Enhancing AI Agents From Basic to Intelligent Solutions

everybody. Hello, this is Jim from Augusto Digital, and today I’m going to talk about how we can make LLM or AI agents more intelligent. which will make it become contextual. We’re going to add search, which allows it to become current, get real live information.

And lastly, we’re going to add tools, which will give it our business information. So you’ll see where it starts and how limited it can be. And we’ll show how we can expand that to make it more useful for your organization, or the things that you’re doing for your home.

So let me switch to our demo here. I’m in a tool called N8N. We use this at our organization for workflow. But it’s a really good tool to showcase how AI agents work. And as we add tools to see how valuable that can be, have. We’ve got a simple chat agent right now that agent has a system

prompt that says you’re a helpful assistant. Please attempt to answer the question is best that you can. I’m currently using chat gtp mini 4.1. We can quickly change that to something a little bit more robust. It will go to gpt 5 mini. This demo doesn’t necessarily matter for that

but it’s fun to be able to see that you can change the different LLM models you’re using. To demonstrate how limited an agent can be and how much information is in a model, let’s start with a really simple question. Who is the president of the United States? This question will hit our agent.

It’ll come back with information and like any LLM, it should answer that. right, you can see it says I don’t have any real time updates, but as of June 2024, that’s when this LLM agent was trained, you can see that Joe Biden was the president. And you can tell me

that you can go as of today, so it knows the current date, you can go check other sources. So right now, we can tell our agent only has limited information about when the date it was trained. We either need to give it more information, go out to the White House and tell it that, or we’ve got to

get some more ways for our agent to become more intelligent I also want to validate really quickly is I’m going to ask it what services does a gusto digital provide. question. Again, this was trained in 2024. It may or may not actually have information on a gusto digital.

I could have asked it what sources were there. I did not. But one of the key ingredients that I want to share here is this LLM is limited to the information that was set at that time. All right, the agent came back and it wanted me to specify, what do you specify the company aimed at gust of digital?

If so, can you share their website? Common services of a digital agency. So, it didn’t necessarily give me back real information. As you can see here, it also didn’t hallucinate. It basically said, but because I can’t understand who a gusto digital is,

it says that common digital agency typically would provide these things, right? So further. One thing to note now is my agent doesn’t actually have memory tied to it. So it won’t be able to follow up a question like who is the president of the United States. I could give it more

tools and if I ask who’s the president, it will probably ask me another question of case which country. It doesn’t have that context of what it is. So what I’m going to do now is add memory to our agent. Now I’m going to be using redis in this particular case and I’ll give it

a context window of 10 in history and life and and we’re going to give it some things for it to remember. So since it doesn’t have access to the web right now we’re going to tell it for this conversation. Remember that Augusto Digital specializes at workflows,

automation and custom GPT’s for mid-size businesses. So I’m gonna go ahead and spit that to it. And so now you can see my workflow here. It added some information to the chat memory. It’s going out now to the OpenAI agent. It’s output was, I got it. For this conversation,

I’ll treat a gusto digital as specializing in workflows, right? So I’d outline today a workflow. It’s asking me other questions I wanna do with that memory and it’s stored. Now, what’s great is we can prove that memory is working. Let’s go ahead and ask it a question based on that,

based on what you remember about Augusto. Notice that this time, I did not say Augusto digital. I’m gonna allow it to make the assumption that Augusto equals Augusto digital because now it has memory in context to be able to make that decision.

So what kind of clients are the best fit for outcomes that we should pitch? So let’s see what it does this time. Now it went and it took its memory, right? and it has memory and information previously that can handle. We’re gonna let this quick run. All right,

it came back. It says, great. Using Augusto’s digital focus, AI workflow and automation at GPT’s, so it spit back what was in memory. Says the best client fit profiles, midsize businesses, the tech stacks around HubSpot, Salesforce, any signals, got some red flags based on it.

So it used all that information to come up with a pretty good enablement and context around what an AI agency that we talked about. And lastly, we’re going to do a quick memory check going back to prove it’s there. I’m just going to say, what services did I say that Augusto Digital specializes in?

Now what’s here, you notice that quickly went to memory, it did touch on the AI agent. It said that you said that we specialized in AI workflow and GPDs. So we’ve done now as we’ve enhanced our agent with memory. It can remember our conversations. It can draw from information we’ve told it.

It can actually help us be relevant in what we’re doing. So next what I want to do is allow our agent to become more relevant with new information. So I’m going to go ahead and add a tool that’s going to let it search the web. This tool is actually called Searching

and G. It’s on my system here on NADN. There’s other tools. Again, I’m using this NADN as an example. This can be done with any other workflow tool or AI agent if you write one yourself or if you’re using tools from whether it’s Azure or whether

you need AWS. now let’s go back to what our original conversation was. I’m going to ask our agent now, who is the current president of the United States? Please verify with sources. So what’s different now, rather than our first query, is it has the ability to go out to the web. Really quick here,

you can see it went to memory, it went to the AI agent, and it went to search. So right you can see here who is the current president of the United States please verify If I was sources, it searched the web twice here in this particular example. It came back with Donald Trump.

It will probably have sources that were relevant to it. I can see Wikipedia in one of these examples here. back. The current president of the United States, starting at January 20th, is Donald Trump. And it gave the sources of the White House.

