Build your own AI Agent with OpenAI Agent Builder

OpenAI has finally joined the race with a new platform that helps you make AI agents and AI workflows. And I wanted to show you exactly how to create your very first AI agent with this brand new platform. And by the end of the video, you'll have a working AI agent that will do exactly this. It will go search the web and find leads for you. Then it will populate a Google spreadsheet with those leads and with their email and all their information. And then it will pass that information to a third agent which is going to be your outreach agent. And that agent will take that information and draft customizable emails in your Gmail account and have it ready to be sent out in draft mode. And the building blocks will carry over to pretty much any AI agent that you want to make. Now, this is how you get to it. Go to platform. openai. com. openai. com. And right here on the left side, you'll see something called agent and build agents. That will bring you to this page, the agent builder. And then you could open up the agent builder from here. And this platform is designed for developers. But just like my other videos, I'm going to simplify this so anybody could follow along. You don't have to be a developer to use this platform here. Now, to get started, you could create your own. And by the way, we're not going to do any coding to create these. And you could also start with these templates. They only have a few right now, but with these templates, I'll be able to kind of show you the building blocks, and then we'll fully build one from scratch. And then by the time it's done, you could go ahead and test it even before launching to make sure everything works as it should. And if you're just brand new to AI agents, I thought this would be a good video to kind of walk you through it step by step, too. So, even if you've never built any type of AI agent in any other platform, this will still make sense to you. So, we'll start at the very top. What are AI agents exactly? Well, you're used to chat GPT, right? That's an AI chatbot. That is a reactive chatbot. You ask a question, it gives you an answer. That's pretty much it. AI agents are more proactive. You give it a goal and then it decides what steps it should take to reach that goal. So, it could gather information. It usually has access to different tools that it could tap into. It could sometimes make decisions and then it could produce a result. And a lot of these AI agents will also have some memory. So this AI agent builder that OpenAI created is a visual workbench basically for creating these agents. So instead of having to write code, you just use these building blocks. These are called nodes on this canvas, right? Each node performs one very clear job and then the arrows connect them to show the flow of data. So you could start with a blank canvas or from the templates. We'll choose this customer service template that's just on the template page. So I could show you the building blocks here. So all these little separate boxes that you see, these are called nodes. And nodes are basically the core of how you create AI agents in this platform. So in this case, there's a start node. This is a guard rail node. I'll skip over this one. It's a little bit more advanced. We'll get back to it. Then you have these AI agents. So this is a node for an AI agent. And then it connects to different conditions. This is another node and so on. Right? So you could see all these different nodes that you could add. Now on the left side you could just carry these over like this and then that will start a node. When we build one from scratch I'll explain a little bit more on how to build those. But this is important to understand kind of what nodes are and you could drag and drop any of them into this platform from the left side. The agent node. This one right here. Let me click on this one. This is basically the brain of your entire operation. And in this template, you could see there are multiple agents. An agent just doesn't have to be one. You could connect these agents to each other to make these AI workflows. And if you click on any of these agents or grab one from the left side here, you could see they have a name. They have a set of instruction that they follow. They also have memory with this chat history here. They could use different AI models depending on which one you choose to give it access to. and they could also access different tools and they have a response format. So this is the core of what makes an AI agent here. You'll also see these notes on top of this template. So that's this note right here. This is just a note. So if you are creating this and you want to remember why you did something or if you're going to share this later with anyone, that's what these notes are for. They're just for internal documentation on what these things are doing. You also have these tools right over here. So you have a file search node. So that comes in really handy. That way you could store searchable documents and it will create a library that your agent could pull from. Guard rails are also an important type because they are a safety check. They could detect unwanted inputs for example or outputs. Make sure personal information doesn't get out there. Make sure no one is jailbreaking it.

