Optimal Rates for Agentic Networked Information Aggregation Established

When autonomous agents pass forward only their downstream predictions instead of raw data through a directed graph, information loss inevitably accumulates. Building on foundational networked learning formulations by Kearns, Roth, and Ryu, this work establishes the tight convergence rate for linear and logistic regression. The authors show that along an M-covered path of depth D, excess prediction error stays constant up to depth M² before decaying at an optimal rate of Θ(M²/D), formalizing the mathematical limits of cascading agent architectures.