A Research Landscape of Agentic AI and Large Language Models: Applications, Challenges and Future Directions

Abstract

Agentic AI and Large Language Models (LLMs) are transforming how language is understood and generated while reshaping decision-making, automation, and research practices. LLMs provide underlying reasoning capabilities, and Agentic AI systems use them to perform tasks through interactions with external tools, services, and Application Programming Interfaces (APIs). Based on a structured scoping review and thematic analysis, this study identifies that core challenges of LLMs, relating to security, privacy and trust, misinformation, misuse and bias, energy consumption, transparency and explainability, and value alignment, can propagate into Agentic AI. Beyond these inherited concerns, Agentic AI introduces new challenges, including context management, security, privacy and trust, goal misalignment, opaque decision-making, limited human oversight, multi-agent coordination, ethical and legal accountability, and long-term safety. We analyse the applications of Agentic AI powered by LLMs across six domains: education, healthcare, cybersecurity, autonomous vehicles, e-commerce, and customer service, to reveal their real-world impact. Furthermore, we demonstrate some LLM limitations using DeepSeek-R1 and GPT-4o. To the best of our knowledge, this is the first comprehensive study to integrate the challenges and applications of LLMs and Agentic AI within a single forward-looking research landscape that promotes interdisciplinary research and responsible advancement of this emerging field.

Keywords: Agentic AI; large language models (LLMs); GPT-4o; DeepSeek-R1; technical, legal, and ethical challenges; future Agentic AI research; multi-agent systems; next-generation AI.

1. Introduction

Large Language Models (LLMs) have emerged as one of the most significant breakthroughs in Artificial Intelligence (AI), attracting considerable interest from researchers, developers, and policymakers, and their accelerated evolution is reshaping the way language is understood and generated. Yet a notable lack of peer-reviewed research remains at the intersection of LLMs and Agentic AI, where terms such as "Agentic AI", "Agentic LLMs", "Agentic Systems", and "LLM-based agents" are often used interchangeably due to the lack of clear, standardised definitions in this emerging field. In the context of this study, these terms broadly refer to LLM-powered Agentic AI, where LLMs serve as the reasoning core while agents rely on tools to take action.

To demonstrate how Agentic AI and LLMs operate in practice, the paper presents a simplified flight-booking example (Dubai to London): the LLM decomposes instructions into subtasks through the orchestrator layer — a search agent retrieves options via APIs such as Skyscanner, a filtering agent applies weather and price data from services like OpenWeatherMap, and a booking agent checks user availability through tools such as Google Calendar. While LLMs provide interpretation and planning, execution depends on agentic components that interface with external APIs, maintain local memory, and track progress toward goals, with minimal human intervention.

Due to this close connection, Agentic AI inherits many LLM limitations — security, privacy, trust, bias, misinformation, energy consumption, value alignment, transparency, and explainability — while introducing new challenges such as context management and memory persistence, goal misalignment, opaque decision-making, limited human oversight, multi-agent coordination, long-term safety, and ethical and legal accountability. To the best of the authors' knowledge, no existing study brings together the conceptual, technical, and ethical aspects of LLMs and Agentic AI in a single coherent analysis.

The core contributions are: (1) a unified examination of LLMs and Agentic AI as deeply interconnected systems; (2) cross-domain application insights revealing real-world impact on productivity, efficiency, and autonomy; (3) empirical illustrations of model limitations through small-scale experiments with GPT-4o and DeepSeek-R1 on bias and cultural inclusivity; and (4) a forward-looking research landscape mapping technical and ethical challenges with corresponding research opportunities for both LLMs and Agentic AI.

2. Research Scope and Method

This study adopts a structured scoping review to explore the breadth of emerging developments in Agentic AI and LLMs, where much of the literature is recent, fast-evolving, and often not yet widely indexed in traditional academic databases. Although not formally a systematic review, the method was carefully planned and executed to maintain rigour and transparency, with clearly defined search strategy, inclusion and quality assessment criteria, and a documented screening process.

The first phase developed a targeted search strategy using keywords such as Agentic AI, Large Language Models, and their challenges across Google Scholar, Scopus, and arXiv — deliberately including preprints to capture early insights in this fast-moving field. Initial inclusion criteria (English-language publications, 2020–2025) yielded 327 documents: 142 from Google Scholar, 110 from Scopus, and 75 from arXiv. In the second phase, duplicates were removed using Zotero and full-text inclusion and quality criteria were applied, retaining studies focused on Agentic AI, LLMs, or their intersection with sufficient analytical depth.

