Rank Aware Strategic Policy Design for Language Based Economic Environments

Games in Language Based Economic Environments (GLEE) evaluates language agents in multiround bargaining, bilateral negotiation, and repeated persuasion. We report an agent that finished second in the 2026 agent track at a rating of 2590.02 after 46,564 officially rated games. Because GLEE ranks payoffs within each configuration and role, maximizing expected money need not maximize the competition score. Our final policy combines game-theoretic benchmarks, fitted opponent-response models, and history-based rules without a language model at inference. Bargaining balances discounted continuation against agreement risk; negotiation separates visible-surplus decisions from hidden-value inference; persuasion uses recommendations, observed quality, and credibility to guide purchases. We trace the development and failures of these strategies. A separate 245-game final-policy audit connects 2,288 returned actions to authenticated replays. Agreement alone was insufficient: all 15 bargaining games in which a rank-proxy rule changed offers agreed, yet 14 lost rating. These results describe a competitive system, not individual component effects.