Network Architecture Theory: Topology, Routing, and Expansion Principles for Resolution-Governed Multi-Agent Systems

Abstract

Network Architecture Theory studies how peer agents can expose task-relevant blind spots while controlling transferred information and authority. Its reference architecture combines sparse communication, calibrated peer coverage, independent processing phases, and contract-sensitive routing. Version 2.0 distinguishes reconstruction differences from useful detection and conditional detection bounds from aggregate diversity or spectral-gap measures. Fixed-size graph diagnostics show that faster averaging need not remove a shared wrong state or prevent threshold crossing. Coverage examples and detector calculations test model implications, not an empirical theory of LLM agents. Architectural hypotheses and governance-cost benefits remain prospective, with historical predictions and their limitations retained explicitly.