Hey Dev.to community! I am Syed Ammad Hussain, a BS Computer Science
student from FAST University Peshawar, Pakistan. This is my
first post here. I am sharing a research paper review I did
for my university AI course. Would love any feedback!
Tagging my course TA @raqeeb_26 who assigned this!
Why These Two Papers?
Our AI course covers agents, search algorithms, CSPs, and planning. When I looked at the paper options, two of them jumped out immediately because they seemed directly connected to things we were already studying in class. I chose:
Paper A: The Rise of Agentic AI: A Review of Definitions, Frameworks, and Challenges (2025)
Paper B: A Survey of LLM-based Deep Search Agents (2026)
I read both papers manually first — no AI assistance — before uploading them to Google NotebookLM for deeper exploration. The difference between what I understood on my own versus what NotebookLM helped me uncover was honestly eye-opening.
Paper A — The Rise of Agentic AI (2025)
What is the paper's goal?
This paper does something important — it tries to clearly define what "Agentic AI" actually means, because the term was being used loosely everywhere. The researchers reviewed over 140 studies and built a proper framework around this concept.
The core idea is simple but powerful. Traditional AI systems just respond to inputs. You ask a question, it answers, done. It has no memory of what happened before, no ability to plan future steps, and no way to recover if something goes wrong.
Agentic AI is completely different. An agentic system:
Plans a sequence of steps to reach a goal
Remembers previous actions and context
Uses tools like search engines, code interpreters, and APIs
Coordinates with other agents to complete complex tasks
Recovers from errors without human intervention
What makes this paper stand out is the data it presents on how fast this field grew. Interest in Agentic AI was nearly zero before April 2024. Then it spiked suddenly and reached its peak in July 2025. This is a field that practically did not exist two years ago.
How does it connect to our course?
In our AI course, Dr. Bilal taught us the classic agent taxonomy — Simple Reflex Agents, Model-Based Agents, Goal-Based Agents, and Utility-Based Agents. At the time, I studied these for the exam and moved on.
Reading this paper made me realize those are not just textbook categories. They are the exact foundation of what Agentic AI builds on. A Goal-Based Agent that can plan, combined with memory and tool-use, running on top of a Large Language Model — that is literally what an Agentic AI system is. The theory maps directly to the real thing.
Paper B — A Survey of LLM-based Deep Search Agents (2026)
What is the paper's goal?
This paper provides the first systematic survey of how search has evolved with the rise of Large Language Models. The central argument is that search is no longer a passive, one-time lookup — it has become an active, multi-step, intelligent process.
The paper identifies three generations of search:
Generation 1 — Traditional Search You type keywords. You get a list of links. You open each one, read manually, and piece together the answer yourself. All the work is done by the human.
Generation 2 — LLM-Enhanced Search An LLM reads the search results and summarizes them for you. But this is still a single-turn process. You ask once, it answers once. There is no dynamic planning or follow-up.
Generation 3 — Search Agents This is where it gets genuinely impressive. A Search Agent takes your question, creates a plan, breaks it into sub-questions, searches multiple sources simultaneously, reads and synthesizes everything, and produces a comprehensive report — all autonomously. The paper describes this as multi-turn retrieval with dynamic planning.
The real-world example they highlight is OpenAI's Deep Research tool, where you give it a complex research question and it returns a detailed, cited report after several minutes of autonomous searching.
The paper also breaks down the agent pipeline into four core stages:
Planning — understand the question and design a search strategy
Developing Questions — break the main query into focused sub-questions
Web Exploration — autonomously search and read multiple sources
Report Generation — synthesize all findings into a coherent answer
How does it connect to our course?
In class we study BFS, DFS, UCS, and A* search. These algorithms follow fixed rules. A* uses a heuristic to guide search, but that heuristic is defined before the search starts. It never changes while the algorithm is running.
LLM Search Agents break this model completely. They evaluate what they find mid-search and dynamically adjust their strategy. If a sub-question is not returning useful results, the agent shifts its approach. It is like A* being able to rewrite its own heuristic function in real time based on what it discovers along the way.
My Personal Insight — Manual Reading vs NotebookLM
I want to be honest about what I actually experienced here.
When I read manually first, I understood the surface-level ideas — Agentic AI is more powerful than traditional AI, search agents are smarter than keyword search. But I missed a lot of the nuance. For example, in Paper A, I initially confused "agentic" with just "autonomous." I thought any automated system counted. It was only after reading more carefully that I understood the key distinction — an agentic system has persistent goals and self-directed decision making, not just automation.
When I uploaded both papers to Google NotebookLM, it helped me ask better questions. I asked things like "How does this paper define the boundary between a model-based agent and an agentic AI system?" and the answers helped me see distinctions I had glossed over. NotebookLM also helped me connect concepts across both papers — for example, how the planning stage in Paper B's search pipeline directly mirrors the goal formulation step in Paper A's agentic framework.
The combination of both approaches — manual reading for genuine understanding, NotebookLM for deeper exploration — gave me a much richer picture than either alone.
The most interesting thing I discovered overall: both papers, written independently, point to the same fundamental shift. AI is moving from systems that follow rules to systems that pursue goals. And everything we are studying in this course — agent types, search algorithms, planning — is the exact theoretical foundation that makes this possible.
Video Explanation
I recorded a short 3-minute video explaining the core ideas of both papers and what I found most interesting. Check it out below:
🎥 Watch my video here
Thanks for reading! I am Syed Ammad Hussain, BS-CS student at FAST University Peshawar. Feel free to connect and share your thoughts in the comments.
Top comments (0)