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

Multi-agent LLM systems fail in production mostly due to coordination defects rather than base-model capability. We treat coordination as a configurable architectural layer, separable from agent logic and information access, and test five reference configurations on prediction markets under fixed models, tools, output caps, and prompts. Murphy decomposition separates calibration from discrimination, while total compute is treated as an endogenous architectural output. On 100 Polymarket binary markets resolved after the model training cutoff, three of five pre-specified predictions are upheld in direction and two configurations dominate the cost-quality Pareto frontier within this regime. Live Foresight Arena deployments provide an on-chain replication channel. The contribution is a methodology-validating first instantiation, not a general cross-model claim.