Basically getting the AI model or the AI agent to do something it shouldn't do. So these are useful. And you have the MCP. It lets you connect your agent to external services like Gmail or Zapier which then connects to bunch of other things or your company database. So this is going to be really useful and I have covered other agent videos on this channel plenty. So I have videos on make Zapier all kinds of different things NAT. So those are the competitors to OpenAI but I'll link all the resources at the end of the video in the description. Now, to see an agent in action, since we're using a template, we're on edit mode right now, but if we press play, you could see the agent actually going to work. This one starts with a chat. So, if I just type in a chat here, the agent will then go to work and try to figure out what's going on. So, it's going through the safety check here. Then, it's going to classify the text. You could see it in real time going to work here. And this is what it looks like internally. If you publish this, it will look like a traditional chatbot. it won't really give you these things, but when you're testing it, you want to see exactly how it's flowing through the information. And you could see this agent after it classified your text, it went through the set of condition like how should I route this information and it decided to send it to our information agent. We also have a return agent, a retention agent, right? So, this is a little bit more complicated, but I wanted to show you here with a template that had bunch of building blocks because we're going to basically go from beginner to pretty advanced in this one video. And we'll also talk about publishing here once we actually build one from scratch that we're going to go ahead and publish. By default, they all get saved to your draft. So, anytime you start one, it will create a draft. You could always come back to it. I'll jump into this data enrichment agent. This is a really simple agent just so you see what a very basic agent looks like because I wanted to show you all the building blocks. But an agent could start just with a start node here. Then this agent searches the web and then it basically gives you an output. Now this is super simple. You pretty much could use chat GPT for an agent this simple. But this agent will have access to all the different tools and things like that you can give it access to. You have your own set of instructions. So you could go a little bit beyond just using chat GPTR if you build these. But the whole point of these is actually that it's modular. You could really build it out, right? Things that you just can't do with a simple chatbot. Okay. Now, let me show you how to actually build an agent from scratch. And I want to build a real agent, a really powerful agent, too. Not a simple one that you could see from these templates, but we'll start simple so you could follow along here. And then you'll see the working and building blocks of this. If it's too complicated, then you might want to look at a different platform like Me, Zapier, or N810. I think they're a little bit simpler. This is really more designed for developers, but you can make really powerful agents here. And it just came out, so they're probably going to tweak this a lot more in the near future, too. Now, before we build this lead generation agent, I also have a resource that shows you how to actually implement AI agents to save you time at work. is called how to use AI agents to save 20 plus hours a week from HubSpot. It covers five AI workflows you can implement at work and it also includes real tools, prompts, and agent templates you could use today. And my personal favorite in this list is the company brief generator agent. It's an AI agent that retrieves company information on your customers and is perfect for any type of B2B type of business like my production company for example. You could also type in a URL and it'll create a report for you. And it's not a simple report that you'll get out of chat GPT. This agent pulls not just from the web search. It also uses another tool called glean search. That's like a company's internal search tool and your customer database via another tool called datab bricks. And it outputs the same result in a consistent format. So once we built our agent with OpenAI agent builder, you could build similar and even more advanced agents using what's covered in this guide. I'll put a link in the description where you could download this guide completely for free. And thank you for HubSpot for sponsoring this video and making this resource available to my audience. Okay, let's build our lead generation agent. You'll just click create here to create your agent. And the best place to start with this agent is actually the tool that it's going to use to populate our leads. So all I did was open up Google Sheets. Google Sheets is the best place for this. And I created this template here. You have a name, a business type, the address, the phone number, the email, a column for description, what makes that business different, and if an email has been sent to them or not. This could actually do that step two. Now, every agent usually comes with these two nodes. You can actually delete these here and start from scratch, but you definitely need a starter node. So that can be deleted here. So an agent can be grabbed from here and then your start

