Guide of AI Agent Types with examples
From a home alarm, to a fleet of robots in a warehouse, to your smartphone’s assistant, AI agents play a pivotal role in today’s technological advancements. In my previous article, I explained to you what are AI agents. In this article, I am going to introduce you the different types of AI agents, with many examples.
Introduction
An AI agent is something that perceives its environment through sensors and that can do action regarding rules, data and algorithms. A simple example is a thermostat that regulate the heating to reach a certain temperature depending on certain hour. It perceives its environment through temperature sensor and clock. It takes action with a switch that can turn on or off the heating depending on the actual temperature or the time. A thermostat can be turned into a more sophisticated AI agent by the addition of AI capability that allow it to learn from the habit of the people who live in the house.
Types of AI Agents Explained
There is a lot of AI agents and agent in general. Here, I’m proposing you a list that allow you to get a better view of this landscape. For every agent, I propose you a name, and you can find them into others names. This list is not exhaustive, because many variants could be created. Let’s start with the first most common agent.
Simple Reflex Agent
If you know the application IFTTT (If That, Then That), this is the same thing. This is an application that I used before to put in place certain automation effortlessly. This is the same as zapier.
A Simple Reflex Agent, as its name suggest, as the simplest agent possible. They are not mad for complex task. It works on the same principle that IFFT or zapier. It takes decision based on what they perceive of the environment and with a pre-defined list of rule and condition. They should be used in a scenario when the environment is stable, because they are not designed to be adaptable.
Two simple examples to make it more concrete:
- A basic spam email filter: it perceives when an email is received, and regarding rules such as keywords list and sender email, it can take the decision to directly move the email into the spam folder.
- A simple thermostat: as I wrote into the introduction, a simple thermostat can be assimilated to a simple reflex agent.
Model-based Reflex Agent
It is a simple reflex agent, but this time it has an internal “memory”. Beyond the rules, conditions and the perception of its environment, it can use the history of the previous events and actions to take a decision of what to do.
Examples:
- A chess player agent: Given of the actual state of the game, the previous moves of the current game and of the previous games, it tries to predict the best move possible with the most win rate probability.
- A stock trader bot: Given the actual state of an asset, the state of the global market, key indicators and the previous market moves through the time, it can take the decision to take a position, to cut loss or to take profit.
Goal-based Agent
A goal base agent, as the name suggest once again, is driven by a pre-defined goal. To reach its goal, it has access to a various set of tools and data. It also evaluates the potential future of its actions and the consequences to take an action.
Examples:
- Autonomous car: It starts at a point A and has to reach a point B by respecting all the highway code, and without producing any incident. For that, it has access to a GPS, to cameras, speed sensors or to a LiDAR (Light Detection And Ranging, a kind of more sophisticated sonar working with light). To reach the destination, it has to know the itinerary, to stop if someone want to cross the road or to not exceed the authorized speed, even if he needs to overtake another car.
- Automated message response system: It has to answer to a message in a short period of time by sending a relevant answer. For that, it could have access to various databases or search engine to gather the relevant data, and then to use it to provide a good quality answer.
Utility-based Agent
It evaluates the desirability of it goals’ outcomes. This evaluation aids it in choosing between multiple goals or actions based on the most beneficial outcome.
Examples:
- Route optimization system: Let’s take back the previous example of an autonomous car. It could have an agent dedicated to the prediction of the best itinerary according to roadworks, weather conditions, passenger choices (speed and eco-friendliness) and real-time traffic conditions.
- Smart thermostat: Let’s take back again the example of the basic thermostat. It could be turned into a smarter version with pre-defined scenarios (heat more during the day than the night, and not during certain periods of time), the weather, machine learning and presence sensor. It can decide to heat certain area, to heat more to compensate if the outside temperature drop down to keep the targeted temperature, or to anticipate the heating process regarding all the data gathered to reach the good temperature on time.
Learning Agent
It stands at the top of AI agent. This kind of agent enhances its performance over time thanks its experiences, and it can adapt to new situation without explicit programming.
Examples:
Spam Filter: A simple example of a Learning Agent is a spam filter that uses machine learning algorithms to detect and filter out unwanted emails. The agent’s goal is to learn from historical data and user feedback to improve its ability to classify emails as spam or not. If you use Outlook or Gmail, when you report an email as junk or spam, it will help the algorithm to better filter the future email, for you, but also for others users. This simple learning agent adapts to new data and user preferences to better achieve its goal of filtering out spam emails.
Customer retention: When you use an application that rewards you when you go to your favorite supermarket, it will learn your behavior and remains all your previous purchases. The goal of this agent will be to provide you at the right time a voucher or email you with products you may want to buy, to ensure you continue to spend money on this supermarket and to increase the income.
Hierarchical Agent
It is more a way of making agent work together than a kind of agent. A hierarchical agent reach its goal by overseeing lower-level agents that each manage a part of the main goal or a subgoal. It decomposes a task into others more little, in a hierarchical way, in order to handle complex decision-making processes and to have a good adaptability in dynamic environments.
Applications in Robotics and Autonomous Systems: It is useful in robotic domain and autonomous systems, where complex task planning is needed, by execution through the decomposition of goals into sub-goals and tasks. This kind of agent can navigate to a specific location by avoiding obstacles, and performing high-level tasks like object manipulation thanks to lower-level motion control tasks. The caveat is when fast execution and real time is needed. Agent need to have a low response time or another system is needed.
Conclusion
What is even more powerful is when you put multiple agent together and when you mix them to create a crew of agent capable of much more than a single alone gent. In this case, the coordination is key. You can either put multiple agent together to reach a goal, or you can make hybrid agent as well. As you see through the examples I provided you, AI agent is not something new, they are everywhere.
In this article, I introduced what are AI agent. We are talking more and more about AI agent for a good reason. With the recent fast improvement of the Large Language Model, it becomes easier for everybody to create their own agents and their own crew of agents. It enables you to do tasks in minutes where it could have been taken you hours or days!
From simple reflex agents maintaining home temperatures to more advanced agents driving a car, AI agents are already everywhere. Understanding these agents’ capabilities and limitations allows us to leverage their potential to address complex challenges, and enhance our daily experiences.
Thanks for reading. If you liked this article or if you want to encourage me to write more, feel free to give me some 👏
