Agentic AI: A Comprehensive Survey of Architectures, Applications, and Future Directions

TL;DR

The paper reframes Agentic AI through a dual-paradigm lens, distinguishing Symbolic/Classical and Neural/Generative lineages to prevent conceptual retrofitting and to enable precise analysis of autonomy, architecture, and governance. It employs a PRISMA-based review of 90 studies (2018–2025) to map foundational theories, contemporary architectures, and domain-specific deployments, revealing that Symbolic systems excel in safety-critical, verifiable tasks while Neural systems dominate data-rich, adaptive settings. The study highlights critical governance gaps for symbolic approaches and argues for hybrid neuro-symbolic architectures as the most viable path forward, supported by a paradigm-aware taxonomy and targeted policy implications. By articulating domain-sensitive evaluation metrics, coordination protocols, and a roadmap toward hybrid intelligence, the work provides a principled foundation for research, development, and policy toward robust, trustworthy agentic systems.

Abstract

Agentic AI represents a transformative shift in artificial intelligence, but its rapid advancement has led to a fragmented understanding, often conflating modern neural systems with outdated symbolic models -- a practice known as conceptual retrofitting. This survey cuts through this confusion by introducing a novel dual-paradigm framework that categorizes agentic systems into two distinct lineages: the Symbolic/Classical (relying on algorithmic planning and persistent state) and the Neural/Generative (leveraging stochastic generation and prompt-driven orchestration). Through a systematic PRISMA-based review of 90 studies (2018--2025), we provide a comprehensive analysis structured around this framework across three dimensions: (1) the theoretical foundations and architectural principles defining each paradigm; (2) domain-specific implementations in healthcare, finance, and robotics, demonstrating how application constraints dictate paradigm selection; and (3) paradigm-specific ethical and governance challenges, revealing divergent risks and mitigation strategies. Our analysis reveals that the choice of paradigm is strategic: symbolic systems dominate safety-critical domains (e.g., healthcare), while neural systems prevail in adaptive, data-rich environments (e.g., finance). Furthermore, we identify critical research gaps, including a significant deficit in governance models for symbolic systems and a pressing need for hybrid neuro-symbolic architectures. The findings culminate in a strategic roadmap arguing that the future of Agentic AI lies not in the dominance of one paradigm, but in their intentional integration to create systems that are both adaptable and reliable. This work provides the essential conceptual toolkit to guide future research, development, and policy toward robust and trustworthy hybrid intelligent systems.

Figures (8)

• Figure 1: Historical Evolution of AI Paradigms: This timeline charts the key breakthroughs and eras in AI, from early symbolic systems to the modern agentic era. It highlights the Transformer architecture as the pivotal enabling technology for large language models (LLMs), which in turn powered the generative AI revolution and provided the substrate for contemporary agentic systems.
• Figure 2: Conceptual Framework of Agentic AI's Dual Lineages. This taxonomy resolves conceptual retrofitting by distinguishing the Symbolic/Classical lineage (left), defined by algorithmic planning and persistent state, from the Neural/Generative lineage (right), defined by stochastic generation and prompt-driven orchestration. While both paradigms target similar applications, their underlying mechanisms are fundamentally incompatible. This framework provides the analytical structure for this survey.
• Figure 3: Classical symbolic reasoning: Comparison between a rule-based MDP scheduler (left) and a belief-based POMDP assistant (right). The MDP agent relies on explicit calendar states and deterministic policies, while the POMDP agent infers hidden user preferences from behavioral feedback. Both represent the symbolic paradigm's approach to decision-making.
• Figure 4: The shift toward learned behavior: Architectural contrast between vanilla DRL (single-task optimization) and meta-DRL (dual-loop generalization). The latter improves adaptability across tasks through meta-optimization loops, moving from explicit programming toward learned, emergent capabilities.
• Figure 5: The journey from symbolic to neural agency: The evolution of a personal assistant from a deterministic rule-based (MDP) system, to an uncertainty-aware (POMDP) system, and finally to a modern LLM-orchestrated agent. This journey bridges the two paradigms, ending with a system that exhibits intelligent behavior through entirely different mechanisms.