Tarot and I-Ching prompts reshape LLM risk-taking in multi-agent games

A new paper by Augustin Chan (arXiv:2606.07552) reveals that injecting symbolic reasoning frameworks into a single LLM agent dramatically reshapes multiplayer game dynamics. In a 7-player Warring States Diplomacy variant, the LLM agents played 41 games across four conditions: control, I-Ching yarrow divination, Tarot, and a scrambled-text ablation. Under control, Yan dominated (7/11 wins). But under I-Ching, Yan and Chu co-dominated while Qin was completely suppressed (0/10). Under Tarot, Qin dominated (5/10, Fisher p=0.006). The framework-receiving agent (Han) never won, but Tarot consistently elevated Han's peak territory (mean 3.0 SCs vs. 2.1–2.5, Kruskal-Wallis p=0.010).

Crucially, the content of the divination prompts didn't predict agent actions—hexagram themes and Tarot card postures were independent of choices (p=0.95 and p=0.69). Instead, the reflective process of reasoning itself modulated LLM risk aversion. This suggests that alignment-framework choice at the agent level produces distinctive system-level consequences in multi-agent settings, with implications for autonomous AI coordination and AI safety research.