Generative to Agentic AI: Survey, Conceptualization, and Challenges

1 Introduction

• Agentic AI extends GenAI with multi-step, problem-dependent reasoning—including planning and reflection—and the ability to pursue broad objectives independently.The introduction contrasts this with earlier models such as ChatGPT 3.5, which typically responded immediately, and describes reasoning models such as ChatGPT o1 as spending minutes processing self-generated prompts.Source: §1 · p. 3
• Unlike GenAI’s direct input-to-output tasks, Agentic AI uses tools to act in environments, receive feedback, and guide subsequent actions.The introduction describes this interaction as incorporating reinforcement-learning elements and instant learning.Source: §1 · p. 3
• Autonomous memory, flexible tool selection, and open-ended exploration can make agents less predictable and harder to control and specify, despite their use of everyday language.Because agents act directly and handle less-detailed instructions, their solutions can be more diverse and difficult to control; the introduction calls for renewed, cautious assessment of AI risks and opportunities.Source: §1 · pp. 3–4
• Reasoning models show striking improvements on ARC and outperform other models in programming and radiology-report generation, but these gains do not establish broad generalization.The cited examples are ARC (Clark et al., 2018), programming (El-Kishky et al., 2025), and radiology reports (Zhong et al., 2024); the introduction cautions that test-time compute alone may not advance AGI, given data limitations and error accumulation.Source: §1 · p. 4, Figure 2 · p. 4
• Agentic AI’s advantages are uneven: it offers no dramatic improvement in every area, and deep-reasoning models can still err on simple tasks.The introduction reports only minor translation gains, a slight MMLU edge at substantially greater computational cost, and that domain-specific fine-tuned GenAI may outperform agents lacking domain knowledge at lower cost.Source: §1 · pp. 4–5
• The survey compares GenAI and Agentic AI, traces Agentic AI’s evolution, examines reasoning, tools, environments, agent specification, and evaluation, and outlines AGI-related research challenges and risks.Its scope includes single- and multi-agent systems, applications, methodology, and related surveys.Source: §1 · p. 5

2.1 Defining GenAI and Agentic AI

• A system with a foundation model, planning, reflection, and interaction or tool use can perform tasks beyond creating digital artifacts, unlike GenAI’s content-focused role.Agentic systems may affect physical objects or manage systems without producing a directly observable artifact.Source: §2.1 · pp. 6–7

2.2 Capabilities Compared

• Agentic AI’s defining capabilities are reasoning and interaction with environments and tools, enabling partial autonomy for multi-step tasks.GenAI is assistive for simple tasks and typically requires specific instructions for each interaction; Agentic AI’s autonomy and adaptability remain limited, including by human-oversight requirements.Source: §2.2 · pp. 8–9

2.3 Why Agentic AI?

• GenAI struggles with moderately complex tasks requiring multiple actions, while reasoning models can derive detailed steps from high-level goals.Sophisticated prompt engineering can partly compensate for GenAI’s execution limits, but reasoning models can improve usability by deriving their own steps.Source: §2.3 · p. 9
• Scaling data, models, and compute is not necessarily optimal because task-specific data can be expensive or impossible to obtain, and anticipating every task is infeasible.The paper identifies a significant lack of training data as a limitation of this scaling approach.Source: §2.3 · p. 9
• Modern Agentic AI systems extend limited early-model memory with larger context windows, retrieval, and trial-and-error learning from simulations or feedback.ChatGPT 3.5 handled about 4,000 words, while modern foundation models can handle millions of input tokens; agents can also use tool feedback such as compiler errors.Source: §2.3 · p. 10
• Agentic AI reasoning can reduce hallucinations through self-verification and external knowledge retrieval, while intermediate results can improve interpretability and reasoning may reduce bias.The paper notes that bias and interpretability remain unresolved GenAI shortcomings, while Agentic AI’s proposed improvements are not stated as guarantees.Source: §2.3 · p. 10
• A 1B-parameter model can outperform an 8B model in MATH-500 accuracy with more computation, and reasoning enables smaller models to approach large-model performance.Agents can allocate computation more finely based on intermediate outputs, supporting cost-efficient specialized models with variable inference effort.Source: §2.3 · p. 10, Figure 6 · p. 11

