Arkitektonisk kontroll i LLM-baserade AI-agenter

Sammanfattning

This study investigates how different agent architectures - single-agent, multi-agent, and human-in-the-loop - affect control, transparency, and reliability in systems based on large language models (LLMs). The study combines a literature review with an experimental prototype in which the three architectures are implemented and evaluated using a set of structured test cases. To enable a systematic comparison, a set of measurable evaluation dimensions is defined, including the number of identified issues, the number of implemented improvements, and the presence of factual errors in generated responses. These dimensions operationalize aspects of control and transparency by capturing how each architecture identifies, processes, and corrects deficiencies in its outputs. The results indicate that multi-agent architectures enable more structured internal review and iterative refinement, while human-in-the-loop provides the highest level of control and traceability through explicit human intervention. In contrast, the single-agent architecture is simpler but lacks explicit mechanisms for iterative validation and improvement. The study also highlights challenges in evaluating AI-generated responses, as the assessment process is partly interpretative and influenced by task characteristics. The findings suggest that there is no universally optimal architecture; instead, the choice involves trade-offs between efficiency, control, transparency, and system complexity.