Large language model (LLM) agents are increasingly operating in multi‑agent systems (MAS) where strategic interaction is essential. This study evaluates four popular LLMs playing four games with distinct cooperation equilibria, comparing natural‑language messages, numerical signals, and random sequences, and examining the impact of assigned personalities. Results show that structured messages significantly alter final payoffs for most games and models, yet the effect follows no predictable pattern, challenging the assumption that AI agents will converge to stable equilibria when endowed with extra communication capabilities. Moreover, when agents are explicitly instructed to communicate, their generated numerical messages deviate strongly from randomness; the symbol distributions align closely with the payoff structure and become more concentrated with repeated exchanges, but remain difficult for humans to interpret. Consequently, monitoring AI coordination should prioritize model‑agnostic message‑level fingerprints over raw behavioural choices.
Review: Numerical signalling offers both a hidden coordination channel and an interpretability hurdle, urging designers of safe AI systems to focus on message patterns rather than just action outcomes.