Autonomous agents operating in shared environments face a fundamental tension between achieving their objectives and concealing their intentions from others who may exploit that knowledge. This research develops a theoretical framework and computational model to study adversarial dynamics in multi-agent systems where agents must balance goal achievement against strategic information management.

We model a sequential game between two classes of agents with asymmetric roles and capabilities. One agent class possesses private objectives and must take observable actions over time to achieve them, while simultaneously minimizing the information revealed through those actions. The opposing agent class observes only the environmental signals generated by these actions and attempts to infer the first agent's hidden objectives to gain competitive advantage. Crucially, the information available for inference is generated endogenously through the goal-directed agent's own behavior, creating a dynamic feedback loop between action selection and information leakage.

Our framework synthesizes insights from classical resource allocation games, signaling theory, and multi-agent reinforcement learning. We introduce temporal constraints requiring agents to achieve objectives within fixed horizons, creating non-trivial strategic tradeoffs.

We validate our theoretical framework through computational experiments using policy gradient methods, demonstrating emergent equilibrium behaviors including strategic timing of actions, adaptive belief formation, and an evolving detection-obfuscation arms race. Our results contribute to understanding strategic behavior in competitive multi-agent systems and demonstrate how reinforcement learning can illuminate complex dynamics arising from information asymmetry and adversarial interaction.