Overview
Language-model outputs are stochastic: repeated runs on the same problem can produce different answers. This variability is not only a source of inconsistency. It can be used deliberately to explore more of the solution space, identify recurring ideas, and surface unusual alternatives that a single run may miss.
Two patterns make this useful. Stochastic consensus collects independent answers and measures where they overlap. Multi-agent debate lets agents inspect one another's proposals and revise their own conclusions over several rounds.