The paper presents the performance of various multi-agent systems, such as AutoGen, CrewAI, and MetaGPT; So, how is it reflected in the code?