MAS-Consensus-v1 is a computational research framework designed to study consensus formation, disagreement persistence, adaptive trust evolution, and emergent interaction dynamics in multi-agent large-language-model (LLM) systems. The framework models five specialized agents -- technical analyst, sentiment analyst, risk assessor, macroeconomic analyst, and contrarian critic that participate in structured multi-round negotiations over financial and macroeconomic scenarios. The system incorporates adaptive trust updating, disagreement propagation, ideological clustering, and consensus-weighting mechanisms inspired by computational social science, collective intelligence research, and decentralized coordination systems. Experiments were conducted across 120 benchmark scenarios involving recession forecasting, inflation shocks, speculative bubbles, banking crises, commodity volatility, geopolitical instability, and policy transitions. The experiments demonstrate measurable and reproducible emergent dynamics across agent interactions. Key findings include: Moderate but adaptive trust evolution with mean trust volatility VT = 0.0342. Bimodal disagreement persistence patterns, clustering around 0.33–0.42 in low-conflict regimes and approximately 0.67 in high-conflict environments. Near-zero coalition density across most experiments, indicating persistent disagreement without stable faction consolidation. Strong SELL-oriented consensus bias in both heuristic and LLM-driven modes, with LLM-enhanced experiments increasing SELL outcomes from 60.6% to approximately 68–74%. The framework provides a reproducible experimental environment for analyzing artificial agent societies and emergent coordination behaviors. More broadly, the work contributes toward computational approaches for studying collective intelligence, adaptive consensus systems, trust-aware multi-agent reasoning, and emergent decision-making dynamics in decentralized AI systems. Repository contents include: Research paper and methodology Experimental framework and source code Simulation outputs and benchmark results Reproducibility instructions and documentation
Vansh Mahajan (Sun,) studied this question.