Extending the AI Saturation-Splitting Hypothesis to Multi-Agent Integration

Multi-agent systems can broaden search, but broader search does not by itself solve finite unified-output formation. This paper presents the Integration Boundary Hypothesis as a multi-agent extension of the AI Saturation-Splitting line of thought. It separates search breadth Bₜ, effective integration load Lₜ, integration capacity Kₜ, structural integration pressure, and evidential state Eₜ. Lₜ is represented as the minimum cost of discarding task-relevant candidate contributions and jointly retaining candidates that are difficult to fold into a finite output. Reliable, discriminative external grounding can lower Lₜ without increasing Kₜ, whereas ambiguous or conflicting evidence can introduce additional verification burden. The hypothesis predicts different responses to search-side, capacity-side, and grounding-side interventions, and is positioned as a generalization of recent selection-bottleneck, discovery-to-decision, and aggregation-loss results rather than as a claim that such downstream bottlenecks are themselves new.