This paper describes a machine-listening based system developed for an interactive composition for soprano saxophone and electronics. The auditory processing stage of the system consists of a feedforward Neural Network trained to perform real-time recognition of different playing techniques (single notes, multiphonics, air tones and slap tones). This classification algorithm is embedded in four interaction scenarios that entail different compositional instructions and listening modes, including selective listening modes and a non-listening state. The integration of a classification task in the auditory processing stage of the system has the purpose of shifting the focus of machine listening from sensory (signal-level features) to symbolic information (composerdefined sound classes), enabling the design of idiosyncratic agent behaviors in the context of composed, scenario-based sonic interaction.