From Rules to Reasoning: Evolving Agentic AI for Strategy Synthesis in Multi-Agent Wargaming Environments

Abstract

Contemporary conflict environments are increasingly characterized by autonomy, decentralization, and rapid adaptation. Traditional rule-based artificial intelligence systems struggle to operate effectively in such settings due to their reliance on pre-specified logic and static assumptions. Reinforcement learning (RL) has improved adaptability, yet many implementations remain reactive and myopic, optimizing local rewards rather than synthesizing higher-level strategies. This paper presents an incremental research program exploring the evolution of agentic artificial intelligence within a custom digital wargaming environment. Beginning with deterministic rule-based agents, progressing through multi-agent reinforcement learning (MARL) with memory augmentation, and culminating in an emerging paradigm of self-directed, research-capable agents, this work examines how autonomy can be extended beyond action selection toward independent strategy discovery. Using a dynamic 3v3 drone “capture-the-flag" simulation, we demonstrate measurable gains in coordination, adaptability, and tactical effectiveness, and outline a next-phase architecture in which agents autonomously explore, evaluate, and synthesize strategies across episodes. These results contribute to ongoing efforts to move from reactive autonomy toward genuinely agentic systems suitable for complex, contested operational domains.