Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems

Multi-agent LLM systems in production fail at rates between 41% and 87%, with the majority of failures traced to coordination defects rather than base-model limitations. Existing solutions—either empirical failure catalogues or declarative orchestration frameworks—lack a principled mapping from coordination configuration to failure-mode signature. In a new paper, researchers Maksym Nechepurenko and Pavel Shuvalov argue that coordination should be treated as a configurable architectural layer, separate from agent logic and information access, enabling architectural reasoning over engineering workarounds.

The team instantiated this with an information-controlled design on prediction markets. Using a single LLM (Claude Opus 4.6), fixed tools, and a fixed prompt template, they tested five coordination configurations on 100 resolved Polymarket binary markets. Murphy decomposition of the Brier score revealed distinguishable signatures even when aggregate scores matched, while two configurations dominated the cost-quality Pareto frontier. Code, traces, and production agents were open-sourced. The work positions itself as a methodology-validating first step, not a general cross-model claim.