A Systematic Literature Review of Agentic AI: Definitions, Architectures, and Challenges

This Systematic Literature Review (SLR) synthesizes the rapidly evolving state of Agentic AI, analyzing peer-reviewed studies published between 2021 and 2025. We organize the current body of knowledge into a taxonomy comprising five distinct research fields: Memory Cognition, Networking Systems, Trust Safety, Evaluation Limits, and Applications Use-Cases. Our intersection analysis reveals a research landscape currently in an "exploitation" phase, heavily skewed toward Applications Use-Cases (31.9%) and the practical implementation of Autonomy. While the community is successfully operationalizing existing autonomous capabilities, we identify a critical "research desert" regarding Adaptability and New Models. The findings indicate a scarcity of architectural frameworks designed to support long-term resilience and dynamic behavior modification in unforeseen environments. To advance Agentic AI from brittle prototypes to robust real-world systems, we conclude that future research must pivot from solely demonstrating autonomy toward engineering adaptability and developing novel modeling paradigms that prioritize responsiveness to change.