Failure Propagation and Self-Correction in Multi-Agent LLM Systems: How Deliberative Consensus, Credit Assignment, and Architectural Isolation Jointly Determine Systemic Reliability

Abstract

This version corrects the use of a withdrawn preprint. Version 2 cited the CONCAT framework (arXiv:2605.29612) in Section 3.4 and reported its results of up to 2.02x higher efficiency and a 50.1% latency reduction. On 2026-09-22 its authors withdrew the manuscript, writing: "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". Those results are removed from the body and the Discussion, and CONCAT is no longer used as evidence; the claim it supported rests, as before, on its two other sources. The Limitations section records the withdrawal. Typographic dashes are removed, mathematical symbols now render in the PDF, and the file carries a neutral name. The abstract below is unchanged apart from punctuation. The problem was found by an automated check and confirmed by hand against arXiv; this version has not had a full claim-by-claim audit. Multi-agent LLM systems are increasingly deployed in settings where individual agent failures can cascade across the collective, yet the mechanisms by which such failures propagate, and the structural conditions under which they are contained or corrected, remain poorly characterised. This paper synthesises seven findings from recent cs.MA and cs.DC preprints to argue a candidate structural pattern (a heuristic reading, not a derivation from a shared formal structure): systemic reliability in multi-agent LLM systems is not primarily a function of individual agent capability, but of the interaction between (a) the topology through which errors can propagate, (b) the credit-assignment mechanisms that identify which components generated the error, and (c) the architectural isolation boundaries that prevent a failure in one execution channel from contaminating others. We argue the analogy by naming specific mechanisms in each case rather than by appeal to a unified formalism. The corpus draws from cs.MA papers on multi-agent deliberation, self-evolution, coordination policy learning, and governance infrastructure, supplemented by cs.DC work on federated orchestration and zero-trust enforcement at physical actuation boundaries. Key findings include: deliberative consensus among LLM agents can degrade accuracy below single-model baselines when high-confidence wrong agents flip correct ones arXiv:2605.30802; architectural channel isolation failures silently block cross-agent memory injection regardless of agent-level correctness arXiv:2606.04896; governance layers at the execution boundary reduce unsafe executions from 88% to near-zero without modifying underlying generators arXiv:2606.04306; and temporal plus structural credit decomposition substantially reduces query complexity in MAS optimisation arXiv:2605.30227. Together these findings suggest that failure propagation and self-correction are topology-dependent phenomena that cannot be addressed by improving agent intelligence alone. Falsification path: if deliberative degradation disappears when inter-agent error correlation is experimentally reduced below 0.3, the error-propagation mechanism is confirmed; if it persists, the topology hypothesis is insufficient and capability variance must be the primary driver. Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted synthesis from an arXiv preprint corpus, originally drafted 2026-06-08, 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.