Interactive real-time virtual worlds benefit from agents with complex shapes and geometry, imbued with a repertoire of expressive actions suitable for engaging narrative settings. Volumetric simulations like position-based dynamics, often used for motion picture graphics or interactive material simulations, offer a way towards increased expressivity. In this paper, we describe a proof-of-concept real-time simulation testbed that integrates animated agent behaviors in narrative contexts with deformable geometries, specifically GPU-accelerated volumetric soft body dynamics, and custom planners that expand the agents’ behavioral action space. GPU-based simulation models are not typically interfaced with agent prototyping testbeds. The aim is to introduce these simulations in a modular hierarchical event-driven agent prototyping framework using minimal message passing between the CPU and GPU. As a result, more diverse animations, beyond skeletal animation and Inverse Kinematics, are afforded to virtual agents that can be incorporated directly in a narrative setting.