2.4 Evolution of GenAI to Agentic AI

• Agentic AI evolves GenAI’s shallow, prompted reasoning into planning, search, and reflection, while extending tool use and memory toward environment interaction and feedback-driven agents.The evolution also incorporates reinforcement learning and agent definitions that specify roles, workflows, and permitted tools; reasoning models can fill in more workflow detail than low-level GenAI prompts.Source: §2.4 · pp. 12–13

• Reasoning encompasses problem decomposition, verification, search, and planning.Source: §3 · p. 13
• Interaction covers engagement with the environment, tools, and memory.Source: §3 · p. 13
• The section also addresses how single-agent and multi-agent systems are specified.Source: §3 · p. 13

3.1 Reasoning

• Agents most often reason by generating stepwise intermediate solutions rather than immediate answers, potentially combining analysis, planning, validation, and refinement (Kojima et al., 2022; Wei et al., 2022b).Early prompted chain-of-thought was shallow and may rely on learned patterns rather than reasoning; simple CoT primarily helps with math and symbolic reasoning. Modern Agentic AI can reason more extensively by integrating planning algorithms and tools such as code interpreters.Source: §3.1 · p. 14

3.1.1 Decomposition

• Decomposition ranges from prompted step-by-step solutions to multi-agent planning and reflection; hierarchical approaches refine high-level plans into concrete actions or retrieve reusable guidelines.Reasoning-chain structure matters more than the content of individual steps, and chain length can matter more than component difficulty, though excessive verbosity may hurt effectiveness.Source: §3.1.1 · pp. 14–15

3.1.2 Reflection

• Reflection covers verification of outputs and refinement of prior solutions, using internal review or external feedback such as compiler results (Chen et al., 2023).Verification can compare alternative reasoning paths or intermediate results, while refinement methods iteratively incorporate feedback and prior outputs; simple prompted reflection without external input has been questioned (Huang et al., 2024; Zhang et al., 2024).Source: §3.1.2 · pp. 15–16

3.1.3 Planning and Search

• Planning and search extend reasoning by generating and evaluating partial solution paths, pruning weak ones, and using controllers or task-specific evaluators to manage the process.Tree of Thought uses depth-first or breadth-first search, while systems may seek diverse paths or use planning tools, Monte Carlo Tree Search, and heuristic planning.Source: §3.1.3 · pp. 17–18

3.1.4 Learning to Reason

• Reasoning can be trained using stepwise examples, expanded CoT demonstrations, and domain-specific data; DeepSeek-R1 combines CoT examples, GRPO, supervised fine-tuning, and alignment training (Guo et al., 2025).DeepSeek-R1’s GRPO rewards solution accuracy, formatting, and language consistency; including high-quality CoT examples made its outputs more human-readable.Source: §3.1.4 · pp. 18–19
• Preference training can improve human-LLM alignment and sometimes reasoning, but feedback quality is critical; oracle feedback on final answers can also yield rewards for intermediate steps (Ivison et al., 2024; Zhang et al., 2024).LLM-generated feedback can reduce collection costs, but it must be high-quality; these preference methods are used more commonly for helpfulness and safety than for reasoning.Source: §3.1.4 · p. 19

3.2.1 Memory Types and Structures

• Memory supports personalization and large inputs, multi-step reasoning, access to post-training information, and learning from failures through stored experiences.Large prompts once exceeded model context windows, and reasoning through sub-steps generates more tokens than simply outputting a solution.Source: §3.2.1 · p. 20
• Parametric memory encodes training data but is generally fixed for instant learning, while ephemeral, retrieval-based, and architectural memory provide other storage mechanisms.Online parameter updates risk catastrophic forgetting or malicious manipulation; ephemeral memory is bounded by context windows, retrieval-based memory offers virtually unlimited capacity, and MemGPT uses a hierarchy to load long-term information and discard temporarily unneeded content.Source: §3.2.1 · pp. 20–21
• Memory may be internal, external, or hybrid, and some systems accumulate the full chat or prior iterative feedback in each prompt rather than retrieving selectively.The chat example includes past user and model utterances; iterative feedback can enhance prompts across iterations.Source: §3.2.1 · p. 21