node will need to get connected to it like how it was in the beginning. So it's really easy drag and drop, right? So we have our very first agent here. This is going to be a complete workflow. So it's going to have multiple different agents doing different things. So this very first agent is going to actually find us leads based on what we search for. Okay. So I'm going to name this agent. This is our very first agent in this sequence. It's going to be three different agents working together. And this is going to be a leadfinder agent. And then you're going to give it a instruction. These are called system instructions. And I'm going to explain this to you right here. Let me open this up. So this is where real prompt engineering comes into play. It's not just basically talking to chat GPT, although they call that prompt engineering too. But prompt engineering at its core is when you build an agent or any type of AI app. What's the instruction that you're giving it? This is where that comes into play. So, I'll make a page for you so they're really easy to copy and paste here. But this first one has a different instruction than my next one. This is you are a helpful lead generation agent. And I give a context on what we are and what I wanted to do. And then I told it to pass the information to our next agent, which I could always come back to once I build this out because our next agent is a data entry agent and I already kind of build this out so I know what the next one is and I kind of specifically told it. The more specific you get with these type of prompts, these system prompts, the better your agent is going to do. And in the middle, I also explained basically the different columns that I had made in my Google sheet because I wanted to go find the information and then eventually put that information there. And I'm going to save this one. Then you also have chat history. So you want to turn this on. It's on by default. So it could have conversational memory. Then you could choose your AI model. So you could use chatpt 5 5. 1. You could also choose the type of tools that he has access to. So the MCP server is going to give us access to pretty much any tool in the world. But in this case, we have file search, we have web search. Web search is what we want. Now chatpt by default has web search, right? So you're used to it. But that's technically chatpt calling a tool called web search. That's what it's doing in the background. Now, I don't have to fill anything out here, but you can actually change the context size to be high. You want it to be able to look for a lot of information. The high lets you do that. So, I'm going to go ahead and add that. And the output format is going to be text. I'm going to get text and I'm going to pass it to the next agent as text. Now, already we have an AI agent. Okay. So, if I press play right on top here, it's gonna bring this chatbot. And if I talk to this chatbot, it's going to go and try to get us leads. Okay. but it will get us leads and respond to us in chat format, which again we could use just about any chatbot for what we want to do now is actually get those leads to nicely format in our Google sheet and then reach out to them with email with our follow-up agent. So the sequence of agents is going to do that for us. So we'll go back to edit mode here. Now let's grab another agent node. We'll drop that here. So again, they are not connected by default. You have to grab this right here and then connect them like that. And with this agent, what we're going to do is this is going to be our data entry agent. So, it's going to take the leads that it found based on the different columns for our sheet that we told it and it's actually going to input those into that. So, with each agent, you want it to have a pretty narrow role. That's how you get it to really do what you want for the most part, right? You don't want it to give it five different things to do like you might with chat GPTs. Kind of one at a time and then you could just stack those agents on top of each other. Okay. Now this data entry agent might seem a little bit complicated because it will have to use an MCP server to do it. Now if you use other versions of AI agent platforms that let you build them for example make Zap year and N8N they directly connect to other tools like our Google sheet. This does not. This requires an MCP server to connect to different things and we'll actually use Zapier in between the two. Okay. So, I'm going to go to tools. I'm going to press the plus sign and I'm going to choose MCP server. Hopefully, they just roll out those other tools like the other platforms have it. So, you don't have to do this middle part here. Now, there are some tools that are related, even Google Drive, but the actions within those is very limited right now. And Google Sheet is just not one we could select here. Even the Google Drive option doesn't really do what I want in this case. But the nice thing is Zapier is an app that connects bunch of different apps. So, I'm going to select this instead. And there is a website here, mcp. zapier. com. I'm going to go ahead and copy this and open this up. It's going to ask you to log into a Zapier account. So, if you don't have one, you'll have to create one here. They do have some free credits available with it, but sometimes you may have to upgrade depending on if this ends up being really useful for you and you're going to use it a lot. And with Zapier, I could basically connect to 8,000 different apps and they have like

thousands of actions across all those apps that you could use. So it could find a Google sheet, it could write in it. All those things are done through Zapier. And then once you log into Zapier, I'm going to choose this one, new MCP server right on top. And then it's going to ask us what our client is. That's just basically the company and the API that we want to use. So OpenAI agent builder. That's the platform we're using. So we're going to choose that. I'll call this lead genen agent here. And we'll create our server. Then it asks you what tool this MCP server has access to. What can it do? Google sheet. And in this case, we want the agent to be able to create rows inside of the Google sheet we already created, right? You can give it access to all kinds of different things that you can do inside of Google sheet. I'll choose this option right here. And you could also choose some configuration here. So obviously