There's a classic set of rules of thumb for how much a network is worth as it grows: broadcast networks scale with N, fully connected ones with N squared, group-forming ones with 2 to the N. This work asks the same question about networks of AI agents, where the nodes are models that talk and coordinate rather than televisions or telephones.

The authors model connection value against coordination-group size, spell out what an ideal collaboration protocol should do, and propose ANet Patu-1, a protocol that keeps reshuffling its own coalitions to ride the best of those regimes in a constant number of consensus rounds. Two claims stand out: a diverse crowd of cheap models can overtake a uniform crowd of a much stronger one as the group grows, and a diverse network handed only its own problem rediscovers ANet Patu-1 unprompted.

Striking, theory-heavy claims, so the paper is essential for the assumptions behind them.