The Rise of Agentic AI: A Review of Definitions, Frameworks, Architectures, Applications, Evaluation Metrics, and Challenges

Analysis

Why This Paper Matters

This paper arrives at a critical juncture in AI development, where the shift from generative models to autonomous, goal-driven systems is accelerating. As AI practitioners move beyond chatbots and content generators, understanding agentic AI—systems that can plan, reason, and act independently—becomes essential. The authors provide a much-needed synthesis of a fragmented field, drawing from 143 studies to create a coherent map of definitions, architectures, and evaluation methods. This is particularly valuable for teams building LLM-based agents or multi-agent systems, as it clarifies the distinctions between agentic AI, generative AI, and autonomic computing.

The paper also addresses the practical challenges of deploying such systems, including coordination, ethics, and security. For Neura Market's audience of AI practitioners, this review serves as a reference guide for designing more robust and responsible agentic systems, especially as the industry moves toward autonomous decision-making in areas like robotics, finance, and healthcare.

Technical Contributions

• Taxonomy of Agentic AI: Clearly distinguishes agentic AI from generative AI, autonomic computing, and multi-agent systems, providing a unified definition centered on autonomy, adaptability, and goal-driven reasoning.
• Architectural Classification: Categorizes architectural models (e.g., LLM-based vs. non-LLM) and input-output mechanisms, with tabular summaries that map to real-world case studies.
• Evaluation Framework: Proposes a dual classification of evaluation metrics into qualitative (e.g., coherence, safety) and quantitative (e.g., task completion rate, latency), along with testing methods for reliability.
• Challenge Mapping: Systematically identifies technical, architectural, coordination, ethical, and security challenges, offering a structured problem space for future research.
• Comprehensive Literature Review: Analyzes 143 primary studies, covering both LLM-driven and traditional agentic systems, ensuring broad coverage of current approaches.

Results

The paper does not present new experimental results but synthesizes existing knowledge into a structured foundation. Key outputs include: a comparative table of agentic AI vs. related fields, a classification of architectural models with example applications, and a taxonomy of evaluation metrics. The review highlights that LLM-based agents excel in planning and memory but face challenges in coordination and ethical alignment. Non-LLM agents remain relevant for deterministic tasks. The authors note that most evaluation metrics are task-specific, with no universal benchmark yet established.

Significance

This review is a foundational resource for the AI community, particularly for practitioners designing autonomous systems. By organizing the fragmented landscape of agentic AI, it enables more informed decisions about architecture selection, evaluation, and deployment. The structured challenge mapping also guides future research toward critical gaps, such as ethical safeguards and multi-agent coordination. For Neura Market's audience, this paper offers a practical roadmap for building responsible, adaptable agentic systems that can operate reliably across diverse domains.