The earlier versions of the EVA project (Evolutionary Virtual Agent) were designed as C++ applications running under Windows. These versions included a simple real-time 3D engine for rendering the agent. The agent face 3D model used a small numbers of polygons (~100-500) focusing on a basic set of emotional expressions. This was inspired by the work done by Ken Perlin a few years ago with his "face demo". In the new implementation, we choose another graphical-rendering approach : a 2D animation engine that render the face in real-time by using a limited set of pre-computed 3D images. There are two main benefits of this approach: the face will be much more realistic without the need of a high-performance 3D engine, and the agent will be able to run in any web environment. The picture shows a preliminary design done by one of our student (thanks to Kevin). In the same meaning, we moved from a C++ “heavy” design to a “light” Java-based software architecture. The Eva “brain” engine is now based on a Scheme kernel (implemented in Java) which is well adapted for evolutionary experiments using the genetic programming paradigm.