Multi-Agent Scaling Peaks at Intermediate Complexity and Only Pays Off Above an LLM Capability Threshold

arXiv·medium signalXBlueskyLinkedInCopy link

This architectural study (arXiv 2607.27942, July 30) distills four design principles for scalable multi-agent systems — simplicity, elastic feedback, sequential workflows with optional loops, and summary-based communication — then formalizes a reference architecture as a constrained directed workflow graph and evaluates four configurations of increasing complexity on terminal-based system engineering tasks with two LLMs of differing capability. Scaling yields measurable accuracy gains with roughly linear cost growth, but only when the underlying LLM clears a minimum capability threshold; performance peaks at intermediate complexity and then degrades from timeouts and evaluation limits. Persistent consistency problems appear at every scaling level, which the authors flag as the central open challenge rather than raw agent count.