Intro To Agentic AI
Agentic AI refers to autonomous goal-driven systems that collaborate with other AI agents and use this multi-agent approach to accomplish goals. They are not like the simple AI Agent which requires continous Human Handling. Agentic AI breaks the main goal into sub-tasks and then assigns the tasks to the sub–AI models present within it, and then collaboratively generates the result for the user in the end. They do not work in loops where continuous human prompting is needed to generate the next output. These systems are smart enough to understand what relatable information and work will be required to reach the goal efficiently, and do the work on that part, along with the solution generation.
Core Technical Components:
Perception and World Modelling: This part takes in external sources and works on them to conclude structures or internal representations for repetitive goal generations
Memory: It uses Short Term Memory STM and Long Term Memory LTM in order to accomplish consistency in sequential tasks and Episodic Tasks
Planning, Reasoning, and Goal Decomposition: This Module breaks the entire main goal into stepable actions where actions are performed gradually, and alternatives are planned to achieve the goal.
Communication and Orchestration: This layer coordinates the task flow and planning between the other AI agents and also integrates the human oversight and guidance when needed on key points.
Why Agentic AI? - Technical Point Of View
Agentic AI shifts the trend to proactive agents over reactive agents, making them more time efficient less human-reliant. Moreover, it provides end-to-end automation, which helps automate complex processes which are usually time-consuming or require human intervention, such as scientific research. Agentic AI also feels the need for time when the industry focuses on operational automation rather than just content generation. Agentic AI is being deployed in various fields such as healthcare systems and financial risk management where reasoning and strong decision-making are required.
Adaptive A* Algorithm
Autonomous driving technology is a strong real-world example of Agentic AI. Self-driving vehicles perceive their environment, build internal models, and plan routes without continuous human input. A key component of such intelligent systems is path planning, the ability to compute safe and efficient routes toward a goal that ensures human safety and legal road performance. Classical search algorithms, especially the A* algorithm, form the core foundation of path planning in these modern autonomous systems.
While searching for the optimal path, sometimes complex paths’ obstacles trigger the explosion of search nodes. In order to demonstrate that the search paper discusses a series of improvements, such as adaptive dynamic weighting coefficients and heuristic reward values. These help escape node explosion by reducing nodes, and also improve search efficiency.
In the adaptive weight improving method, there was an addition of a function in the formula.
f(n)=g(n)+f(x,y)∗h(n)
f(x,y) is the adaptive dynamic weight function
This function modifies nodes closer to the starting point to have a higher dynamic weight. Afterwards, the weights decrease gradually towards 1 when the search is taking place farther from the start point.
Why Better? - Technical Point of View
It reduces the computational load as mentioned in the research paper, by reducing the number of expanded nodes by 40.4%. Moreover, it is more efficient in static environments when it is combined with other improvements, like heuristic reward values, which reduces the total number of search nodes by an average of 76.4%, as per the research paper. However, one trade-off is that increasing the adaptive weight may sometimes produce a slightly longer path compared to classical A*, especially when the heuristic becomes more aggressive. This means that while node expansion is reduced, the strict optimality of the solution can be slightly affected in certain cases.
Course Connection
In my AI Course, I studied different types of search algorithms, including uninformed search methods. However, the informed search method, A*, was considered better as it combines path cost and heuristic estimation for expansion. Exploring this research paper, I have realised that recent improvements have been made to this heuristic formula by adding adaptive weighting functions, which enhance efficiency and reduce unnecessary expansion of nodes.
Similarly, I was already familiar with different types of agents, such as simple reflex agents and utility-based agents. However, I did not initially understand how these concepts connect to modern generative AI systems. After studying Agentic AI, I recognised that many generative systems now operate using advanced versions of these agent models, integrating planning, memory, and reasoning into more autonomous frameworks.
Personal Reflection
I initially viewed A* purely as a mathematical formula, but after studying adaptive weighting, I realised how small heuristic modifications can significantly impact computational efficiency in real systems. NotebookLM made it easier for me to understand and differentiate between concepts. The major limitation was to relate both research papers and draw mutual concepts. NotebookLM helped me understand both topics with interconnected examples, conceptual explanations and also with reference to my course content when I provided it in prompts.
Submitted as part of AI coursework under @raqeeb_26
Top comments (1)
We need to link our academic studies with the latest research so we can realize how these things work in practical life. Well explained @abdullah-mangrio. One more thing whenever you are sharing your review like this you should start with how you think not starting with definition. Make it sound personal next time and then move towards technical stuff...