I need to choose an account for my Google Drive. Then I'm going to choose right here, set a specific value from my Google Drive. I'm going to choose my Google Drive here. And then we want a specific spreadsheet, right? because I only give it access to that spreadsheet. So, I'm going to say right here, set a specific value and then I picked that specific spreadsheet. So, that's why I created the spreadsheet as the very first task inside of the same Google account. So, make sure that was the first thing you did here. And I also made sure the worksheet is set to sheet one. That document only had one sheet. So, this is just more extra here. But everything else I'm going to leave on default, right? The AI needs to fill out all those different things. But you can see it pulled in that information at least the header for all those different sections of that sheet. So I'm going to save this. Now we have to connect it to OpenAI. So right on top it's going to tell you that's your next step. Now this creates a API key or basically a secret key. It's a handshake between these two platforms. So they could actually talk to each other. So copy this with APIs. Anytime you see APIs just that's a key that lets two different apps talk to each other. in this case, the agent builder and our Zapier account. So, I'm going to copy this one. And this is the Zapier MCP server that we've created. I'll go back to agent builder and we'll type in that API key here. This is our Zapier MCP. So, I'm going to connect the two. Now, this part is actually asking you, do you want it to access and add other tools and edit other tools? I actually don't want that. I want to limit the control that this server has. I don't want it to just add bunch of tools in Zapier. doesn't need to do that in this case, but you can add that and it's on by default. I only want it to have access to the thing I'm telling it to do so it doesn't get confused. So, very specifically, I check this on and I'm going to add it. And then you'll see it added it over here. One more thing I actually forgot to add here. So, I'll go back to it. Under approval, always require approval for all tool calls. In this case, I could say never require approval. Now, you can turn this on if you're going to give it access to bunch of tools and it might have too much control. In this case, I limited it, so I'm going to say never. So, it doesn't ask me and it could go to work and I'm going to update it. So, anytime you could click a tool to update it settings. Okay. Now, every single time you create an agent, the instruction part is a very important part. So, let's go ahead and open this up. So, in this case, I really spelled out exactly what I want to happen. So if I was teaching someone for the very first time how to do this for me, I would walk them through and they would watch me do it. So while I was doing it manually, I just wrote down all the steps, right? Retrieve the lead information from the previous agent and then pass it to this agent. And these are the things I want you to put in these columns. And that's it for this agent. We'll go ahead and save this one here. And we'll have our final agent. So, so far it's going to go search the web, find the leads based on what we're going to give it in chat. Then it's going to go and add that to our Google sheet. Then we're going to actually have the agent outreach. So, I'll call this one outreach agent. So, this will literally do the outreach for us, right? It's not just finding leads, giving us a sheet. I wanted to take that even a further step here. And since I already showed you Zapier, which is probably the most complicated part of using this is that MCP server. So, it can't find tools. I just felt like without tools, this would not be that good of a video. I feel like custom GPTs could almost do that inside of chat GPT. Now, this also needs that tool first. Let me just change the agent and then we'll come in and put in our instructions here, but let's go ahead and choose our Zapier tool. So, it's going to be MCP server Zapier again. So, same kind of thing we looked at before by going to this site again, this portion. Okay, it took us to the same place again, but we're going to add a new tool and this is going to be Gmail. So you could send out an email and draft it. And with Gmail, we want to create a draft. So you could search up here, create, create a draft is what we want. In this case, I do want to manually look at this, but then I could just go through a bunch of different email drafts that is created for us and then press send. I find that to be a little bit safer. And for configuration, again

add to your right accounts. Now, the body needs to be generated with AI, so we're going to leave that. But we do have some advanced options. One of them that I do recommend is body type. You could actually set this to be HTML. So, I'm going to change it to specific and HTML just because HTML just creates better formatting, the headings and things like that. But the plane sometimes does work. This is one thing I do like to change. So, I'll save that. The Google sheet also needs to be able to look up a Google sheet because the previous one only let us write something in a Google sheet. So, I'm going to add a tool. We'll use Google sheet again, but this time I want to look up a specific row in a sheet. So, we'll select this option. And then again, I'm going to choose specifically that sheet, right? That spreadsheet. Even though AI might be able to guess that, I don't want it to make any mistakes. And the worksheet, again, I'm going to choose that one to be our only sheets there. Okay. And everything else I'm going to leave on default. So, pretty much every time I'm just choosing that specific sheet here to make sure it doesn't look elsewhere in my Google sheet or my Google Drive. And then one more tool inside of Google Sheet that I need to add, which is updating the sheet. I'll choose this one