Recent developments in Artificial Intelligence have marked the transition from purely generative models to autonomous agentic systems (Agentic AI), capable of planning, reasoning, and executing complex tasks by integrating external tools. However, the rapid proliferation of orchestration frameworks has created a significant gap between stated capabilities and empirical evidence, leaving software designers without objective criteria for technology selection. This thesis presents a systematic comparative analysis of three market leading frameworks: LangChain, CrewAI, and AutoGen. The research methodology adopts a rigorous experimental approach, based on the implementation and measurement of a benchmark application of real complexity: an automated system for correcting open ended university exams, powered by RAG (Retrieval-Augmented Generation) pipelines. Keeping the foundation model and domain logic constant, the study isolated and quantified the impact of different orchestration architectures: procedural, hierarchical, and conversational on system performance. Empirical results highlight distinct architectural trade-offs: LangChain demonstrates lower orchestration latency, making it ideal for high-speed batch processing, but at the cost of greater code complexity. CrewAI excels in development ergonomics and operational reliability, achieving perfect success rates thanks to structured validation, positioning itself as the optimal choice for mission-critical processes. AutoGen, despite having longer total execution times, offers the best token consumption efficiency at scale, representing the most economical solution for massive deployments. The work concludes by providing decision making guidelines based on quantitative data, offering software engineers structured criteria for selecting the most suitable framework based on specific constraints of cost, robustness, and maintainability.
