OpenAI Update: Building Agents using NEW Agent Builder!

Category: AI Tools

Tags: AgentsDeploymentOpenAIOptimizationSecurity

Entities: Agent BuilderChatkitGoogle CalendarHubSpotMakeMCPNADNOpenAIZapier

Transcript

00:00

OpenAI just released their new agent builder and after building hundreds of AI agents on tools like NADN, Make and Zapier, I was super excited for this update. Now, in today's video, I want to review this, talk about its position in

00:15

the market, and actually build a live agent with you. And by the end of this video, you should have a good idea on whether you should use agent builder or not.

Now, I want to start by saying this is far different from an NAND or a make or a Zapier. And honestly, it solves a

00:31

lot of the problems that those applications bring when it comes to actually deploying chat bots and real-time agents. It solves a lot of those problems for you and it makes it look really nice too, the agent.

So, let's quickly review everything that's included with this new update. Now,

00:49

everything takes place on platform.openai.com openai.com and they've developed something called agent kit which is a toolkit for building, deploying and optimizing agents. So they kind of break it down into those three things building agents on OpenAI, deploying them and optimizing them.

And they have tools to help you

01:05

along the way through all of this. So you get to build in something called the agent builder.

If I click into the agent builder, this is just an open canvas that you can drag and drop different nodes. You can add like if else statements and it's very similar to

01:20

tools like NADN, make and any other thing that you've seen. We're going to be building an agent live in here, but I'm just trying to show you what it kind of looks like and the UI of this agent builder.

It's very nice in my opinion. I really like the simplicity of it.

And they don't just overload you with a

01:36

bunch of tools. They give you access to very little, but you can do some amazing things with this.

So, I want to show you. So after you build your agent in the agent builder, then you get to deploy them.

And they have two ways of deploying these agents. And by deploy, I mean like actually making them available

01:53

for public use. Not just keeping them in your workflow, but actually making them available to the public.

So you get to do that with Chatkit, which is another toolkit that they've developed that allows you to embed chat widgets into an existing website or application. And so

02:09

if you want to talk with your agent and embedded them to a website, this is a perfect way to do it. And they make it very simple.

Or you can write your own code and deploy it with the agents SDK. And this is what I really love about this update is the optimize agent performance because a lot of the time you build workflows, the agent stays the

02:26

same and it's really manual when it comes to optimizing the performance of the agent for a client or for yourself. So when you have this optimize agent platform with full evaluation and trace grading and data sets and optimizers, it

02:42

makes it really powerful when it comes to actually deploying agents that are useful in businesses. In my opinion, this is one step forward into the future when it comes to agents actually starting to make an impact in the market.

Before deployment was very

02:57

thrown together. It was very confusing, but they make it very straightforward.

Now, that's not to say that this is a good agent building platform, but I will say when it comes to like a secure, structured deployment and optimizing your agents performance, it seems like it could do a very good job. So, now

03:14

that we've reviewed everything that's included, let's dive into the agent builder and actually build out an agent. So, I'm at platform.openai.com/agent-builder.

I will leave this URL in the description if you want to come try it out. If you

03:30

create a new project, if you go to your default project, you're not going to be able to see templates. But if you create a new project uh in your OpenAI dashboard, I can just call this testing.

Then you're going to be able to see all of the templates that they've provided you as well. And some of these work out

03:47

of the box pretty nice. You can do a structured data Q&A, customer service, planning helper, data enrichment, document comparison, and internal knowledge assistant.

Uh, you can use all of those or you can just hit this create button and build your own. When you begin, it automatically imports a start

04:03

node and a my agent node. And when you click into these nodes, you configure them on the right.

So, you get to start in a couple of different ways. You can add things called state variables, which is the input that this agent will receive or that this workflow, what they

04:20

call them, will receive. And so it can receive a string, a number, a boolean, an object, or a list.

So pretty much very standard right now. Over here on the left, you get to see all of the nodes you get to work with.

So you have

04:37

an agent node, an end node, which is the very end of this agentic workflow. You have file search.

So you can automatically just throw this in here and then create a vector store, which makes Rag super easy, which is honestly

04:52

what I'm most excited for in this entire thing is the ability to just easily throw in a vector store and just go to the OpenAI platform and create one. So there's so many cool ways that you could use file search and I'm very excited for that.

But let's go over the other nodes.

05:07

You have guardrails. And guardrails are interesting.

Guard rails solve a big problem in the AI agent world. So if I select guardrails, we can actually connect this start node to this node, the guardrails node.

