How Behavioral Structure Shapes Agent Strategy in Language-Based Economic Games

In competitive multi-agent economic games, agents continuously interact, strategize, and make decisions within a shared market. Consequently, every message produced by a language agent changes both the current economic state of the market and the opponent's subsequent behavior. Although terminal rewards can indicate whether a decision was effective and broad behavioral labels can provide a coarse interpretation of the interaction, neither specifies the concrete response that should be taken next. We therefore ask whether explicitly providing an opponent's operational structure to a focal agent changes the value of its task-specific executable response strategies and the resulting game outcome. Using Games in Language-based Economic Environments (GLEE) as a controlled environment, we investigate this question across bargaining, negotiation, and persuasion games. We analyze 298,357 trajectories using seven relational profiles and decompose each task's dominant profile into ten Act·onomy structures, finding that dominant profiles have task-specific operational compositions. We then compare structure-blind and structure-aware policies in matched offline simulation and prospective online play. In fixed 334-game online cohorts for each task, the aware-minus-blind official-percentile point estimates are +2.52, +1.90, and +5.91 points in bargaining, negotiation, and persuasion, respectively. Taken together, the results distinguish strategy selection from response realization: structure re-ranks strategies, while in persuasion the online point estimate becomes positive only when the policy also changes the submitted message or purchase action.