Take the case of parallel, derivative-free optimization. A set of agents independently traversing a solution space need some kind of shared information to base their decisions on, eg. the current best solution found by all agents, together with its score and location, and what areas of the search space have low values (or high ones). I'm sure there's a way to do this without shared data, but I doubt it beats the ease of slapping a "#pragma omp parallel" on a code block and adding a critical section for resource access.