If we delete this AI agent node, and what this acts as is

05:24

a filter, a filter for sensitive information. So if I open up personally identifiable information and I select the settings gear, just take a look at the sensitive personal data that this workflow actually blocks.

if you wanted

05:39

to before it gets to the AI agent, the thing that's using the chat model. So security is always a question when it comes to building agents.

So you could block out anything you want before it ever reaches the model. So this makes it very professional grade, very secure,

05:56

and just a very good step to have in between information that people are typing into the agent and information that gets to the agent itself. So you could select all of those entities if you'd like and you could add and whether it passes or fails it could

06:12

go to the next thing. So it passes, you could have it go somewhere.

If it fails, it could go somewhere. Meaning if it has some of the information in here, you could just end the workflow.

So I could just add the end node and the workflow would just end. Guardrails also has a bunch of other things like you can block

06:28

harmful content in here and it has all of the filters on here that would really be helpful to have in other agent builders. Jailbreak, you can prevent things uh from from jailbreaking, AI safety rules, so prompt injection and things like that, hallucination,

06:44

continue on error. You have so many different ways that you can use guardrails.

And it's another one of those nodes that I am super excited for. Now, one interesting thing about the agent builder, if we keep on going down the list, is there's not really anything like web hooks, HTTP requests, or

06:59

anything like that in here. They focus heavily on MCP.

And so MCP allows you to connect to a bunch of different tools just like the connectors or similar to the connectors, you could connect to Plaid, Stripe, PayPal, Gmail, Google

07:14

Calendar, all of the usual tools. You can go to the ones by OpenAI or other developers and add a server.

And this is going to allow you to take actions in other tools that you are connected to. So you could do something like create Google calendar events for yourself,

07:30

which is very basic, but you could have it update your CRM and HubSpot. So many cool things that you can do with MCP servers.

And this is also a tool that you can use within your AI agent. So if you connect an AI agent to the start node, you can add it as a tool, an MCP

07:46

server. So, this agent can be instructed on how to use your Gmail properly and you could connect to the Gmail MCP.

And this agent could have instructions on how to use your Gmail if you need it to or if it needs to send an email for you. So, now that we've went over everything, let's go ahead and build out a quick

08:03

agent. Now, this is going to be a very simple agent.

I just want to show you how some of the tools work, some of the nodes work, and how it feels moving around this AI agent builder with OpenAI. So, I'm going to throw in an agent node after the start node, and we are going to connect it by dragging and dropping and connecting it to here.

Now,

08:21

in the agent node, we have tons of options. And this is one of my favorite nodes.

And by the way, all of these tools that you're seeing over here on the left, I'm just going to refer to them as nodes. But these are one of my favorite nodes of this AI agent builder.

You get to give it a name. And what I

08:36

want to do is I want to build out an agent that can look over all of my YouTube videos and give people answers to their questions. So if people have a question about my YouTube video, they can go to this agent, ask it, and this agent will only respond with information that's from my YouTube video and even

08:51

site where it got that information from. So I'll just name this like YouTube Drakebot or something like that.

And then I can create some instructions. So what you can do is you can expand this instructions prompt right here in order to pull it up in the middle of the screen and it looks a lot nicer to use.

09:09

So I gave very simple instructions. I say you are an agent that helps users get questions answered from AI foundations YouTube videos.

The user will be asking questions about artificial intelligence agents and automations. If someone has a question, your goal is to search the file store and answer using the transcript content.

09:25

Now you might be wondering what is the file store? Where is the transcript content?

Well, that's what I'm going to show you how to do next. Because what you have the ability to do, I'm going to save these instructions, is you can add something called tools.

You have tons of different things you can select here. Include chat history model.

They have

09:41

all of the models on here. For now, I'll just select GPT5.

You can select reasoning effort. I'll select low.

And they have a lot more stuff in here that you get to modify. But what you get to do is you get to select tools.

And when you hit this plus button, you get access to all of this fun stuff like chatkit

09:58

client tool, MCP servers, which would allow you to connect to Zapier, NADN, and pretty much anything that has an API documentation hooked up to an MCP server. You get this thing called file search, which is what we're going to use.

You get to give it web search access, code interpreter, functions, custom. You can do so many cool things

10:14

with these agents on the agent builder, but for now, I'm going to do file search. And what this is going to allow me to do is upload transcripts from my YouTube videos that I can then talk with or have other people talk with.

This could be embedded on a website or in my

10:30

custom application. So I'll hit upload.

And I've already prepared all of my transcripts from my last 15 YouTube videos. So I will open those.

And now OpenAI vectorizes those for us, which is very nice. So I'll call this the AI

10:46

Foundations vector store. And then I will attach all of those documents.

