BRIDGING CLASSICAL INTELLIGENT AGENTS AND LLM-BASED AGENTS: A SYSTEMATIC REVIEW OF HYBRID ARCHITECTURES

Abstract

This article analyzes and synthesizes different forms of integration between classical intelligent agents and large language model (LLM)-based agents through a systematic literature review of hybrid architectures. The methodology was based on Kitchenham and Charters' guidelines, using the PICOC framework to guide the research questions and the PRISMA guidelines as a reference for reporting the study selection process. Twenty scientific studies were analyzed, originally selected with a focus on recommender systems while also including studies from other domains relevant to the architectural investigation. The analysis identified three main configurations: integration between agents belonging to different paradigms, hybridization of classical and LLM-based mechanisms within individual agents, and partial combinations of computational components. The findings revealed different levels of integration, particularly in smart manufacturing applications, as well as limitations related to experimental validation, reliability, and computational costs. In the complementary context of recommender systems, none of the selected studies demonstrated direct integration between independent classical intelligent agents and LLM-based agents. The review concludes that characterizing agent roles and interaction mechanisms is essential to understanding the possibilities and limitations of these architectures, while more comprehensive comparative evaluations remain necessary.