ANet Patu-1: The Value of Connection in the Agent Network

As AI development shifts from single-model scaling to multi-agent systems, the authors investigate how the topology of agent connections influences collective value. Drawing an analogy to Internet value laws (Sarnoff, Metcalfe, and Reed), the study asks whether AI agent networks can adaptively optimize their structure to maximize collective output without relying on rigid, centralized coordination.

The authors model the net value of agent connections as a function of coordination-group size, accounting for synergy, task conflict, and coordination costs. They define six properties of an optimal collaboration protocol (P1–P6), including group-forming value, adaptive decomposition, and bottleneck-free decision-making. To evaluate these, they introduce ANet Patu-1, a protocol that uses parallel consensus and self-organizing coalitions to avoid central bottlenecks. They test this by comparing a heterogeneous crowd of smaller models against a homogeneous crowd of a stronger model, and by observing whether the network can independently rediscover the optimal protocol structure (reflexivity).

The study demonstrates an 'emergence crossover' where a heterogeneous network of cheaper agents, by virtue of their diverse specializations, eventually produces more value than a homogeneous network of stronger agents. Furthermore, the researchers find that when given an open-ended task, a heterogeneous network of agents naturally converges on the ANet Patu-1 protocol, effectively reconstructing the high-dimensional rules governing its own connective efficiency. The protocol maintains O(1) parallel consensus rounds, allowing it to scale effectively without the performance degradation typical of centralized architectures.

This work suggests that the future of AI intelligence may lie in the 'science of connection' rather than just the scaling of individual models. By proving that diverse, self-organizing networks can outperform monolithic systems, the authors provide a framework for building decentralized, highly efficient collective intelligence systems that do not require a central authority or rigid organizational hierarchy.