The process produced 84 final sources: journal articles 48%, preprints 27%, conference papers 20%, and selected websites 5% — with a sharp increase in research activity from 2023 to 2025 (Figures 6–7). In the third phase, thematic information was extracted (the Agentic AI–LLM relationship, applications, challenges, and research directions), supported by keyword co-occurrence analysis using VOSviewer: explainable AI, transparency, misinformation, and ethics showed growing interest, while sustainability, security, and safety remain least explored. In the final phase, applications were grouped into six sectors and challenges divided into two categories: those arising from LLM design and limitations, and those unique to Agentic AI systems.

3. Applications of Agentic AI Powered by LLMs

Agentic AI represents a significant evolution in the design and deployment of intelligent systems, capable not only of following instructions but of pursuing goals, adapting to changing contexts, and exhibiting autonomy in decision-making. The paper examines six application domains revealing the real-world impact of these systems.

3.1. Education

Agentic AI is reshaping learning through tailored personalised support that adapts instructional strategies in real time, offering personalised feedback, multiple explanations of complex concepts, and encouraging metacognitive reflection. Recent deployments such as Khan Academy's Khanmigo use GPT-4 for Socratic-style tutoring that avoids direct answers, integrating external tools like search engines and calculators to address knowledge cut-offs and calculation weaknesses. These systems also support students with diverse needs, including learning disabilities and language barriers, offering a promising scalable solution for global education equity — particularly where qualified teachers are scarce.

3.2. Healthcare

LLM-powered Agentic AI is reshaping diagnostic processes and patient engagement, enhancing clinical efficiency and decision support by streamlining administrative burdens, automating approvals, and updating care plans in real time. These systems analyse vast patient records, summarise complex case histories, and provide preliminary diagnoses, while LLMs alleviate the cognitive load of imaging data and EHR complexity. A notable example is Med-Flamingo, a few-shot multimodal AI agent integrating 2D medical images with textual inputs, demonstrating up to 20% improvement in diagnostic accuracy. Transparency, robustness, and accountability remain pressing concerns as these systems assume more autonomous medical roles.

3.3. Cybersecurity

As cyber threats intensify in scale, speed, and complexity, Agentic AI enables real-time threat detection and prevention with continuous learning beyond static detection — agents rapidly assess incidents, prioritise risks, and execute containment actions such as revoking credentials or patching vulnerabilities. The study notes AI will soon transform from chatbots to agent-driven cybersecurity. Leading firms such as ReliaQuest and CrowdStrike already deploy Agentic AI to automate Security Operations Centre (SOC) tasks and accelerate response, alongside proactive vulnerability protection across the software development lifecycle. However, interaction with multiple external platforms broadens the attack surface, demanding robust governance and risk mitigation.

3.4. Autonomous Vehicles

Agentic AI constitutes a qualitative leap toward the genuine autonomy required for SAE Level 5 vehicles, empowering self-driving cars to perceive complex environments, plan manoeuvres, execute actions, and continuously learn without constant human oversight. LLMs serve as the cognitive engine for advanced reasoning, contextual understanding, human–vehicle interaction, adaptive learning, and interpretability. The technology remains at testing or pre-commercial stages, with primary concerns around reliability and predictability in safety-critical scenarios, data privacy, and the opacity of complex models.

3.5. E-Commerce

Agentic AI is transforming retail through hyper-personalisation, operational efficiency, and autonomous decision-making, with goal-oriented agents adapting to market conditions and optimising operations with minimal human intervention. LLM-powered chatbots and shopping assistants provide real-time support, reduce wait times, and enhance product discovery by interpreting user behaviour. This is among the most mature Agentic AI deployment domains, already in commercial use — though handling vast sensitive customer data keeps security and privacy risks prominent.

3.6. Customer Service

Organisations deploy LLM-backed agents to automate routine customer interactions around the clock, with intelligent escalation of complex cases to human staff. Conversational and contextual reasoning enables a more human-like interface that lowers service barriers and improves satisfaction, while challenges around output quality, accuracy, and trust persist.

4. LLMs: Challenges and Research Directions Impacting Agentic AI

The paper identifies five core LLM challenges that can propagate into Agentic AI systems, each paired with a research direction (Table 1 of the original paper).

4.1. Security, Privacy, and Trust Concerns: Toward Trustworthy LLMs

As LLMs enter critical domains such as healthcare, education, and finance, understanding how they handle privacy, security, and reliability is crucial. Future research requires privacy-preserving architectures, robust security auditing protocols, and trust frameworks promoting transparency and user confidence.

4.2. Misinformation, Misuse, and Bias: Toward Ethical LLMs

LLMs generate incorrect or fabricated content due to design limitations or flawed training data, and can be misused for misleading content, spam, and political propaganda — with no well-defined governance frameworks in place. Models also reflect training-data biases: testing GPT-4o (USA) and DeepSeek-R1 (China) with three rephrased prompts on one topic (treatment of religious minorities in China) across three different days produced divergent responses (Figures 9–14), concretely showing how limitations propagate from LLMs into Agentic AI and underscoring the urgency of addressing them at both model and system levels.

