Search papers, labs, and topics across Lattice.
This paper explores the value of connection in networks of AI agents by modeling the net value of connection as a function of coordination-group size and deriving properties for an optimal collaboration protocol. The authors introduce ANet Patu-1, a self-organizing consensus protocol that adaptively forms coalitions while achieving efficient consensus in parallel rounds. Key findings reveal that heterogeneous groups can outperform homogeneous ones in terms of collective value, and that such networks can autonomously converge on the protocol that maximizes their connective value.
Heterogeneous AI agent networks can evolve to outperform homogeneous counterparts, revealing a new scaling law for collaboration that defies traditional strength metrics.
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.
