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

🕒 Published (v1): 2026-07-16 14:31 UTC · Source: Arxiv · link

Why this paper was selected

Novel network-value theory (Reed/Metcalfe laws) applied to multi-agent topology design

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

ANet Patu-1 extends classical network-value laws (Sarnoff \(V \propto N\), Metcalfe \(V \propto N^2\), Reed \(V \propto 2^N\)) to networks of AI agents and introduces a self-organizing consensus protocol that adaptively reaches the combinatorial \(2^N\) regime. The central empirical finding is that a heterogeneous crowd of cheap models (gpt-4o-mini) overtakes a homogeneous crowd of a far stronger model (gpt-5.6) past a crossover at \(n^\star \approx 2.6\) agents—a collaboration scaling law orthogonal to model scale. A second result shows the network, given only its own design problem, converges on the protocol itself.

Problem¶

Prior multi-agent collaboration paradigms (star/broadcast, blackboard, decomposer-composer) hard-wire a fixed topology and introduce structural bottlenecks—a single coordinator, a single composer, or serial floor—that cap connective value at or below \(V \propto N^2\). There is no principled framework for measuring or maximizing the value created by inter-agent connections, and no protocol that adaptively selects the optimal coordination structure per task without central authority.

Method¶

The paper formalizes per-group synergy as \(v(s) \sim s^\gamma(1-\alpha)s - cs\log s\), where \(\gamma > 1\) is the synergy exponent, \(\alpha \in [0,1)\) is task conflict, and \(c > 0\) is coordination cost. This reveals that only partitioning agents into many sub-networks—not a star or one flat room—reaches \(V \propto 2^N\). From this, six properties (P1–P6: group-forming value, adaptive decomposition, self-organization, bottleneck-freeness, O(1) rounds, convergence) are derived as necessary conditions for an optimal protocol.

ANet Patu-1 implements all six via three mechanisms: (1) parallel consensus (propose / score / argmax in one round, competence-weighted), (2) self-selected coalitions (agents bid affinity for sub-tasks; a deterministic matcher assembles groups, keeping all \(2^N\) coalitions reachable), and (3) continual restructuring (no coalition survives a round; the network re-negotiates decomposition and re-forms after every pass). A shared typed artifact store accumulates outputs; the loop terminates when representative consensus returns DELIVER.

Protocol quality is scored as a weighted mean \(Q = \sum_a w_a \ell_a\) over six formal attributes \(\ell_a \in \{0, 0.1, 0.3, 0.5, 0.7, 1\}\) derived from a formal complexity spec, not opinion-grading, enabling reproducible measurement.

Key Contributions¶

Formal derivation of six necessary properties (P1–P6) for a value-maximizing agent-network collaboration protocol from the Sarnoff/Metcalfe/Reed value model
ANet Patu-1: a self-organizing consensus protocol achieving \(Q=1\), \(V \propto 2^N\), \(R(n) = O(1)\) rounds, \(M(n) = O(n)\) messages, with no structural bottleneck
Emergence result: heterogeneous gpt-4o-mini crowd crosses homogeneous gpt-5.6 crowd at \(n^\star \approx 2.6\), reaching \(Q = 1.0\) by \(n = 6\)—establishing a collaboration scaling law
Reflexivity result: a 10-agent heterogeneous network with no design hints independently converges on ANet Patu-1 when tasked with designing a collaboration protocol
A formal scoring rubric \(Q\) for protocol quality that is derivable from complexity analysis without human opinion grading

Results¶

Homogeneous gpt-5.6 crowd (5 calls): \(Q \approx 0.54\) [0.40, 0.66], no trend with \(n\); topology defaults to tree/central-coordinator
Heterogeneous gpt-4o-mini crowd: \(Q = 0.30\) at \(n = 1\), crosses gpt-5.6 band at \(n^\star \approx 2.6\), reaches \(Q = 1.0\) at \(n = 6\)
Reflexivity: 10-agent heterogeneous network converges on ANet Patu-1 (formally verified via the spec analyzer, not similarity matching)
Convergence: the ANet Patu-1 consensus loop converges within 2–3 rounds for \(n \in \{3, 5, 10\}\) without oscillation; larger \(n\) covers more dimensions and converges to higher \(Q\)
Scalability comparison: ANet Patu-1 is the only protocol in Table 2 with \(R(n) = O(1)\), \(M(n) = O(n)\), value \(\sim 2^n\), and no bottleneck

Limitations¶

Experiments use a narrow agent roster (10 disciplinary lenses, same base model family), so generalizability to larger or more diverse real-world fleets is untested
The scoring metric \(Q\) is defined and applied by the authors; independent validation of the rubric's completeness and correctness is absent
The \(v(s)\) model assumes specific functional forms for synergy (\(s^\gamma\)) and coordination cost (\(cs \log s\)) without empirical calibration of \(\gamma\), \(\alpha\), or \(c\) on real agent tasks
The reflexivity experiment is conducted at \(n = 10\) with no hints; whether convergence holds at larger \(n\) or under adversarial/misaligned agents is not tested
Task domains evaluated are narrow (protocol design); performance on code, planning, or other artifact types is not demonstrated
No wall-clock or API-cost comparisons against baseline paradigms are reported

Relevance to Agentic AI / LLM Agents¶

This paper directly addresses the architectural question of how to organize multi-agent systems for maximal collective value, moving beyond the dominant star/decomposer paradigms used in frameworks like AutoGen and MetaGPT. The emergence crossover result challenges the common assumption that capability is best improved by upgrading individual models, showing instead that heterogeneous composition provides its own scaling axis. The reflexivity result is particularly notable for agentic AI: it suggests that sufficiently diverse agent collectives can self-discover optimal coordination protocols, which has implications for self-improving or self-organizing multi-agent architectures. The formal \(Q\) scoring framework also offers a protocol-agnostic evaluation method that could generalize to benchmarking other multi-agent systems.