3.2.2 Retrieval Augmentation

• Retrieval augmentation selects relevant knowledge for model input, reducing hallucinations and enabling private or post-training data and dynamic learning from experience.Classical RAG indexes document chunks as vectors and appends retrieved snippets; chunk size, encoding, and number of chunks remain design challenges. Adaptive retrieval, relevance filtering (Zhang et al., 2024), and evaluation-driven retrieval (Yan et al., 2024) address these challenges.Source: §3.2.2 · p. 21
• Agent memory can store and abstract interaction experiences, while memory hierarchies, GPU key-value caching, and search integrated into reasoning address retrieval or computation needs.Zhao et al. (2024b) further processes stored experiences into abstract insights, while irrelevant details can impair retrieval efficiency (Zheng et al., 2023); MemGPT manages context like an operating system, and Chen et al. (2025) integrated search into reasoning without labeled data and GPRO.Source: §3.2.2 · p. 22
• Using an entire large context can outperform retrieving subsets but costs more to generate; alternatives include SQL databases, BM25, knowledge graphs, and specialized retrieval agents.BM25 is useful for exact matches, speed, interpretability, and simplicity; knowledge graphs represent interconnected or hierarchical information, and agents have been tailored to SQL and real-time web retrieval.Source: §3.2.2 · p. 22

3.3 Tools

• Tools complement agents by reducing errors, increasing efficiency, enabling new interactions, and embedding interpretable rules for transparency and oversight.Examples include calculators for basic math and browsers for online tasks such as shopping.Source: §3.3 · p. 22

3.3.1 Tool Creation

• LLMs can create tools from code, including Python utilities and tools derived from academic papers, with agents also able to create and execute scripts dynamically.Cai et al. (2023) propose tools from three examples, verify them with unit tests, and wrap them with documentation and examples; W¨olflein et al. (2025) convert code-containing papers into LLM-accessible tools, while Richards (2023) supports dynamic script creation and execution.Source: §3.3.1 · pp. 22–23

3.3.3 Tool Use

• Agents use software APIs, interpreters, generated code, SQL, and specialized LLMs as tools to execute tasks and route subtasks.Fine-tuned API calling covers calculators, translation systems, and search engines (Schick et al., 2023; Parisi et al., 2022a); later work improves call accuracy and supports out-of-distribution APIs (Qin et al., 2023), while self-play expands usage examples (Parisi et al., 2022b). Interpreters run generated code (Gao et al., 2023), code-based reasoning can be refined using environmental feedback (Wang et al., 2024; Qiao et al., 2023), SQL handles relational-database analysis (Li et al., 2023), and tasks can be routed to specialized LLMs.Source: §3.3.3 · p. 23

3.4.1 Reinforcement Learning

• Reinforcement-learning agents learn world models and policies from experience and feedback, balancing exploration with exploitation; agentic AI adopts these concepts without being limited to narrow reward objectives.Agentic systems may use actor-critic architectures; Voyager (Wang et al., 2023) illustrates continual Minecraft learning through exploration, skill building, feedback, and self-verification.Source: §3.4.1 · pp. 23–24

3.4.2 Interacting in (Virtual) Worlds

• Multimodal models may serve as world models, while virtual-home exploration has produced fine-tuned LLMs that perform better than much larger untuned models.LLMs can also guide planning with world-model rewards (Hao et al., 2023), and vision-language-action models learn from robot observation-action trajectories (Brohan et al., 2023).Source: §3.4.2 · p. 25
• Agents interact in simulations, games, and computer interfaces, but still struggle with human-like 3D spatial reasoning without extensive task-specific training.Approaches include LLM-planned subgoals paired with RL policies in Minecraft, self-explaining interactive plans, and browser agents that use screenshots and controls; LLMs can also reason over textualized driving scenes (Mao et al., 2023).Source: §3.4.2 · pp. 24–26
• On Workarena++, humans score close to 100%, while GPT-4o and Llama3 score close to 0% on tasks that remain challenging for agents.Tasks can require interpreting a chart and interacting with an ordering page; specialized agent-friendly interfaces can improve success rates in software engineering (Yang et al., 2024).Source: Figure 13 · p. 26, §3.4.2 · pp. 26–27
• Tool use is treated as an agent-system component rather than extended environmental interaction, though reinforcement learning may suit complex tools lacking training data.Agents may use documentation, examples, and immediate feedback to learn tools; specialized agent-friendly interfaces can also improve success rates (Yang et al., 2024).Source: §3.4.2 · p. 27

