AI agent is an AI system that perceives the environment and uses the resources (such as data, computational power, algorithms, models such as machine learning and deep learning, sensors, and APIs) at its disposal to act for decision making and to accomplish goals with the lowest possible level of human supervision.
These systems are considered to be autonomous/semi-autonomous due to their ability to learn and take actions with minimal human intervention. They combine Perception, Reasoning, Decision Making and Actions to ensure that the AI agent learns, adapts and optimizes its performance over time.
The concept of intelligent agents has evolved through time. Early research started off with rule based and expert AI systems in 1950s and 1970s pioneered by John McCarthy and Alan Turing. AI Agents today can interact with their environment, solve problems, and perform tasks that extend beyond just content generation that Generative AI models are meant to do. Although the foundation of both agentic and generative technologies comes from foundations models (FMs) or LLMs, generative AI generates new material by using the patterns it discovered in the training data. In contrast, agentic systems can act on the information they receive from their surroundings in addition to producing content, which can include performing tasks that have intermediate output that becomes input to other such agents, collectively which as a system generate the final end-user output. Because of their goal oriented and reasoning behaviour these systems can be used in real world scenarios, ranging from autonomous completion of a single task to managing complex system of systems, such as process automation or even self-driving cars.
Characteristics of Agentic AI
.
Figure 1: Word cloud of the characteristics of Agentic AI
During 1950s Simple Reflex and Reactive Agents emerged as some of the earliest rule based systems, capable of operating without human supervision. The gradual evolution of these systems into machine learning and deep learning-based models has led us to today's advanced and complex AI-driven automation and intelligent assistants. By 2020s we have advanced to Learning Agents which can continuously improve their decision making through experience, reducing or even eliminating human intervention.
AI agents can be understood through multiple ways of categorization. This article bifurcates the current AI Landscape into three primary categorization.
- Objective Based Agents
- Model-Based Agents Using present perception and memory, it constructs an internal model of the physical world which gets updated as new information becomes available. The model further identifies rules that align with the present circumstances and new information to chart a direction that may differ from past rules. For example, self-driving cars continuously monitor real-time sensor data to revise existing internal map of their model environment to forecast the movements of other vehicles and pedestrians in the present and near future state.
- Goal Based Agents make judgements depending on how far they are from achieving the pre-determined goals set by the user. With the help of search algorithms, they choose the optimal next move since every move will bring them closer to the goal which gives them the ability to choose from a variety of options. They are appropriate for tasks like NLP and robotics
- Utility Based Agents evaluate multiple possibilities and select the action that maximizes the desired outcome. This enhances their flexibility and ability to balance trade-offs between different outcomes. For an instance a stock agent can analyse the data, market, stock price, trends etc. and can evaluate multiple investment options. Based on which it can choose the trade that can minimize the risk while maximising the profit.
2. Learning Agent
- Complete Human Supervision Agent are the agents which learns but with constant human intervention during their learning process, they require regular labelled data, correction and reinforcement from human experts. Example include Chat GPT which regularly takes human input for its learning process.
- Learning Agents have the capabilities to learn from the past itself, by employing techniques like Reinforcement learning they can adjust to changing circumstances to enhance their performance. Example include self-learning agent like Deep seek which learns from Self Supervised data and do not rely on human supervised data.
3. Function Based Agents
- Single Agent System are those where only one agent performs the task independently on its own. Decision making process occurs independently in this system.
- Sequential Agent System is a system where a complex task is broken down into a series of ordered steps where each action influences the next. Here the system considers the long-term consequences of its decision rather than acting upon the current situation
- Multi Agent System are said to be those where group of different agents work together and interact with each other in a shared environment to perform a certain task. This system is able to work in parallel as well as sequential manner.
- Hierarchical Agent System is a system where different AI Agents are organized in layers which consists of High-level, Mid-Level and Low-Level agents. Here higher levels agents break down a complex task into sub tasks and assign them to lower-level agents. Here the Agents works parallel to each other in each Level.
AI technologies which forms an Agentic AI system:
Figure 2: Core AI technologies that Make an Agentic AI System Agentic AI
Merit and Demerits of Agentic AI
|
Merits of the Feature
|
Evolving Features of Agentic AI Systems
|
Demerits of the Feature
|
|
The current level of autonomy improves efficiency by automating complex workflows, facilitating quicker decision making and reducing human workloads
|
Level of Autonomy
|
Autonomous AI Agents systems raise risks of error and unintended actions due to limited human oversight.
|
|
This parameter of agents enables continuous act of self-improvement with the help of Reinforcement Learning, allowing it to optimize on its own.
|
Learnability
|
Unchecked learning can lead to model drift in which AI Agents may acquire unwanted biases that deviates it from its goal.
|
|
Because of AI Agent’s current scalability, it is deployed among various industries, handling enormous volume of dataset and automating complex tasks.
|
Scalability
|
Scaling agents leads to the demand of complex infrastructure, energy usage and computing expenses, making resource allocation major issue.
|
|
Nowadays, its high level of personalization add up to enhanced user experience and tailored responses for recommendation and better decision making based on individual’s preferences.
|
Level of Personalization
|
Excessive personalization leads to privacy issues, data security concerns and potential biases by excessively tailoring the user’s behaviour
|
|
The current level of human supervision allows the AI Agent to mitigate risks by providing oversight, which is an important element since at every level of autonomy some human intervention is required.
|
Level of Human Supervision
|
A decreased level of human supervision increases the risk of unpredictable, unchecked and unethical responses, leading to dangerous real-world consequences.
|
Table 1: Merits and Demerits of core evolving features of Agentic AI
We are still progressing toward creating a truly autonomous Agentic AI systems capable of real-time learning, independent and complex decision-making, and seamless adaptation to unpredictable environments. While technologies like large language models, reinforcement learning, and multi-agent systems have advanced AI’s capabilities, they still require structured training, human intervention, and controlled scenarios to function effectively. Challenges such as continuous learning, ethical considerations, system reliability, and real-world unpredictability remain as significant challenges to be addressed.
Reaching the stage of fully autonomous AI will demand breakthroughs in general intelligence, self-evolving algorithms, and most importantly, trustworthy AI frameworks making it a future goal rather than an immediate reality.
