GAME: A Governed Agent MacroEconomy As A Realistic Policy Sandbox

Macroeconomic simulation has progressed from agent-based models to systems whose agents are driven by large language models (LLM), yet existing LLM-based frameworks focus on reproducing economic phenomena or individual decision-making while paying little attention to the fidelity and robustness of simulated outputs, and hardly any supports policy analysis that jointly embeds central-bank and fiscal interventions. We propose the Governed Agent MacroEconomy (GAME), which couples a stock-flow-consistent economy of households, firms, banks, a central bank, and a government with LLM household cognition. Parameters are calibrated on historical sequences via the simulated method of moments, household groups are condensed from survey microdata, and each simulation step runs a four-layer cycle: perception layer from memory and neighbor signals, decision layer with LLM over the assembled information set, governance layer that anchors decision parameters to realistic conditions, and settlement layer through labor, goods, and credit market equilibria——along with the policy authorities periodically survey the economy and intervene with their instruments. On US and Canadian macroeconomic time series, under a protocol that strips all temporal , monetary and regional identity cues from every prompt, GAME attains state-of-the-art alignment with real series in statistical similarity and result robustness, acts as a reliable policy sandbox whose responses to policy changes match empirical evidence, replicates exogenous shocks of distinct types in line with classical economic principles, and its component ablations confirm that every mechanism is necessary. We believe that GAME can advance deeper research in the social sciences.