When performing goal sampling for multiple agents, the goal likelihoods are shared between agents somehow. This may need some detective work.