3.4.3 Interacting with Humans

• Human-agent interaction spans research, assistance, and domain-specific applications, with some systems seeking clarifications and conversational strategies still under study.AI scientists have generated hundreds of medium-quality papers within a week, while back-channeling increased conversational engagement.Source: §3.4.3 · pp. 27–28

3.4.4 Interacting with Other Agents

• Agents can simulate systems, serve as analytical components, or collaborate autonomously; multi-agent interaction can improve reasoning but also reduce diversity or spread manipulated knowledge.Independent collaborators may benefit from critics, precise prompts, and hierarchical coordination with leader feedback.Source: §3.4.4 · p. 28

3.5.1 Specifying an Agent

• Agent definitions can specify identity, motivations, roles, permissions, delegation, workflows, and interactions; simple role-goal-constraint descriptions may suffice, while partial autonomy can require human consent checkpoints.The appropriate definition depends on the intended level of autonomy and may reflect technical or legal constraints.Source: §3.5.1 · pp. 28–29
• Agent design may be manual or automatic, using search-based optimization, reflective code-generating meta-agents, evolutionary algorithms, few-shot fitting, trait sampling, or LLM-expanded profiles.Automatic design requires a search space, search algorithm, and evaluation function.Source: §3.5.1 · p. 29

3.5.2 Specifying Multi-Agents

• Multi-agent systems can specify agents and tasks independently, describing agent goals, roles, tools, delegation, and task outputs and procedures, including agent interactions.This separation supports flexible assignment of agents to tasks, with textual descriptions used for both.Source: §3.5.2 · pp. 29–30
• Multi-agent design dimensions include agent homogeneity, cooperation, communication direction and centralization, flat or hierarchical organization, learning arrangements, and communication content.Communication may be push or pull and centralized or decentralized; learning may occur independently or through shared experience.Source: §3.5.2 · pp. 30–31

3.5.3 Evaluating Agentic AI

• Agent evaluation is difficult because behavior is nondeterministic, interactions complex, tasks longer, and capabilities broad; evaluation spans skills, desiderata, domain checks, metrics, datasets, frameworks, games, and web tasks.Capabilities include planning, learning, error handling, tool use, and spatial reasoning; measures include completion and success rates, solution quality, failure awareness, and tool-call accuracy.Source: §3.5.3 · pp. 31–32

4.1 Challenges

• Reasoning can reduce errors relative to GenAI, but complex, lengthy tasks risk cumulative errors; readable Chain-of-Thought may aid interpretability without faithfully explaining decisions.GenAI and deep learning broadly continue to struggle with explainability, and agent systems add complexity through planning, memory, and tool use (Schneider, 2024a; Longo et al., 2024).Source: §4.1 · p. 32
• Agents operate in changing environments, select tools, and collaborate dynamically, while partial observability and opaque APIs can limit visibility into environments and agent operations.Agents may collaborate in parallel, and their behaviors and tool operations may not be fully transparent.Source: §4.1 · p. 32
• Agentic AI faces heightened safety and privacy risks from attacks, tool invocation, multimodal vulnerabilities, and sharing information with other agents and tools.Even seemingly minor information, such as an agent’s role as a psychological counselor, can represent a privacy breach.Source: §4.1 · p. 32
• Agentic AI evaluation remains incomplete, while agents may act against human intent and behave unpredictably, making the consequences of human or independent approvals difficult to anticipate.Evaluation includes reliability, task performance, and potential harms (Shavit et al., 2023).Source: §4.1 · pp. 32–33
• Agentic AI makes resource management more demanding because agents dynamically consume varying and potentially arbitrary amounts of information (Cheng et al., 2024).Source: §4.1 · p. 33