And these are just the files, but within the files, it's literally every word I spoke in each one of these videos that I've recently uploaded. So now I could ask a question.

I could leave output format on

11:01

text. I could hit preview in the upper right hand corner and I could ask something like what are the four stages of AI automation and agent building that

11:17

Drake brings up and I can send that off. I have a video titled the four stages of AI automation agent building where I go over four stages of doing so.

And what it's going to do is go through and actually answer that question and use the file store in order to do so. And it

11:33

says Drake's four stages are prompting with frameworks. That's correct.

Creating custom instructions, building automations, then building AI agents. And it cites where it got its information from.

So this all came from one of the videos. And if you go actually watch this video on YouTube,

11:48

the one that it's citing, you'll notice that these are the exact four stages. So, in this AI agent builder, with only two nodes and one tool, I've already created an amazing, very, very powerful workflow.

But let's say we want to get a little crazier. How about we add another

12:04

creator? My brother, the productive dude on YouTube.

As you can see, Productive Dude over here has his own YouTube channel. So, how about we combine this and make the agent make the user select which creator they want to get answers from.

So, what I could do is I could add

12:20

some instructions in here in order to make this agent understand that I said you have access to two different vector stores. One from AI Foundations, one from Productive Dude.

Each contains direct transcripts from their YouTube videos. If someone has a question, your goal is to search the file store of the creator they select.

So, now I'm making

12:36

the agent make the user select a creator. And then I say, make sure to ask the user which creator they'd like advice from.

I can save that. I can add another tool which is another file search and you can use two file searches within the same agent which is very cool

12:52

and this is just one use case and one tool that you can use and two nodes on the AI agent builder. So when I hit upload I will actually upload productive dude's latest video transcripts that I have prepared.

And here there's 14 of them. And so once they're all finished I'll hit attach.

And now this AI agent

13:09

has access to both of them. And I'll actually come in here and I'll edit this and I'll label it productive dude vector store.

Perfect. And I'll close that.

And so now the agent is not only the YouTube Drakebot, but it's just the YouTube

13:25

questionbot because it has access to two channels transcripts now. So if I hit preview, I could go here and I could just say I have a question and I could send it off.

And then it should ask me which creator do you want uh to get information from. So it's like it

13:42

has two consultants in this AI chat thread and it's going to walk you until you reach the end goal of your agent. It's going to walk you through that process.

It says great. What's your question and which creator should I pull the answer from?

AI foundations or productive dude. I'll say AI foundations.

13:59

And maybe this is embedded on a website or your application. Then it says great.

What's your question for AI foundations? I could ask, I want to know about his seven AI agent tools he likes and what they are.

I could send

14:17

that off because I have a video going over seven AI agent tools that I like to use. And then it gets all seven tools in order correct from that video in a matter of seconds.

Very quick, faster than I would have even remembered them. So it cites its information where it got it from and it gave me a good response.

14:34

You could have this walking people through a sales process and doing so many cool things, adding them to a CRM because what you could do is you could hook this up to an MCP and this MCP could be connected to your HubSpot right here or Pipeream or anything and you

14:52

could have it walking people through a process until they are warmed up and until your free bot actually helps them. But for now, I'm just going to leave it at this very simple agent.

And I just wanted to show you some of those capabilities. We're not going to be building out some crazy workflow in this

15:08

video. And then when you're ready to actually publish this thing, in the upper right hand corner, you can hit publish.

And then you can just select deploy to production and publish. And that will give you a couple options to actually deploy this agent.

You can deploy it using chatkit or the agents

15:25

SDK, which is a little bit more custom. Chatkit's like their out of the box.

You can embed this into a website today. And to be honest with you, deployment is an entirely new video on its own, but it is the next step.

And today, I just wanted to show you the AI agent builder, go

15:40

over my thoughts and opinions on this new feature in OpenAI. I think it's going to be amazing.

Honestly, I think this solves a lot of the problems that tools like Zapier and NADN have, and you can still connect to those tools via MCP in the agent builder. So, I think this

15:56

is a great spot to deploy, an amazing spot if you need a chat widget. I think this is still very early, but I would love to hear your thoughts in the comments on what do you think of this new feature.

What are your thoughts and opinions? This was just one very small example.

If you'd like me to come out with more, I would be happy to. And if

16:12

you would like to stay uptodate with AI, then I recommend joining our AI Foundations community using the link in the description. We love AI there.

We talk about it all day. Plus, you get unlimited AI support, courses, calendar calls, literally everything you need to master AI.