here with line item support. I'll choose advanced and then do the exact same thing. Basically, choose the right spreadsheet and limit it to that one sheet on that spreadsheet. Let's save that. Okay. So, Google Sheets has three different actions that it could take inside of that tool. and Gmail has one now. And all we have to do is go to connect here, copy our secret key, and then take it to our agent builder. Again, I'll name it the same thing here. Let's connect. I changed the approval to never require. And I also uncheck these two that are checked by default. And I'll add it now. And for our instruction, I really spelled out what I want to happen here. And you could ask chat GPT to kind of help you expand it out if you just use a few couple of sentences maybe to describe it. But I like to really spell out step by step what's happening. I just found it that prompt engineering with these system prompts makes for good AI agents. Otherwise, it's going to be a little bit frustrating because it's just going to go off the rails here. But again, I'll have a sheet with what I'm using here with those three different agents and the system prompt you could copy and paste. And it's ultimately saying to find the leads and then draft a tailored email. That's why I had a one row that had the description field that made that difference. So, it has that information to actually make a better email outreach, right? It actually finds things that are relevant to that business. It's not just a plain email it's writing every time. That's the whole point of the lead genen to customize it for that business here. And it will also be able to change what's on that row. So, I know an email has been sent for example. So, right now everything was set to no. it will change it to yes once it drives that email. So we'll save that. Okay. I'm going to go back to that same sheet. I'm just going to erase everything. So then we'll populate here with new information. Okay. Now with the agent complete, all I have to do is go up here and it creates a chatbot for us. Right. So this is the same as talking to chat GPT except it's now going to go through this workflow, reference everything we have. And my prompt actually in this box is really simple. Find me five video production companies in Chicago. So, you just have to type in the type of business and the location. And could be anything, dentists, law firms. It will still work with all the system prompts that I've set up. So, this is going to first, you'll see every time an agent gets called, it will go to work. So, the start node was just the chat is finding me those video production companies here. And then now, okay, it found me five. The data entry agent is going to go ahead and add that to our sheet, which I've cleared out here. And let me let this one go to work here. This may take a few minutes. Okay. So then your second agent should populate this Google sheet with the all the right information and with the email address. And in this case, all these emails that are set to be sent should say no. And the descriptions all should be different here because it's going to use that information to create a custom email for every single person. It's not copy and pasting the same exact email. And then the third agent is the outreach agent that should create the drafts for you in your inbox. So these are the different drafts that I have here. And I'll click on one of them. And it's nicely formatted here. It has the subject line that is different based on that production company here. And this part of the email is different every time. That's what the agent extracted. Now you could personalize this a lot more a lot differently than mine based on your system prompt, right? So, make sure your system prompt tells you how you want the email. You could even give it an exact type of a email template. And you could always publish these. And I actually forgot to name this one. You could always name it right on top here. So, we'll name it now AI Legent Agent and publish. Now, this is going to be published in your workspace. So, this is how I would use it every time. You'll come back and use it here with press and play and use it here. Now this is where

it gets a lot more complicated and definitely requires a lot more advanced knowledge and even a developer. They have something called chatkit which then you would have to share to add this to your website. So this part does require a lot more advanced knowledge. If you want to put this in your app or your website, I recommend getting a developer for this step. Just build the agent, test it out. You could use it here anytime you want. But if you do want to add it to your website, definitely not simple right now. I still use Chatbase. That's my favorite platform for creating agents that could easily get embedded to your website. I have plenty of videos about that on my channel as well. Now, as I'm recording this, the Zapier MCP that is in beta. The OpenAI agent builder also in beta. So, you may come across some issues. I just wanted to make this video as this came out, but it does have issues. It sometimes took me three or four different chats to get it to work. That's why I don't think it's ready for publishing publicly, but it's great as an internal tool. So, hopefully this gives you a good understanding of what the Open AI agent builder is and how to get something out of it that is useful. And there are plenty of other AI agent builders out there. Some of them are a lot easier like Zapier, Make Natn, and even Google has a new one too. And I've covered most of them on this channel. So, I'll put some resources in the description. And on our platform, Skill, we also have an AI agent builder course along with 30 other courses that you get access to. I recommend you watch that with the free trial and I'll link that in the description below this video, too. Thanks so much for watching this video and I'll see you on the next