4.2 Agentic AI to AGI?

• AGI generalizes beyond narrow domains to unfamiliar situations; benchmarks emphasize novel tasks learned from few examples with limited computation, while current agents remain constrained to relatively simple tasks.Agentic AI enables applications beyond GenAI and makes increasingly complex reasoning and interaction tasks more feasible, but technological and regulatory constraints currently limit agents.Source: §4.2 · p. 33
• The path and timeline to AGI are uncertain: scaling may suffice, but limited training data or the need for fundamentally new approaches could stall progress.Predictions cited range from Ray Kurzweil’s 50% chance of human-level AI by 2029 to estimates of 50–70 years, 50–100 years, and approximately 180 years (Ford, 2018; Mok, 2025; Bengio and Hu, 2023).Source: §4.2 · p. 33
• If AGI surpasses human intelligence, outcomes are difficult to predict, and existential risks include loss of control and value alignment as systems pursue self-preservation or resources.Such risks are distinguished from economic risks such as unemployment and wage erosion; improvement could accelerate beyond exponential rates.Source: §4.2 · p. 33

5.1 Methodology and Scope

• The survey takes a high-level capability perspective for broad interest, while addressing recent Agentic AI innovations for technically focused readers and assuming basic familiarity with deep learning and GenAI.It does not focus on in-depth transformer knowledge and directs readers to other works for more extensive coverage of specific Agentic AI aspects (Schneider et al. (2024); Schneider (2024b)).Source: §5.1 · pp. 33–34
• The review follows Wohlin (2014), beginning with a meta-survey and then using backward, forward, and targeted searches to refine the survey’s structure and content.The authors assessed more than 30 surveys, selected five on Agentic AI and two on GenAI, focused later searches on works from 2024 onward, and quality-checked recent preprints and selected reputable blog posts.Source: §5.1 · p. 34

5.2 Related Surveys

• The survey distinguishes itself by directly comparing GenAI and Agentic AI, clarifying agent definitions, and expanding discussion of autonomy, motivations, practical agent and memory aspects, and conceptual frameworks.The authors note that the five related surveys did not directly distinguish GenAI from Agentic AI or provide a clear, multi-angled definition of Agentic AI; their second part also expands on practical aspects and offers a different conceptual framework.Source: §5.2 · pp. 34–35
• Wang et al. (2024) and Xi et al. (2025) use architecture- and brain/perception/action-focused framings, while this survey emphasizes capabilities, reasoning, interaction, and GenAI–Agentic AI differences.Acharya et al. (2025) compares Agentic AI with traditional AI; this survey instead starts from GenAI for readers familiar with systems such as ChatGPT, and its conceptualization differs substantially, with only one common aspect between the surveys’ Table 1s.Source: §5.2 · p. 35
• Plaat et al. (2025) and Cheng et al. (2024) differ in framing: this survey treats retrieval augmentation as tool interaction and emphasizes novel capabilities, rather than reasoning or traditional RL-agent concepts.It emphasizes two rather than three core capabilities, gives broader motivations for Agentic AI, and discusses agent definitions in greater depth than Plaat et al. (2025); Cheng et al. (2024) compares RL and LLM agents.Source: §5.2 · p. 35

• Agentic AI extends beyond Generative AI with reasoning, interaction, and autonomy at a new scale.Source: §6 · p. 35
• The survey systematically contrasts Agentic AI and GenAI while covering technical foundations, practical specification, and open challenges.Source: §6 · p. 35
• Understanding Agentic AI’s capabilities, risks, and agent-specification nuances is essential for advancing research and responsible deployment.Source: §6 · p. 35

Conflict of Interest

• The authors declare no conflict of interest.Source: Conflict of Interest · p. 36

Funding

• No funding was received for conducting the study.Source: Funding · p. 36

Data Availability Statement

• Data sharing is not applicable because no datasets were generated or analyzed, and all references are included in the paper.Source: Data Availability Statement · p. 36