The Integration-Boundary Hypothesis in Multi-Agent Intelligence

Multi-agent systems can broaden the search space through parallel reasoning, specialization, mutual criticism, and debate. However, expanding search does not by itself solve the problem of forming a single finite output. This paper proposes the Integration Boundary Hypothesis. When multiple possibilities must be transformed into finite representations, judgments, or actions, selection, rejection, compression, and integration remain necessary and may become system-level bottlenecks. The hypothesis distinguishes simple search breadth from effective integration load, which is determined by task-relevant loss when individual candidates are excluded and by additional distortion arising from relations among candidates. It further distinguishes integration pressure as a structural constraint imposed by the requirement for finite output, integration capacity as the ability to control task-relevant loss, and the effective integration load ratio as the severity of the load relative to that capacity. The hypothesis predicts that boundaries move with the available class of improvements, but that boundary movement saturates when integration capacity fails to keep pace with the non-redundant, high-value candidates discovered by search.