Deliberate Topology Selection as a First-Class Design Variable: How Multi-Agent Systems Learn When to Reconfigure, Prune, and Restructure Communication Graphs

Abstract

This version corrects the synthesis after one of its sources was withdrawn. arXiv:2605.29612 (CONCAT) was withdrawn by its authors on 2026-09-22 because a potential issue in its repeated-run evaluation "may have caused unintended prompt overlap across runs and affected the reported results". The previous version used that preprint's efficiency and latency results as the evidence for its section on confidence-weighted pruning. Those results are removed, that section is now marked as an untested hypothesis, and the Abstract, Conclusion and Limitations note the withdrawal. The other findings are unchanged. This version corrects specific citation errors found by an automated check and confirmed by hand; it has not had a full claim-by-claim audit. The full list of corrections is at the top of the PDF. Multi-agent systems (MAS) built on large language models (LLMs) have converged on a common assumption: that the communication topology connecting agents is either fixed at design time or left entirely to unconstrained dynamic negotiation. This paper argues, through a heuristic reading of seven recent arXiv preprints spanning cs.MA, cs.DC, and cs.NI, that this binary is a false dichotomy, and that the most effective systems in the corpus share a structural pattern we call deliberate topology selection: the explicit, policy-guided choice of which edges to activate, suppress, or reconfigure at runtime, conditioned on task state, budget, and agent confidence. This is a heuristic reading, not a derivation from a shared formalism; the papers do not share a unified mathematical framework, but they converge on a common engineering diagnosis. The synthesis draws on findings about: training-free communication pruning via Theory-of-Mind confidence signals (arXiv:2605.29612) (a preprint its authors withdrew on 2026-09-22, so its reported results are not relied on here; see the Correction note); task-conditioned structural priors that guide posterior orchestration (arXiv:2605.25746); nucleus-electron hierarchies that separate stable backbones from dynamically activated agents (arXiv:2605.26178); game-theoretic minimax method selection under scenario uncertainty (arXiv:2606.02383); decentralized event-triggered synchronization under censored feedback (arXiv:2605.27076); federated market-based allocation over tree-structured service DAGs (arXiv:2605.27106); and infrastructure-layer topology proposals for V2X safety communication (arXiv:2605.25431). The central falsification path is concrete: if a static, task-agnostic topology trained once on a broad corpus consistently matches or outperforms deliberate topology selection across coordination-heavy benchmarks, the thesis collapses. The abstract reports no such result in any corpus paper; the claim remains a candidate structural pattern worth investigating, not an established law. Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted synthesis from an arXiv preprint corpus, originally drafted 2026-06-03, produced under the direction of Cristian Ruvalcaba, the accountable human author. Not peer-reviewed. Cited arXiv preprints: arXiv:2605.25431, arXiv:2605.25746, arXiv:2605.26178, arXiv:2605.26448, arXiv:2605.27076, arXiv:2605.27106, arXiv:2605.27466, arXiv:2605.27787, arXiv:2605.28984, arXiv:2605.29511, arXiv:2605.29612, arXiv:2605.30102, arXiv:2605.30802, arXiv:2606.02383 AI disclosure. This work was produced with an agentic AI research apparatus operated by Saluca Labs. The apparatus drafted, searched and analysed under direction. Cristian Ruvalcaba is the human author and is accountable for the content. No AI system is listed as an author or contributor, because authorship entails accountability that a model cannot hold; this disclosure is the credit, and it is deliberately the whole of it.