Program Code Generation: Single LLMs vs. Multi-Agent Systems

This paper investigates the effectiveness of single large language models (LLMs) versus multi-agent systems in generating program code. Using AutoGen, we examine whether assigning detailed role descriptions enhances code quality in both setups. Key metrics such as lines of code, cyclomatic complexity, runtime, memory usage, and maintainability index are assessed. Python code generated by both systems is compared to human-written solutions across varying difficulty levels. Results show multi-agent systems hold potential for improving code quality, though the impact of detailed role descriptions warrants further exploration. Of the LLM-generated codes, 11 (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 6 \%}$</tex>) successfully solved the tasks, while 5 (21%) required minimal modifications, saving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2. 9 \%}$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 7. 9 \%}$</tex> of time compared to coding from scratch.