ANet Patu-1: The Value of Connection in the Agent Network
Networking and Internet ArchitectureArtificial Intelligence
Summary
The authors explore how the value of networks made up of AI agents depends on how those agents connect, similar to classic network value laws like Sarnoff's and Metcalfe's. They create a model for collaboration value based on group size and develop ANet Patu-1, a protocol that lets AI networks form coalitions efficiently. Their findings show that a diverse group of simpler models can outperform a uniform group of stronger models as the network grows, and that such networks can independently discover this optimal protocol without external guidance. This work connects network theory with distributed AI collaboration.
Authors
Mu Yuan, Jinke Song, Zhaomeng Zhou, Lan Zhang
Abstract
The Internet taught us that the value of a network depends on \emph{how} its nodes connect: broadcast stars scale as $V\!\propto\!N$ (Sarnoff), fully-connected meshes as $N^2$ (Metcalfe), and group-forming networks as $2^{N}$ (Reed). We ask the analogous question for networks of AI agents. We model the net value of connection as a function of coordination-group size, derive from it the properties an optimal collaboration protocol must have, and introduce ANet Patu-1 -- a self-organizing consensus protocol in which the network continuously re-forms its own coalitions, adaptively riding the upper envelope of all three regimes at $O(1)$ parallel consensus rounds. To measure value without opinion-grading, we score an emergent protocol by formally specifying it and deriving its complexity, the way distributed algorithms are analyzed. Two results follow. (i)~Emergence -- a crowd of the \emph{cheapest} model, when heterogeneous, starts weak but its collective value compounds with $N$ and \emph{overtakes} a crowd of a far \emph{stronger} model that is homogeneous: a crossover that marks a scaling law for collaboration rather than for scale. (ii)~Reflexivity -- a heterogeneous network, given only its own problem and no design hints, converges on ANet Patu-1 itself, reconstructing the high-dimensional law that governs its own connective value.