A2: Architectures for Code Development with LLMs

Comparative analysis of single-agent vs multi-agent LLM architectures for code generation tasks.

Project overview

• A (Application): projects in which you will analyze the effectiveness of the application of LLMs in various Software Engineering tasks.
• Responsible: prof. Riccardo Coppola

Project Track:

• LLMs can generate code, but single-prompt interactions often fail on long or complex development tasks Multi-agent architectures may improve quality by splitting responsibilities (design, planning, writing, reviewing, debugging).

Task Selection Select 10-20 programming tasks from:

• Public datasets (HumanEval, MBPP, CodeNet subsets)
• Past course assignments
• Open-source code snippets

System Implementation

• Single-Agent Baseline: One LLM generating complete code solutions
• Multi-Agent System: ≥2 specialized agents with distinct roles (e.g., planning, development, review, testing)

Evaluation Metrics Choose at least one evaluation method:

• Functional correctness (unit tests or provided test suites)
• Static code quality metrics (complexity, maintainability indices)
• Debugging performance (fault detection and fixing capabilities)
• Readability and maintainability assessment

This project evaluates the effectiveness of different LLM architectures in software engineering tasks, comparing:

• Single-Agent Baseline: One LLM generating complete code solutions
• Multi-Agent System: Multiple specialized agents with distinct roles (planning, development, review, testing)

Research Questions

1. Which architectures produce higher-quality and more maintainable code?
2. How do agent coordination strategies impact correctness?
3. Does modular role separation improve code generation?