With the terms AI Agent, Agentic AI, and Multi-Agent Solutions being bandied around (including by me!), it is worthwhile to take a step back and try to define/describe what an AI Agent is. For instance, when is something an Agent versus something that I just chat with. The term Agent implies a degree of autonomy and the ability to make decisions, a good way of looking at this is to think of a spectrum of AI behavior:
So at some point in its growth our AI acquired enough autonomy and decision making capability to be considered an agent.
Let’s take a quick peek under the covers and see what things might exist inside an Agent:
Note the AI model, we can’t really have an AI Agent without an AI Model! Could be a LLM, some other type of model, or perhaps more than one. This is the brain of the AI agent.
One important thing to always be aware of is that we are telling the agent what we want it to do, not how to do it. It is the job of the Agent’s brain to figure out how to go about satisfying our request.
What I’ve termed Configuration may vary from model to model. But it is in large measure the instructions that tailor the AI’s behavior to the particular task at hand or give it the desired personality.
Tools are the things, external to the Agent, that it can make use of. Tools are typically described in some way such that the Agent can use its brain to figure out which ones to use and in what fashion.
As for all of us, memory can be important in being able to do a better job based on learning from past experience.
A deeper dive into some of the types of AI Agents and their internals will be the topic of a future post.