The study reported here centers around the question of how to evaluate the performance of creative intelligent machines. In this case, the prototype is a music improvising agent performing Free Music – a genre characterized by the lack of rules. Given this circumstance, evaluating the aesthetic quality of the agent's performance becomes even more difficult than for traditional music. The degree of ensemble communication between the agent and human musicians is used as a performance measure. Using foot pedals, the participating musicians are asked to track their perceived ensemble states by considering everybody's tension curves. The correlation over time between the ensemble states recorded by the musicians and by the machine is then evaluated. Additionally, an introspective approach to testing the agent is introduced by simulating a trio consisting of the agent and two other instruments (saxophone and bass pedal), both performed by the author.