Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems: An Information-Controlled Empirical Study on Prediction Markets

Abstract

Multi-agent LLM systems fail in production at rates between 41% and 87%, mostly due to coordination defects rather than base-model capability. Existing responses split between cataloguing failure modes empirically and shipping declarative orchestration frameworks as engineering tools; neither delivers a principled mapping from coordination configuration to predictable failure-mode signature. We argue that coordination should be treated as a configurable architectural layer, separable from agent logic and from information access, enabling architectural reasoning rather than only engineering productivity.

We instantiate this with an information-controlled design on prediction markets: a single LLM, fixed tools, fixed per-call output cap, and fixed prompt template across five reference coordination configurations, with total compute per question treated as an endogenous architectural output. The Murphy decomposition of the Brier score separates calibration error from discriminative power, so configurations leave distinguishable signatures even when aggregate scores coincide.

On 100 Polymarket binary markets resolved after the model's training cutoff (claude-opus-4-6) we report observed Murphy signatures, a cost-quality Pareto frontier, category-conditioned analysis, and a bootstrap power-projection that quantifies which architectural contrasts are resolvable on the existing sample. Three of five pre-specified Murphy-signature predictions are upheld in the predicted direction; two configurations dominate the Pareto frontier within this implementation and information regime; exploratory bootstrap intervals suggest separation primarily for consensus alignment versus other configurations, although pairwise tests do not survive Bonferroni correction at n=100. We additionally deploy the same five configurations as live agents on Foresight Arena under web-search-enabled conditions on real future events, providing an independent on-chain replication channel whose data accumulates in parallel. The harness, the open trace dataset, and the production-agent deployment are released as three public repositories.

We position this work as a methodology-validating first instantiation of the architectural-layer framework, not a general claim about cross-model or cross-domain architectural laws.

Keywords: multi-agent systems, LLM agents, coordination architectures, declarative orchestration, proper scoring rules, Murphy decomposition, prediction markets, observational power analysis JEL Classication: C53 (Forecasting Models

JEL Classification: C53, C18, G14, O33, D83