Silent Entropy and Structural Fragility: How Communication Isolation, Coordination Debt, Adversarial Symmetry, Consensus Illusions, and Topology-Memory Coupling Jointly Define a Failure Taxonomy for Deployed Multi-Agent LLM Systems

As large language model (LLM)-based multi-agent systems (MAS) move from research benchmarks into operational deployment, a set of recurring structural failure modes is emerging that cannot be attributed to any single component defect. This paper synthesizes seven findings from recent cs.MA and cs.DC preprints into a candidate taxonomy of MAS failure patterns, organized around a central thesis: **MAS failures are predominantly structural rather than component-level, arising from mismatches between coordination topology, memory architecture, communication channel assumptions, adversarial scaling dynamics, and consensus semantics.** This is a heuristic reading across the cited sources, not a derivation from a shared formal framework. The synthesis draws on: (1) empirical evidence that entropy accumulates monotonically in LLM agent systems across interaction rounds [corpus:arxiv:2606.08162]; (2) a demonstration that scheduled cross-agent memory injection silently fails due to hardcoded architectural isolation [corpus:arxiv:2606.04896]; (3) findings that model scale creates a compliance-correction symmetry in adversarial linear pipelines, where larger models become more obedient to malicious instructions [corpus:arxiv:2606.12709]; (4) evidence that answer-level consensus in multi-agent debate masks reasoning-level divergence [corpus:arxiv:2606.08457]; (5) the counter-intuitive result that memory depth and network topology interact to flip the sign of coordination speed [corpus:arxiv:2606.04197]; (6) a proof that deliberative consensus degrades oracle accuracy below single-model baselines through error propagation [corpus:arxiv:2605.30802]; and (7) characterization of the cost of learning under censored feedback in threshold-activated cooperative settings [corpus:arxiv:2605.27076]. Together, these findings suggest that MAS deployment safety requires co-design of topology, memory depth, channel verification, and consensus semantics — none of which is sufficient in isolation. Falsification path: a controlled experiment holding task fixed while independently varying topology class, memory depth, and channel architecture should produce predictable failure-mode signatures if the taxonomy is structurally grounded. --- Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted from arXiv preprint corpus on the date in the filename. Cited arXiv preprints: 2605.27076, 2605.30802, 2606.04197, 2606.04896, 2606.08162, 2606.08457, 2606.12709, 2606.13068, 2606.13594