Final Thesis: A Modular Framework for Scalable and Auditable LLM-Based Multi- Agent Systems

Abstract: Large Language Models have recently enabled new forms of intelligent systems that go beyond traditional question answering or text generation. When embedded into agent-based systems, these models can reason, plan, and interact with external tools. However, deploying such systems in real-world and industrial contexts remains challenging. Existing agent implementations often lack transparency, scalability, and reliable mechanisms for
human oversight. As a result, their adoption in safety critical or operational environments is limited. This thesis presents a modular framework for scalable and auditable LLM-based multi-agent systems. The framework is
designed to treat orchestration, memory management, observability, and human intervention as first class system concerns. Rather than focusing on model level improvements, the framework emphasizes system level design that enables controlled execution, traceability, and reuse across different applications. The framework is evaluated through two applied case studies. The first case study focuses on document driven analysis for a methanol factory construction scenario, where multiple specialized agents collaborate to interpret requirements and generate
structured outputs. The second case study addresses operational diagnostics and communication support for a Terminal Management System, where selective human approval is required before executing external actions. The results show that the proposed framework improves transparency, supports scalable knowledge access, and
enables safe integration of automated reasoning into real workflows. The thesis demonstrates that effective
deployment of LLMbased multi-agent systems depends primarily on architectural and engineering decisions rather than on model capabilities alone.

Keywords: LLM, AI, Agents, Agentic Software Development

PDF: Master Thesis

Reference: Mohammad Ayaz Alam. A Modular Framework for Scalable and Auditable LLM- Based Multi- Agent Systems. Master Thesis. Friedrich-Alexander-Universität Erlangen-Nürnberg: 2026.