Abstract
Autonomous agents and multi-agent systems have rapidly advanced with the integration of large language models (LLMs), enabling increasingly capable reasoning, planning, tool use, and collaboration. However, the field remains fragmented by inconsistent architectural terminology, limited cross-domain comparison, and heterogeneous evaluation practices. This review addresses these gaps through a two-stage PRISMA-informed search combining Scopus with structured manual screening, covering 107 peer-reviewed original studies published between 2018 and 2026. We propose a unified five-module architecture comprising profile, memory, planning, action, and security; systematically compare agent capabilities across ten application domains; and organize evaluation practices into six dimensions: task performance, efficiency, quality, robustness, safety, and interpretability. The review identifies two recurring patterns across domains: explicit integration of domain knowledge consistently improves performance, while planner–executor–reflector role specialization appears across diverse applications. However, technical performance often exceeds validated real-world utility, with robustness, safety, and interpretability remaining particularly under-evaluated. These findings provide a common basis for understanding, comparing, and developing autonomous agents and highlight priorities for more reliable and practically deployable systems.