Taylor Swift: Tailoring Strategy to Other Agents

We describe our agent for the GLEE competition (final rank 36th) through its squad's 100,631 logged games. A 24-hour probe showed that opponents acted on the text of a message in 1.2% of games: the arena is economic, not textual. Trace analysis showed that an opponent's offer depended on ours in a net 2–3% of turns: opponents play near-fixed schedules. We therefore built a nonparametric opponent model that identifies the counterpart from its opening offer and message, forecasts its remaining offers from its last three matching games, and waits only when a dominance test over the surviving candidate identities allows. In a leak-free back-test the forecast covered 37% of bargaining offers and landed within one pot point 95% of the time, well above a persistence baseline where the decision is made. The Agent Behavior Analysis then measures the gap between design and behavior: a forecast existed in 39% of games and fired in 11%, and under inflation the games we closed on the forecast route banked 9.9 pot points less than on other routes. The lesson: even a perfect acceptance threshold is only half a strategy; without tailored demands that get the opponent to accept ours it buys little, and under inflation waiting to accept its offer costs more than closing on ours.