This paper presents a novel VR (Virtual Reality) trained BDI (belief, desire, intention) software agent used to construct crowd simulations for emergency response. The BDI framework allows modeling of human behavior with a high degree of fidelity. The proposed simulation has been developed using AnyLogic software to mimic crowd evacuation from an area under a terrorist bomb attack. The attributes that govern the BDI characteristics of the agent are studied by conducting human in the loop experiments in VR using the CAVE (Cave Automatic Virtual Environment). To enhance generality and interoperability of the proposed crowd simulation modeling scheme, input data models have been developed to define environment attributes. Experiments are also conducted to demonstrate the effect of various parameters on key performance indicators such as crowd evacuation rate and densities.