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

Battle Count: 50

Moderately relevant. The paper directly addresses prediction markets (Polymarket), which are closely related to financial markets and event-driven trading. The Murphy decomposition framework for separating calibration from discriminative power is directly applicable to probabilistic forecasting in trading. The finding that no configuration beats market consensus with web search disabled has implications for LLM-based trading signal generation. The cost-quality Pareto analysis is relevant for production trading systems. However, the paper focuses on binary event prediction rather than continuous price forecasting, and the negative-Alpha-everywhere result suggests limited direct edge for trading strategies in the tested regime. The architectural layer framework could inform multi-agent trading system design.

Moderate complexity. The harness is implemented in TypeScript as a config-driven wrapper around the Anthropic SDK. Five coordination configurations require implementing different communication topologies, synchronization regimes, and aggregation rules. The Murphy decomposition and bootstrap analysis add statistical complexity. However, the paper explicitly states the specification is implementable atop existing frameworks (AutoGen, CrewAI, LangGraph, AWS Strands) without modification. The released code and pinned commits significantly reduce implementation burden for reproduction.

Exceptional reproducibility: three public repositories released (harness, trace dataset, production agents) with pinned commit hashes and version tags (paper-v05). Complete reasoning traces (~17.1M tokens) released as open dataset under CC-BY 4.0. TypeScript harness is config-driven and allows swapping LLM backend, tool stack, or question source. Pre-specified predictions timestamped in public Git history. Analysis pipeline including bootstrap and power-projection scripts included.