The Integration-Boundary Hypothesis in Multi-Agent Intelligence

Abstract

This working paper proposes the Integration-Boundary Hypothesis for multi-agent intelligence. Multi-agent systems may expand the range of hypotheses, interpretations, reasoning paths, and partial results that can be explored or maintained in parallel. However, when the system as a whole is required to converge on a finite output or action, the problems of selecting, rejecting, compressing, and integrating multiple possibilities do not necessarily disappear; they may instead reappear at the system level. The hypothesis distinguishes raw exploration width from effective integration load. A large number of candidates may impose little additional integration difficulty when most can be easily discarded. In contrast, even a small number of candidates may create substantial integration difficulty when they are highly relevant, similarly plausible, mutually dependent, or difficult to reject. The paper therefore introduces the conceptual variables exploration width (B_t), effective integration load (L_t), integration capacity (K_t), and integration pressure (P_t). A central prediction is that successful exploration can itself increase integration difficulty when effective integration load grows faster than integration capacity. The paper further treats integration boundaries as relative to the class of available system improvements rather than as permanently fixed limits. The hypothesis is presented as a falsifiable conceptual framework rather than an impossibility theorem or an empirically established quantitative model.