智能体化人工智能：架构、应用及未来方向的综合综述

Mohamad Abou Ali1,2,3· Fadi Dornaika1,4· Jinan Charafeddine Received: 20 July 2025 / Accepted: 7 October 2025© The Author(s) 2025 Abstract Agentic AI represents a transformative shift in artificial intelligence, but its rapid advance-ment has led to a fragmented understanding, often conflating modern neural systems withoutdated symbolic models—a practice known asconceptual retrofitting. This survey cutsthrough this confusion by introducing a novel dual-paradigm framework that categorizesagentic systems into two distinct lineages: the symbolic/classical (relying on algorithmicplanning and persistent state) and the neural/generative (leveraging stochastic generationand prompt-driven orchestration). Through a systematic PRISMA-based review of 90studies (2018–2025), we provide a comprehensive analysis structured around this frame-work across three dimensions: (1) the theoretical foundations and architectural principlesdefining each paradigm; (2) domain-specific implementations in healthcare, finance, androbotics, demonstrating how application constraints dictate paradigm selection; and (3)paradigm-specific ethical and governance challenges, revealing divergent risks and miti-gation strategies. Our analysis reveals that the choice of paradigm is strategic: symbolic KeywordsAgentic AI· Artificial intelligence· Systematic review· Neuralarchitectures· Symbolic AI· Multi-agent systems· AI governance· Neuro-symbolic AI 1Introduction The field of artificial intelligence (AI) is undergoing a paradigm shift from the developmentof passive, task-specific tools toward the engineering of autonomous systems that exhibitgenuine agency. Modern Agentic AI systems (Wissuchek and Zschech2025; Viswanathanetal.2025) are defined by capabilities such as proactive planning, contextual memory,sophisticated tool use, and the ability to adapt their behavior based on environmental feed-back. These systems operate not as mere solvers but as collaborative partners, capable of To establish a precise conceptual foundation, we distinguish between the field’s coreconcepts. AnAI Agent(or asingle-agent system) is a self-contained autonomous systemdesigned to accomplish a goal. It operates primarily in isolation, though it may interact withtools and APIs. Its agency is defined by itsautonomy,proactivity, and its ability to complete For example, a single, powerfulLLM-based (large language model-based)agent taskedwith “Write a full project proposal for a new mobile app” would autonomously break down In contrast,Agentic AIis the broader field and architectural approach concerned with cre-ating systems that exhibit agency. Crucially, this often involves the orchestration ofmulti-agent systems (MAS), where multiple specialized agents work together, coordinating and For example, an Agentic AI system designed for the same task would employ a teamof specialized agents: aproject manager agentto break the goal into tasks, aResearcherAgentto gather market data, awriter agentto draft content, and aquality assurance agenttoreview the output. Their collaborative workflow is the embodiment of Agentic AI. In summary, one can conceptualize anAI Agentas a single, sophisticated worker, whileAgentic AIrepresents the principle of leveraging agency, frequently by architecting and This rapid evolution, however, has led to a fragmented and often anachronistic under-standing of the field. A critical issue identified in prior reviews isconceptual retrofitting—the misapplication of classical symbolic frameworks (e.g., Belief–Desire–Intention (BDI)(Archibald etal.2024),perceive–plan–act–reflect (PPAR)loops [Zeng etal.2024; Erdoganetal.2025)] to describe modern systems built onlarge language models (LLMs)(Plaatetal.2025), which operate on fundamentally different principles of stochastic generationand prompt-driven orchestration. This practice obscures the true operational mechanics of Several recent reviews have explored aspects of Agentic AI, but most fall short of captur-ing its full scope or addressing the core challenge of conceptual retrofitting. As summarizedin Table1, existing surveys are often limited in scope, focusing on specific technical aspects, Several recent reviews have explored aspects of Agentic AI, but most fall short of captur-ing its full scope or addressing core challenges. Table1summarizes their focus, contribu- This paper addresses these gaps by first establishing a clear historical context (Fig.1),which delineates the evolution of AI through five distinct but overlapping eras. TheSymbolic AI Era (1950s–1980s)(Liang2025) established the foundational ambi-tion of artificial intelligence, grounded in logic and explicit human knowledge. This periodwas dominated by rule-based systems and expert systems such as MYCIN and DENDRAL(Swartout1985), which operated on carefully hand-crafted symbolic rules. Intelligence was TheMachine learning (ML) Era (1980s–2010s)(Thomas and Gupta2020; Nithya etal.2023; Trigka and Dritsas2025) marked a pivotal sh