A Scoping Review of Agentic AI: Applications, Emerging Trends, Risks & Future Directions

Agentic AI describes systems based on foundation models that extend generative AI from passive response generation toward executing tasks to achieve specific goals through planning, tool use, memory, reflection, multi-agent coordination, and environmental interaction. This scoping review maps recent peer-reviewed research on Agentic AI to clarify its conceptual foundations, application domains, technical patterns, risks, and future directions. Following PRISMA-ScR guidance, searches were conducted in IEEE Xplore, Scopus, and ScienceDirect for studies published between 2017 and June 2026. A total of 3549 records were identified, 3308 records were screened after duplicate removal, and 163 included studies were synthesized. The findings show that Agentic AI is commonly operationalized through LLM or foundation model reasoning combined with multi-step task decomposition, tool or API use, RAG, memory, reflection, and multi-agent coordination. Major application areas include cybersecurity, privacy and compliance; networks and telecommunications; robotics and embodied AI; software engineering, DevOps and EDA; and knowledge, data, and multimodal analytics. A post hoc ACM Digital Library and Scopus conference-coverage sensitivity analysis showed that application-domain proportions were sensitive to database and document-type coverage, whereas the main recurring architectural signals remained qualitatively stable. Recurring technical patterns include governance middleware, collaborative and hierarchical multi-agent architectures, memory-enabled agency, executable reasoning, and closed-loop validation. Reported risks include hallucination propagation, unsafe tool use, memory poisoning, privacy leakage, prompt injection, coordination failure, cost, latency, scalability, and human-oversight challenges. Future research should prioritize real-world validation, standardized benchmarks, runtime governance, secure memory management, human oversight, and cross-domain generalization.