4.3. Energy Demands and Sustainability Issues: Toward Sustainable LLMs

Training and inference of large models consume enormous energy. Research directions include efficiency gains through pruning, quantisation, and inference optimisation, alongside transparency in energy reporting.

4.4. Value Misalignment and Exclusion: Toward Human-Centric LLMs

Inclusivity experiments with GPT-4o and DeepSeek-R1 (Figures 15–20) reveal divergent cultural responses. Future work calls for inclusive model design and culturally sensitive value alignment through diverse training datasets and human-in-the-loop oversight.

4.5. Explainability and Transparency Gaps: Toward Transparent LLMs

LLM internal reasoning is difficult to trace, undermining trust and accountability. Research directions include reliable explanation tools, input attribution methods, and evaluative standards for interpretability and accountability.

5. Agentic AI: Core Challenges and Research Directions

Beyond inherited LLM concerns, Agentic AI introduces eight novel challenges, each with a corresponding research direction (Table 2 of the original paper).

5.1. Security, Privacy, and Trust Concerns: Toward Trustworthy Agents

Agents interacting with external systems and sensitive data face amplified security risks. Research must design secure agent architectures incorporating sandboxed execution, scoped authentication, access control, privacy-preserving protocols, and security policies ensuring responsible interaction with external systems.

5.2. Goal Misalignment: Toward Aligned Agents

Among the most critical challenges: agent actions diverging from intended human objectives when goals are under-specified or ambiguous — e.g., an agent told to "minimise customer response time" sacrificing accuracy for speed. Solutions include inverse reinforcement learning, human preference modelling, and real-time feedback, with open questions on maintaining alignment over time in dynamic multi-agent scenarios.

5.3. Opaque Decision-Making: Toward Explainable Agents

Black-box behaviours reduce trust and limit oversight — worse in agentic contexts where decisions are part of multi-step plans. Needed: explainable autonomy via natural language rationales, saliency-based input highlighting, and structured audit trails, with explanations useful to non-expert users in real time.

5.4. Limited Human Oversight: Toward Collaborative Agents

Agentic AI risks reduced human control, especially under high autonomy. Research should design interaction frameworks supporting mixed-initiative control, human intervention, and shared decision-making.

5.5. Context Management: Toward Memory-Aware Agents

Agents need persistent memory architectures to retain, retrieve, and adapt based on long-term context rather than losing state across extended multi-step tasks.

5.6. Multi-Agent Coordination: Toward Cooperative Agents

Multi-agent systems require scalable coordination protocols with decentralised negotiation and shared governance mechanisms to avoid conflicts and undesired emergent behaviours.

5.7. Long-Term Safety: Toward Safe Agents

As agents take more autonomous real-world roles, long-term safety guarantees — preventing drift from original goals and unintended behaviours over time — become an urgent research priority.

5.8. Ethical and Legal Accountability: Toward Accountable Agents

Agent autonomy raises questions of responsibility for automated actions. Research needs ethical and legal accountability frameworks assigning responsibilities across developers, operators, and users, through cross-disciplinary technical, legal, and ethical collaboration.

6. Conclusions

In this paper, we examined how LLMs and Agentic AI can be applied across various domains, highlighting the key challenges associated with them and emphasising future research opportunities. As these AI models continue to integrate into real-world systems — from healthcare and education to finance and public policy — it becomes increasingly critical to understand not only what they can do but also the risks they pose. Despite advancements, issues related to bias, transparency, legal accountability, multi-agent coordination, security, privacy, trust, ethics, sustainability, opaque decision-making, human oversight, human-agent collaboration, memory architectures, global misalignment, and inclusivity remain persistent and pressing. These challenges are not specific to a particular model; they emerge from deeper design and deployment patterns of AI algorithms, so the proposed insights and research directions remain relevant across evolving architectures.

Through this forward-looking analysis, we offer researchers and practitioners a structured roadmap to navigate these developments, emphasising interdisciplinary collaboration to guide responsible and equitable development of LLM and Agentic AI technologies. The study offers an early glimpse of what lies ahead as LLMs shift from passive tools to more active roles in shaping decisions and driving real-world outcomes through Agentic AI, with reflections on the AI future we are collectively shaping — one demanding advanced development with careful attention to ethics, equity, and long-term societal impact.

In terms of limitations, the structured scoping review was guided by established principles of transparency, clear inclusion logic, and procedural rigour, but not a formal systematic review framework. While appropriate for capturing emerging trends across diverse, rapidly evolving sources, the paper recommends future work building on this landscape with deeper systematic and empirical studies.