So not only is it more intelligent with the information we gave it for memory. Now we’re stacking on information that’s coming from the web. Let’s continue down what we’ve done with Augusto. Find Augusto Digital’s website and summarize the services listed there in

six bullets. So I’ve told it what I believe that Augusto Digital’s services are. Now it’s gonna go out to the Augusto Digital website and it’s gonna pull those information from there. points. While this is running, I just want to quick show the response that came

out of searching is it went to the Augusto website and it followed it and it pulled the snippet information back. So let’s actually take a look at the answer. It says based on the information from Augusto’s website and it gave the URL. There’s a summary of six items. AI acceleration,

custom software engineering, user experience, digital transformation and data analytics. So I can provide more detail from specific pages if you You’ve asked it. So now, again, it knows more about Augusto Digital. It knows more about the presidency.

It’s got more relevance. this is a pretty solid agent. You could use this for a lot of things right now. You could increase and change the model to have a more thinking power. What we’re gonna do now is give it contextual information based on your organization.

We’ll call this knowledge. To do that, I have created two different Excel spreadsheets. spreadsheets. In this case, I’m just using Google sheets. So let me go ahead and connect these up. So I’ve got two sheets and we’ll quick walk through them. Let me enable them.

So one Google sheet has got Augusto products. And I just made fictitious products. Let’s go take a look here. So we’ve got a skew. We’ve got a product name. It’s called flow pilot, some kind of a dashboard, it’s analytics with a category, it’s got a price, a status and some notes.

And then I also tied in kind of some sales information, Q2 sales, you know, the SKU sold from the Midwest, the unit sold and who the sales rep is. So now. Our agent has the ability to reason with the LLM model. It’s got memory, so if you remember previous answers,

it can go out to search the web. And now I’ve given it company information it on the gusto products and our sales. So let’s start asking it some questions that might be more relevant. with what products do we sell and which ones are performing best

this quarter? Now I could have said Augusto Digital Products but I’ve given it two different sheets to be able to look up so it needs to look up products and it needs to look up information from the quarter. So let’s watch what the AI agent does with these tools.

You can see that first it went to memory didn’t have that information. It went out to the products list and it also went to sales. So Augusto Digital sells the following products. Now I did not ask it or tell it what products do we made

the assumption I’m with Augusto Digital because I told it that earlier for memory. now what products do we sell? It went to the first sheet and pulled back the products it also told me what category it might be and a little bit about it including the price. It then pulled out the quarterly sales.

And then it came back, it decided to tell me the top selling products with the number of units. Now, it can do some deeper analysis. one thing I will tell you is LLM’s are challenged a little bit with doing math and calculation. is I’m going to give it a tool of a calculator. Instead of the LLM,

which is a large language model, doing that calculation right, I’m going to give it the ability to use a tool, a calculator. This will allow the LLM to pass off any hard calculations Now there’s a couple of reasons to do that one. You’re going to get more accurate and

detail information. But two, a lot of times the agent, the LLM’s, will use more tokens, tokens are crossed, to generate and make statistical and mathematical claims. So the calculator just gives it more energy to do that. So let’s go ahead now. Let’s go from the products information,

products information, Google Sheet list all the products in the AI category include their status and the monthly price. So let’s just go ahead and watch what it does. In this case, it did not have to go to either one of these. This was able to pull it back from memory.

Again, saving time by going not having to reach out to these data elements, it pulled back that information for me. So here are the products in the AI category, and here is their monthly press. That’s all relevant because of our memory. The system can do that.

Let’s question. Now, we’re going to do a little bit more cross look up. So which product category generally the most total ARR for this quarter? And based on trends, what should we expect in the next quarter. now hopefully we’re going to get it to both

look up sales products and do some information research using calculations. You can see our output here, it came back, it actually went to memory to generate this. It said let’s calculate the ARR, which it did for each one, gave back with the summary, and then it said which category.

And then based on what we should expect for an X-quarter, it says automation products, and it used that information now limited as it was. It did generate that. And lastly, it says focus on AI offerings for Q3. All right, for my last demo, I’m going to go ahead and add Q3.

Q3 sales sheet. So similar information on Q2, Q3. And now we’re going to ask it another question. All right. I’m going to do a question around which sales rep benefited the most in Q3 growth, and what products contributed to that? Are there products we should remove

or let’s just leave it that and let’s let it do some calculations. now what we should see is it’s comparing the products versus the quarter. It already has Q2 sales and memory and it’s going to do this comparison for us. All right,

let’s take a look. Here’s the analysis for Q3 sales growth by sales rep and products. Sales rep for Parmix. Alex sold this many units. Jamie sold 155 and then and calculating the ARR, you can see what it did here. Alex was 347,000 and you can see here that Riley was at 509. Now,

Power Performance will Riley with 509 and it primarily contributed to the AI assistant, the dashboard, and others. Products to consider removing. Here’s a couple that it listed and it’s probably gonna give a few recommendations. We should evaluate some feedback and others.

So I hope this example shows how a basic agent can be powerful. But by adding tools, by adding your information, you can enhance this and really make it a strong second brain to be able to make decisions for your business and other operations. If you have questions, feel free to reach out. Thank you.