Scholarlink

<p>This paper presents a PRISMA-based systematic literature review of recent developments in AI agent technologies, with a particular focus on large language model (LLM)-based agents. Based on 30 peer-reviewed publications from 2019–2025, the review examines architectural paradigms, methodological practices, evaluation strategies, and emerging multi-agent frameworks. The analysis identifies recurring limitations across the field, including the absence of shared benchmark standards, limited long-horizon evaluation, insufficient robustness in open or dynamic environments, and weak reproducibility of agent trajectories. The core position of this paper is that robust and trustworthy agent systems cannot rely on LLMs alone. Instead, they require hybrid architectures that integrate language-based reasoning with reinforcement learning, symbolic control, and explicit planning mechanisms. While LLMs provide flexibility and semantic reasoning, reinforcement-learning and symbolic components contribute stability, verifiability, and safety-oriented control. Trustworthy agent systems must further be evaluated not only by final outputs but also by their execution processes, including planning quality, tool use, error recovery, and auditability. Overall, the findings position hybrid architectures and standardized evaluation as essential foundations for advancing LLM-based agents from isolated demonstrations toward reliable real-world applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-based agent systems: a systematic literature review and position on hybrid architectures and evaluation frameworks

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Christoph Gabauer

摘要

This paper presents a PRISMA-based systematic literature review of recent developments in AI agent technologies, with a particular focus on large language model (LLM)-based agents. Based on 30 peer-reviewed publications from 2019–2025, the review examines architectural paradigms, methodological practices, evaluation strategies, and emerging multi-agent frameworks. The analysis identifies recurring limitations across the field, including the absence of shared benchmark standards, limited long-horizon evaluation, insufficient robustness in open or dynamic environments, and weak reproducibility of agent trajectories. The core position of this paper is that robust and trustworthy agent systems cannot rely on LLMs alone. Instead, they require hybrid architectures that integrate language-based reasoning with reinforcement learning, symbolic control, and explicit planning mechanisms. While LLMs provide flexibility and semantic reasoning, reinforcement-learning and symbolic components contribute stability, verifiability, and safety-oriented control. Trustworthy agent systems must further be evaluated not only by final outputs but also by their execution processes, including planning quality, tool use, error recovery, and auditability. Overall, the findings position hybrid architectures and standardized evaluation as essential foundations for advancing LLM-based agents from isolated demonstrations toward reliable real-world applications.