Coordination Under Pressure: How Topology Coupling, Fairness Constraints, Censored Feedback, Deliberation Failure, and Communication Efficiency Jointly Define a Candidate Design Framework for Scalable Multi-Agent Coordination

Abstract

This version corrects a wrong arXiv identifier and removes the results of a withdrawn preprint. Version 2 cited the CONCAT framework as arXiv:2605.29511, which is a different paper (DynaGraph); CONCAT is arXiv:2605.29612, and all three citations are corrected. CONCAT was then withdrawn by its authors on 2026-09-22, who wrote: "We identified a potential issue in the repeated-run evaluation of our method that may have caused unintended prompt overlap across runs and affected the reported results. We therefore withdraw the manuscript for further investigation and re-evaluation". Version 2 used CONCAT's reported results (up to 2.02x higher efficiency and a 50.1% latency reduction) in Section 2.3 and as corroboration for the communication-entropy claim in Section 3.5. Those results are removed; that claim now rests on a single study, and the text says so. The five-mechanism thesis and the seven retained papers are unchanged. Typographic dashes are removed, Greek letters and mathematical symbols now render in the PDF, and the file carries a neutral name. The abstract below is unchanged apart from punctuation. The problems were found by an automated check and confirmed by hand against arXiv; this version has not had a full claim-by-claim audit. Multi-agent systems (MAS) are increasingly deployed at scales where coordination quality, not individual agent capability, becomes the primary performance bottleneck. This paper synthesises seven specific findings from recent arXiv preprints across cs.MA, cs.DC, and cs.NI to argue a candidate structural thesis: scalable multi-agent coordination is constrained by five interlocking mechanisms, topology-memory coupling, fairness-exploitability duality, censored-feedback identifiability, deliberation error propagation, and communication entropy, and that optimising any one in isolation predictably degrades the others. This is a heuristic reading, not a formal derivation; the sources share vocabulary and structural analogies but not a unified formalism, and the cross-domain bridges are argued by analogy rather than by shared mechanism. The corpus spans agent-based simulation, multi-agent reinforcement learning, LLM-based MAS, distributed network systems, and cooperative game theory. Key findings include: memory depth and network topology interact non-monotonically in LLM consensus formation arXiv:2606.04197; fair cooperative MARL policies are systematically exploitable unless contention geometry is carefully structured arXiv:2606.06162; censored feedback in threshold-cooperative tasks creates an identifiability problem that dominates coordination regret under the specific binary-censorship structure studied arXiv:2605.27076; deliberative consensus in multi-agent oracles degrades accuracy below single-model baselines on prediction market questions due to error propagation arXiv:2605.30802; and communication entropy reduction can be embedded as a training objective without sacrificing task performance in the MARL benchmarks studied arXiv:2606.07200. Two additional sources provide structural context: the compliance-correction symmetry in linear adversarial workflows on HumanEval arXiv:2606.12709 and the topology-dependent horizon trade-off in self-organised railway traffic arXiv:2606.13068. A further source, the Librarian study arXiv:2605.27787, is retained in a weakly-connected addendum; its connection to the communication entropy claim is analogical rather than mechanistic. The falsification path for the central thesis is concrete: if a MAS architecture independently optimises topology, fairness, feedback observability, deliberation protocol, and communication budget and achieves Pareto improvement across all five axes simultaneously, the interlocking-constraint claim is falsified. No such architecture is reported in this corpus. Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted synthesis from an arXiv preprint corpus, originally drafted 2026-06-14, produced under the direction of Cristian Ruvalcaba, the accountable human author. Not peer-reviewed. 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.