What are AI Agents and How Do They Work? Explained

The artificial intelligence landscape is evolving rapidly, and we are now moving past usual chatbots. After the introduction of ChatGPT in late 2022, which is powered by large language models (LLMs), the focus is now shifting towards action-driven AI agents. While AI chatbots like ChatGPT and Google’s Gemini can process text and visual information and respond in natural language, AI agents can perform complex actions. Thus, let’s learn everything about AI agents in detail, along with how they work, types of AI agents, and more

What are AI Agents?

The term ‘AI Agent’ refers to an AI-powered software system that can plan, reason, make decisions, and perform multi-step actions to achieve goals autonomously. The purpose of an AI Agent is to complete tasks by interacting with external systems, whereas AI chatbots can process information and remain isolated in their own environment.

However, just like AI chatbots, AI agents are also powered by large language models (LLMs) under the hood. But LLMs are fine-tuned to make them action-driven. Currently, in the AI field, companies are using reinforcement learning and advanced reasoning on visual language models to develop AI agents. Companies are also integrating a lot of external tools including APIs, functions, databases, etc., so that AI agents can accomplish a range of tasks.

Hence, AI agents are not just an AI model, but represent an ‘AI system’ that supports tool calling, long/short-term memory, and can interact with third-party systems to complete a given task. For example, the Operator AI agent, launched by OpenAI, is a Computer-Using agent (CUA) which has been trained to interact with graphical user interfaces (GUIs) on the web.

Basically, the Operator AI agent can browse the web, order groceries for you, fill out forms, book flights, and more. It can do any action you want on the web. And it uses GPT-4o’s vision capability to analyze the screen, and to find out where to click next. However, it’s not fully autonomous yet. Often, it gets stuck in a loop, and sometimes, human supervision is required to complete a task.

And since AI agents are in initial phases, to complete a critical task like making payments, the controls are given back to the user. Put simply, after processing and generating information with AI chatbots, the next evolution of AI applications is going to come from action-driven AI agents.

Types of AI Agents

Stuart Russel and Peter Norvig, in their book ‘Artificial Intelligence: A Modern Approach,’ have classified AI agents into five broader types. It includes Simple Reflex agents, Model-based Reflex agents, Goal-based agents, Utility-based agents, and Learning agents.

Simple Reflex agent basically works on conditional logic. It’s the most basic form of AI agents: perform the action if a particular condition is true. It doesn’t care about past information and doesn’t have memory. During the process, it doesn’t learn or remember patterns.

On the other hand, Model-based Reflex agents maintain a memory and builds a basic understanding of the world by observing how the world evolves with their actions. For example, a robot vacuum cleaner receives new information about obstacles and updates its internal model to avoid them while cleaning. While it has memory, this kind of agent is limited by a set of rules.

Next, Goal-based agents are not bound by rules, but have to achieve a set of specific goals. It can plan and reason to find the best way to achieve a given task. The agent has to consider multiple factors before making a decision, ensuring that it will bring the agent closer to the final task. For instance, a chess-playing AI has to consider all possible moves to achieve a desirable outcome.

Now, Utility-based agents have to choose a sequence of actions that will maximize “happiness” or “satisfaction.” Basically, there is a reward function associated with these kinds of agents. Finally, Learning agents have the same capability as other AI agents, but they can also gain new knowledge from an unknown environment. They get better over time and learn new preferences the more you use these agents. We’ve explained all types of AI agents in our dedicated guide if you wish to learn about them in detail.

Examples of AI Agents

As mentioned above, OpenAI has released the first consumer AI agent called Operator (visit). It can go to the web in a cloud web browser and perform tasks for you. You can ask the Operator to order food for you, find hotels, book a concert ticket, and more. As the agent is in early research preview, it’s only available to users who have subscribed to ChatGPT Pro, which costs $200 per month.

Apart from Operator, OpenAI has launched the Deep Research AI agent that can dive deep into any topic you throw at it to generate a comprehensive report. It also adds citations so you can verify the information by clicking on the source link. You can also try Gemini’s Deep Research AI agent which does the same thing, and it’s available for free.

Anthropic has launched the Computer Use AI agent that can operate a computer by visually analyzing the screen. I have personally tested this AI agent in a Docker instance. While it’s slow to respond, it does work and operates the computer for you. Not to mention, Anthropic’s MCP standard is being adopted by Google, OpenAI, and Microsoft to connect AI agents with AI models.

Recently, Manus, a general AI agent from China went viral, and it can browse the web, run code, and interact with a cloud computer to accomplish tasks. While the demo was quite cool, it turned out that the agent was powered by Anthropic’s Claude 3.5 Sonnet model.

Lastly, on the consumer side, Google is working on Project Mariner that can perform tasks for you in the Chrome browser, just like OpenAI’s Operator. Google is currently testing the agent with trusted testers and will release it in the coming months.

Overall, I would say the agentic AI era is still a year or two away. We have not reached a stage where AI models can be fully trusted to perform critical tasks autonomously. AI companies are themselves adding human oversight as the default way to interact with AI agents. Nevertheless, the future is going to be action-driven, and major AI labs like OpenAI and Google DeepMind are working to turn the vision of agentic AI into reality.