Comparative Analysis of Multi-Agent Frameworks for Requirements Engineering: Characterizing and comparing open-source frameworks for supporting requirements engineering tasks

Abstract

Large language model (LLM)-based agents are increasingly discussed as support for software engineering and requirements engineering (RE), but it is not always clear how practical multi-agent frameworks should be characterized for RE tasks. This thesis compares LangChain and AutoGen as open-source frameworks for building agent-based LLM applications. The work is analytical and literature-based: it synthesizes research on generative AI in RE, LLM-based multi-agent systems, and framework documentation, and compares the frameworks through agent architecture, workflow composition, memory, integration and tool support, and human-in-the-loop support. The analysis finds that LangChain is best characterized as a graph and middleware-oriented framework for controlled agent workflows, while AutoGen is best characterized as a runtime and messaging-oriented framework for conversational agent teams. LangChain appears more suitable for RE workflows that require artifact grounding, repeatable process stages, review checkpoints, and governance. AutoGen appears more suitable when role-based collaboration, stakeholder-like dialogue, and flexible agent-to-agent interaction are central. The thesis concludes that framework choice for RE should be treated as conditional on the task context rather than as a universal ranking. The contribution is conceptual and should be validated through future empirical implementation studies.