Paper 1: The Rise of Agentic AI: A Review of Definitions, Frameworks, and Challenges
Artificial Intelligience is rapidly evolving and one of the emerging concepts of modern AI research is agentic AI. Unlike traditional AI systems which only responds to user prompt, agentic systems are designed to act more independently by planning tasks, using tools and interacting with environments.
In this blog I will discuss insights from the research paper "Rise of Agentic AI" published in 2025 and "A Survey of LLM-based Deep Search Agents" published in 2026
What is the Paper's role ?
The main goal of the paper is to clarify the meaning of agentic AI as it is a relatively new concept and different researchers define it in different ways. These varying definations can create confusion, therefore, the paper reviews the existing defination and frameworks to explain what agentic AI really means.
Agentic AI refers to systems that behave like intelligent agents. Unlike standard large language models that mainly answer questions, agentic systems can plan tasks, break down complex problems into smaller steps, and interact with tools or environments to accomplish goals.
The paper also explains frameworks used to build such systems. These frameworks typically combine several components such as large language models, memory systems, planning mechanisms, reasoning capabilities, and external tools. For example, an agentic system might divide a large task into smaller subtasks and complete them step by step.
However, the paper also highlights several challenges associated with agentic AI. These include reliability issues, safety concerns, and hallucinations, where the AI generates incorrect information. Because these systems may act independently, even small mistakes in reasoning could lead to significant problems.
Connection to Concepts from Our AI Course
Several ideas from the paper connect to concepts we studied in our AI course. We learned about different types of agents such as simple reflex agents and utility-based agents. We also studied search algorithms like A* Search Algorithm, Breadth-First Search, and Depth-First Search.
In these algorithms, the system searches through a state space to reach a goal. For example, in A* Search Algorithm, a heuristic function helps guide the search toward the goal.
Agentic AI systems extend this idea by allowing the large language model to participate in the reasoning process. Instead of relying only on predefined heuristics or fixed algorithms, the LLM may dynamically generate reasoning steps, decide which tools to use, or determine the best strategy for solving a problem.
While traditional algorithms like Breadth-First Search and Depth-First Search follow strict deterministic rules, agentic systems are more flexible because their reasoning is generated dynamically. This flexibility can make them more powerful for complex tasks, but it also makes them less predictable.
Insights from Manual Reading and NotebookLM
While reading the paper manually, some sections were difficult to understand at first because the paper used several technical terms such as frameworks and hallucinations. It also sometimes felt like similar ideas were explained repeatedly using different wording. To better understand the paper, I explored it using NotebookLM, which helped summarize sections and explain difficult concepts.
One important thing i learned from manual reading was realizing that Agentic AI is not just a large language model. Instead, it is a combination of multiple components such as memory, reasoning, tool usage, and interaction with environments. The intelligence of the system comes from the integration of these components rather than the LLM alone.
The NotebookLM exploration also revealed that researchers are increasingly combining traditional AI concepts with modern LLM-based approaches. Many agentic systems still rely on ideas like search and planning, but the LLM now helps generate reasoning steps dynamically instead of relying solely on fixed algorithms.
**
Paper 2: A Survey of LLM-based Deep Search Agents
**
What is the Paper's role ?
It focuses on how large language models can be combined with search techniques to create deep search agents.
Traditional search systems usually retrieve information based on queries. However, deep search agents go further by performing multiple rounds of search and reasoning. For example, the system may generate a query, retrieve information, analyze the results, and then generate new queries based on the findings.
The paper surveys many research systems that combine LLMs with retrieval tools, planning modules, and reasoning processes. These systems aim to improve the ability of AI models to solve complex information-based tasks.
However, the paper also discusses challenges such as high computational cost, hallucination issues, and difficulties in evaluating the reasoning process of these systems.
Connection to Our Course Learning
The idea of deep search agents can also be related to search algorithms such as A* Search Algorithm, Breadth-First Search, and Depth-First Search.
In such search algorithms, the system explores nodes in a state space until it reaches a goal. In deep search agents, the exploration happens across information sources such as documents, databases, or web pages rather than simple graph nodes.
The LLM may act similarly to a heuristic by deciding which query to generate next or which information source should be explored further. This creates a search process that is guided by reasoning generated by the language model rather than fixed algorithmic rules.
Manual Reading and NotebookLM Exploration
While manually reading I realized that deep search agents do not simply retrieve information once. Instead, they continuously search, analyze results, and refine their queries, which creates a multi-step reasoning process.
NotebookLM also helped reveal that these systems behave somewhat like iterative search algorithms, where the model keeps exploring new information until it finds a satisfactory answer. This made it easier to relate the concept to classical search techniques we learned in class.
Conclusion
Both papers highlight an important trend in modern AI research. Agentic AI focuses on building independent systems that can plan tasks and interact with tools, while deep search agents focus on improving how AI systems explore and analyze information.
This blog was written as part of an AI research assignment. @raqeeb_26
Video
Top comments (0)