Agentic AI builds upon Generative AI (GenAI). It constitutes the next major step in the evolution of AI with much stronger reasoning, learning, and interaction capabilities that enable more autonomous behavior to tackle complex, long-running tasks. Since the initial release of ChatGPT 3.5, Generative AI has seen widespread adoption, giving users firsthand experience. However, the distinction between Agentic AI and GenAI remains less well understood. To address this gap, our survey is structured in two parts. In the first part, we compare GenAI and Agentic AI using existing literature, discussing their key characteristics, how Agentic AI remedies limitations of GenAI, and the major steps in GenAI's evolution toward Agentic AI. We also distinguish between early Agentic AI aiming at reasoning and interacting and the increasingly emerging paradigm of self-improving Agentic AI that operates continuously and learns from its experiences by updating itself. This section is intended for a broad audience, including academics in both social sciences and engineering, as well as industry professionals. It provides the necessary insights to comprehend novel applications that are possible with Agentic AI but not with GenAI. In the second part, we deep dive into novel aspects of Agentic AI, including recent developments and practical concerns such as defining agents. Finally, we discuss several challenges that could serve as a future research agenda, while cautioning against risks that can emerge when exceeding human intelligence.
Figures & tables
Figure 1: Reasoning models perform extensive problem-dependent computations, commonly employing problem analysis, planning and reflection, while non-reasoning is shown by immediate responses without intermediate steps.
Figure 2: On the ARC challenge ( Chollet, 2019 ) , Agentic AI with growing reasoning effort involving planning and reflection dramatically outperform.
Figure 3: Overview: The high-level discussion of GenAI to Agentic AI is followed by an in-depth treatment of Agentic AI, followed by challenges and outlook
Figure 4: Agentic AI and GenAI performs similarly on MMLU ( Hendrycks et al., 2021b ) measuring a wide range of capabilities across disciplines(data from ( Paperswithcode, 2025 ) ). The benchmark is considered “solved”.
Figure 5: From Generative AI to Agentic AI. Early GenAI such as GPT-3 ( Brown et al., 2020 ) showed with minor effort basic reasoning and tool usage capabilities. These became much more profound over time, in particular, with Agentic AI, which includes elements from reinforcement learning such as interacting with the environment and other agents. Also the harness coordinator between memory, tools and the environment became more elaborate.
Aspect
Generative AI
Agentic AI
Reasoning
✗immediate responses
✓iterative planning and reflection
Interaction
✗(mostly with user)
✓user, tools, real-world, other AI agents
Execution capability
✗single-step tasks
✓Workflows, sequence of actions requiring diverse expertise
Adaptability
✗no self-improvement, bound to training data
✓Collecting and leveraging experiences
Autonomy
✗user-driven
✓self-directed, pursuing goals
Table 1: Comparison of Generative AI vs Agentic AI characteristics
Autonomy Level
Paradigm
Description
0 - No Autonomy
Classical Machine Learning
Tackles narrow tasks it was explicitly trained for.
1 - Assistive Autonomy
Generative AI
Handles simple tasks with direct instruction.
2 - Partial Autonomy
Agentic AI: Agent-Oriented Workflow
Manages multi-step tasks with human oversight and intervention.
3 - High Autonomy
Agentic AI: Goal-Oriented Collaboration
Achieves complex tasks with occasional guidance.
4 - Full Autonomy
Agentic AI: Autonomous Decision-Making
Given goals, handles all aspects of tasks independently.
Table 2: Autonomy levels for GenAI and Agentic AI based on ( Agent.ai, 2025 )
Figure 6: A small model with 1 billion parameters (1B) can outperform a larger 8B model in accuracy as shown for the MATH-500 benchmark using more computation, i.e., generations meaning calls to the model. Different “reasoning” strategies strongly impact the effectiveness of generations. Figure based on ( Beeching et al., 2024 ) .
Aspect
GenAI Limitation
Agentic AI Advantage
Execution capability
✗Failing complex multi-step tasks. Limited to generation of digital content with limited tool usage.
⚫Performs multiple steps to solve tasks using planning, interaction with an arbitrary environment and tools.
⚫Goals without detailed instructions are sufficient.
Training and few-shot data
✗Relies often on task-specific data, which may be infeasible or expensive.
⚫Leverages logic and external tools to operate with less data.
Learning & Memory
✗Limited by small context-windows
✓Trial and error learning through simulation and real-world interactions; Larger context memory with external storage paired with ability to change own instructions enable learning from past experiences
Errors
⚫Errors such as hallucinations are common
⚫Less errors due to extensive reasoning and validation at test time
Transparency & interpretability
✗Hard to interpret
⚫Shows intermediate reasoning, making outcomes explainable, though still hard to interpret.
Table 3: Limitations of GenAI (partially) overcome by Agentic AI; ⚫ indicates areas where further improvements are needed
Figure 7: A few milestones from Generative to Agentic AI. Self-improving AI agents are increasingly emerging.
Figure 8: Chain-of-thought(CoT) detailing steps to solve a task in the input elicit a CoT in the response thanks to in-context learning (Figure adjusted from ( Wei et al., 2022 ) )
Figure 9: Evolution of reasoning from direct input output prompting, to Chain-of-thought to a forest and graph of thought (Figure enhanced with GoT from Bi et al. (2024) )
Figure 10: Architecture for CoT combined with search with a separate controller containing also Scoring and validation.
Figure 11: Retrieval augmented generation (RAG) ( Lewis et al., 2020 ) enhances prompts with external information. It requires a vector database where vectors summarize text chunks and serve as index for retrieval.
Characteristic
Memory functions
Reasoning, personalization, processing large data and learning
Memory Persistence
Short-term (Working memory, model context)
Long-term (parameters, databases)
Memory Type
Architectural, retrieval-based, parametric, and ephemeral memory
Memory Location
Internal (parameters, context window)
External (databases, APIs, accumulation of inputs/outputs), and hybrid
External, non-agent knowledge (general world knowledge)
Memory access
Non-retrieval (direct access of relevant information)
Retrieval-based
Table 4: Characteristics of Memory and Stored Information
Figure 12: Instruction to (Anthropic’s) agent controlling a computer
Figure 13: (Anthropic’s) agent controlling a computer through mouse, keyboard and sensing through screenshots as illustrated in the left panel
Figure 14: Task from Workarena++ ( Boisvert et al., 2024 ) . Given a ticket asking to refill inventory below four items (top panel), the system must understand a chart to identify items to order(left lower panel) and interact with an ordering page (right lower panel).
Figure 15: Specifying a multi agents system using agents, tasks and their interconnection as abstractions specified by a user (Figure from ( Moura, 2024 ) )
Figure 16: Example definition of a content planner agent, the content planning task and the writer agent providing procedural guidance including interaction with other agents (Figure from ( Moura, 2024 ) )