Agentic AI
Agentic AI is essentially giving AI a specific job that it will execute on a regular basis and giving it the autonomy to do so.
There are five main categories of agentic AI today:
- Simple Reflex
- Model-Based Reflex
- Goal-Based
- Utility-Based
- Learning
Simple Reflex Agents
The most basic agentic AI models are reactive and are built to analyze an input from the environment then use condition based logic/rules to make a decision on what to do next. For instance, your thermostat has a sensor that evaluates the temperature of the air (environmental input) and checks it against the conditional rules (if the temp is above 72 degrees F, then turn on cooling until it reaches 72 degrees). Once it evaluates the data against the conditional logic, it performs an action.
Model-Based Reflex Agents
These agents are still reactive in decision making, but are more advanced than the previous example. They also receive environmental data as inputs via sensors, but the way they evaluate the data is where the differences lie. They evaluate what the current state of the environment is, if that changes without agent action, and what change the agent’s actions would cause. Then it runs this updated data through the conditional logic and makes the decision on next actions based on said data.
An example of model-based reflex agents are robot vacuums. These machines have sensors that evaluate the evironment around them (the floor) and consider how clean or dirty it is. Will the current state change on it’s own? How will it change if I act? The agent then looks at the if/then rules it is programmed with and makes the choice on what action to take. If you watch a robot vacuum move, then stop and spin for a second, then move… you’ve witnessed model-based reflexive AI in action.
Goal-Based Agents
Goal-based agents are the first in our list that don’t simply react to environmental inputs with a simple if/then logic. Instead, these agents have specific goals built in and the evaluation logic is computed based on these goals. In other words, instead of asking “what action matches this condition?”, but asks “what action helps me acheive my goal?” The fundamental difference is that goal based agents are not limited in actions they can take based solely on if/then logic, but which one best helps them acheive the goal.
Self driving cars are an example of this type of agentic AI. If a self-driving vehicle is given an address as it’s goal, then it will analyze the environment (where it is at currently vs the goal) and ask things like:
- “If I turn left at the next intersection, will that help me reach my destination faster?”
- “If I continue straight, will that help me reach the goal faster?”
- “Will the interstate or this side street be faster?”
This logic is constantly being evaluated and then car then takes turns accordingly until it reaches the destination.
Utility-Based Agents
Utility-based agents build upon goal-based agents, but instead of solely evaluating what the goal is and accomplishing it as fast as possible, they consider the utility of their actions. In other words, will this action both satisfy my goal and do so in the most efficient manner possible?
An example of this would be a package delivery drone. The goal is to deliver the package to a delivery address given. The utility considerations may be how much battery the drone currently has vs how much it would have after making the delivery. It would then decide the best route and course of action based on the utility of battery savings. After all, it wouldn’t do to deliver the package and then not be able to make it back to the starting point.
Learning Agents
The most robust form of agentic AI doesn’t just evaluate data from the environment and make decisions based off a goal. Learning agents include a feedback loop that allow them to evaluate previous decisions and learn from them. The agent has specific learning goals they wish to acheive and will store the knowledge from prior experience in memory to be used later as they advance towards the learning goal. Once the feedback loop and learning evaluation process has taken place, then the agent will take action.
An example here would be an AI chess bot. The bot plays the game and stores all the moves taken and various outcomes in memory. When it wins or loses, it then has several data points to learn from. Next time the game is played, it will now enter the feedback loop and critically evaluate the data in memory, plus the environmental data and make decisions based on both. It will then store the new data in memory as well. This process repeats every time the agent plays the game until it becomes a chess expert.
Sources