In critical infrastructure environments, we argue that both adversaries and operators will utilize agents to manage dynamic attack/defence interactions in future. Agent behavior and, in particular, agent interaction require adequate modelling tools to reason over such situations in distributed environments where the state (malicious or non-malicious) of a channel or process can vary dynamically depending on the actions of opposing sides in attack and defence. For this purpose, we propose an extension to applied $\pi$-calculus to model agent behavior. We apply this extended calculus to the formal analysis of a class of agent-based attacks and its detection to demonstrate its utility..