Rise of Agentic AI

This paper that i just read provides a detailed review of agentic Al , a paradigm shift moving AL from the typical reactive tool to autonomous ,goal-driven systems. Unlike the typical generative AI that simply responds to prompts, agentic AL systems have the ability to understand the objective they perform the task by breaking the the main goal to be achieved int to sub-goals and then choose tools accordingly and finally execute the multi-step plans with minimal human supervision. According to the research 143 studies have analyzed the core architecture and application domains ranging from the healthcare to the military security .It concludes that while agentic Al are highly efficient yet significant challenges still remain specially in terms of reliability, ethical alignment and multi-agent coordination.

Agentic AI

Agentic AI vs Al Agents : An AI agent is a single autonomous software that perceives the environment and acts accordingly . Agentic AI refers to the boarder multi-agent system where specialized agents collaborate and then together plan to achieve goal that are high-level , these goals include travel planner coordinating flights, hotels and itineraries autonomously.

The BDI Architecture: This models an agent's mental state using three basic components: Beliefs(Information about the world ),Desires(all possible goals),Intentions(The specific goals that the agent commits to pursuing). It is highly valued for tasks requires explainability and traceability, though it can be brittle in unpredictable environments.

layered Neuro-Symbolic Architecture: This includes the combination of two different AI strengths : Neural perception(for handling uncertain data like images or text) and symbolic planning (for logic-based ,traceable decision-making ).This "best of both worlds" approach handles real-world uncertainty while keeping the AI's reasoning transparent.

Memory Systems : The sources distinguished between Short-term memory(STM),which maintains the context of current task and Long-Term Memory(LTM)which stories from the past and to inform future decisions it also stores users preferences.

Algorithm (Wang et al .,2025)*

GOAL: The primary goal is to address the traditional algorithms limitations that includes long computation times ,excessive search nodes and paths dangerously graze obstacle corners.

Course Connection (Heuristic Search): This paper provides a to the point application of modifying logic studies in course. It introduces three critical optimizations:

Diagonal-free five-way search: This includes modifying the expansion logic to prevent collisions.

Adaptive Dynamic Weighting: It helps by replacing the static heuristic weight with radial basic function, allowing the agent to search faster when far from the goal and more precisely as it gets closer.

Heuristic Reward Values: It adds a "rewards" to the cost function f(n) to help the agent escape "local traps" (local optima ) where a traditional A* might stop.

Personal Insights :

Before going thorough these research papers I believed that the Agentic Ai was advanced form of AI Agents. But I realized that the difference lies in coordination and autonomy. Traditional AI agents follow predefined logic, but agentic systems dynamically plan, delegate subtasks, and select tools without constant human input. This was very helpful for me to understand what "Intelligence " meant in AI.

One concept that I found particularly interesting was the integration of memory systems. In our course, we usually study search algorithms like A* where the algorithm does not remember past sessions. But agentic AI systems maintain both short-term and long-term memory, making them more human-like in decision-making. This made me realize how important memory is for building truly autonomous agents.

Notebook LLM:

I did knew about Notebook LLM before. Since I have discovered this platform I can say this platform is for learners. All the features that this platform provides for free, I do not think any other platform provides . This is a platform where you can learn with no restrictions. Learners can go deep into the topic that they want to explore.

Using Notebook LLM helped me a lot in clarify the layered neuro-symbolic architecture. At first, I was confused about how neural and symbolic systems could work together. But after exploring examples, I finally understood that neural networks handle perception (like interpreting images or text), while symbolic planning ensures logical reasoning and explainability. This combination connects directly to our discussion of model-based and utility-based agents in class.

Another insight I gained was regarding adaptive heuristic modification in the A* optimization paper . In class, we studied A* with a fixed heuristic weight. This research showed how dynamic weighting can improve efficiency and reduce unnecessary node expansion. It helped me see that real-world AI systems continuously refine classical algorithms instead of replacing them completely.

Overall, this paper changed my perspective from seeing AI as isolated algorithms to understanding it as a coordinated ecosystem of agents, memory systems, planning strategies, and adaptive heuristics working together.