What is an AI Agent? How AI Agents Work, With Examples

Hey, my name is Mahan. I'm the CMO of Nairon. My name is Luka and I'm the CEO of Nairon. Luka, AI agents is a term that gets thrown out a lot these days. So, in this video, let's take some of the confusion out. Let's uh break it down, give the audience an introduction into AI agents and what the use cases actually are. So, an AI agent is a product just like how CHPT is the product that you interact with the LLM. An AI agent is a product that does some sort of action for you.

So if an LLM in a chatbot was more informational, more textbased, more answer question type interactions, an AI agent actually does things for you. What you can tell an AI agent is, hey, go through my emails, find me all of my most recent emails, and you know, reply to them. An AI agent will actually be able to do that with the tool calls, with the memory, and with the information that it knows about you. And we talked about all of these, right? We talked about what tool calls are, the ability to use Gmail and other tools, what memory is, the ability for it to understand everything about your business and context. And so, the biggest distinction as I just made is a chatbot responds whereas an agent decides and acts. part of acting and part of deciding what it needs to do is a lot of those reasoning type queries which we talked about in the last video right the agent is essentially like a looping mechanism of an interaction between an LLM to validate whether what you're asking for was done if not continue or whether it is done then present the final you know summary or solution for you you touched on some of the core components of what makes an AI agent be able to work let's dive deeper into those so the three ingredients that every AI agent needs is a goal, clearly stated goal with a specific outcome that's actionoriented, a tool set. So we mentioned Gmail for example, and then memory, this additional information that's associated to the tool of how to use the tool, what are best, you know, practices of how the tool is used in your company, in your organization, and about the goal itself. So if this was essentially a goal that you asked multiple times before, sort of the fifth version of that, you sort of documented the best case on what do you do if things go wrong and all that stuff. All that stuff is in memory. If we go a little bit deeper into what this loop really looks like and we give it a goal, which is, you know, research 10 companies and draft some outreach.

Notice this is not a question and answer type query. This is an actual multi-step process essentially like a workflow that we ask the agent to do. It's also not about one company but about 10 companies. So think about it. It needs to do the same thing 10 times. And so what it would do is literally in step number two it would use search web tool which is essentially a tool that uses Google puts a bunch of keywords and then finds the results reads the results. It then checks its progress multiple times as it's going. So it does the tool call, checks its progress, reads all the information about the company, keeps going until it does research on all 10 companies, and then what it does is it perhaps picks another tool called draft email, which is a Gmail tool in this case. Notice how step number two was a Google tool for example, and you can use any any different type of providers.

There's a bunch of different web search tools and a Gmail tool. So multiple tools from different companies. So we're now drafting an email and we're using a template and our research to fill out that email. And now we repeat that 10 times and we get to our final outcome which is the goal is achieved. We have outreach for 10 companies.

Okay. So essentially we're also looking at this being you know tasks versus workflows. When you look at a query that you give to a chatbot it's typically a task. The task could be just reply to me but the task could be also research this company, find this information, draft the email for me. But that's still viewed as like one task. With an agent, you can do that one task times 50. So it can essentially reproduce that task 50 times. The beautiful thing is that today we don't just use one agent at a time.

We've now started to develop multi- aent systems. So think about it this way. If I create an agent that drafts up some email copy, for example, replying to some leads, you can have an agent that does, let's say, the marketing side. And our agents can now start collaborating together. We can have skills and different workflows set on how our agents need to collaborate together to execute work that is not triggered by a human, but rather by a schedule or what we call a chrome job. And we'll get to all of these things in later videos, but it's really important to understand fundamentally an AI agent is a product that you build. There are completed agents that you can buy online and you can use online but most of the times these agents need to be built. They're not like ChatGPT where the chatbot is essentially built once you get to it.

Of course, something we'll be touching on as well is um multi- aent systems and those work really well for more complex tasks for bigger organizations. So for the time being guys, this has been an introduction to AI agents. We'll go much more in depth in future videos. You can subscribe to our newsletter links down in the description below. You can also find us on LinkedIn. We're very active on a daily basis. And until then, see you in the next video.