Transcription
Chachi Bati just launched their brand new AI agent builder, and it's absolutely insane because this now allows you to build out AI agents and AI workflows that you can publish on your own website, in your own tools, or anywhere that you want.
By the end of this video, you're going to know exactly how this works, how to set them up, and I'm going to share with you a few crazy use cases that I'd strongly suggest that you jump on.
Okay, so here's exactly how this new tool works. What you're going to want to do is you want to come to platform.openai.com/agent-builder. openai.com/agent-builder. And you're now going to see this agent builder right here. And what this allows you to do is build out your own workflows and your own agents that you can test immediately. You can run evaluations on immediately. You can optimize immediately. And then you'll be able to publish these to your own website either via code or via chatkit.
Now, here's exactly how they work. Essentially, what you're able to do is you're going to see a start node right here. And then we can come over here and we could define exactly what the workflow inputs are going to be. We could see that we could add in input variables right here. We could add in state variables. They could be things like strings, a number, a boolean, an object, a list. And you could see you can actually add these in right here.
Now, for the sake of this, I'm just going to do input as text, and it is going to be a string. So essentially, the way that this is going to look is this is going to be a chat similar to ChatGPT, except we could give it access to custom tools. We could train it in a custom way. Basically, this allows us to manipulate ChatGPT or any of OpenAI's agents however we want. Check this out.
So, if we come over here, we have this start, which again, is just going to be somebody chatting. And then we can actually build in agents here. So we come over here and we can name the agents. For example, I'm going to put "YouTube Title Generator." Now, we have instructions here. So, we have to come in here and actually give this system instructions. We could say something like, "You are an expert at generating YouTube titles when a user tells you the topic for their video." And as you could see, we could actually come through and we can enhance this prompt right here. We could add in a new prompt message, and we can adjust a bunch of other things. For example, we can include chat history right here. We could change what model this is actually using. We could change how high or how low the reasoning effort's going to be. We could connect this to a bunch of other different tools, which I'm going to share with you in just a second. We can change the output format right here. Right now, it could be text, it could be a JSON, it could be a widget. We could come over here and change the different model parameters. For example, how big the summary is going to be, how verbose it's actually going to be, aka how many words it is actually going to use in its response.
Then, in addition to that, we could set up for chatkit whether or not we want to display the responses in the chat, whether or not we want to show the search resources, and we could choose advanced things here. For example, write to conversation history or continue on an error. So, if this runs into an error, we can make it so that it just continues to chat over here. And basically, this is where you're going to go through and this calls the AI model with your instructions and your tools, and then it can actually give the user an output here.
In addition to that, what you can do is add in a bunch of things. For example, if we come over to the left-hand side right here, we could add in notes. So we could come over here and we could type in here and say, "This is a YouTube title generator agent," and then we can actually drag this and we can put this so we know that this is actually with this node right here. In addition to that, we could call on different tools. For example, maybe we have a bunch of files that are a bunch of different titles that have done really well. We can actually add this in right here and we could add that this is actually going to call that. We can create the vector store. We can create the query right here and see how many results are here. Or we could add in guardrails if we wanted to.
And again, this is as simple as us just clicking into these things and us telling it what to do. For example, if somebody's trying to jailbreak it, or if it hallucinates, or whether or not it should actually continue on an error, we could get this to call MCPs. If we click into this, you could see that they don't actually have that many. In fact, they only have a few. There's eight by OpenAI MCPs that this hooks up to. So, I want you to think about this. This can actually call your agent that you're going to give whoever you want access to. It can call on Gmail, Google Calendar, Google Drive, Outlook, Outlook Calendar, SharePoint, Microsoft Teams, and Dropbox.
Now, that might seem incredibly limiting until you come into "All" and scroll down and you notice Zapier right here, because guess what Zapier allows you to do? Zapier allows you to connect your AI to any app with Zapier MCP. In fact, this allows you to connect to more than 8,000 different apps, which makes your agent way more powerful because with just these integrations, this is all right. I mean, this is okay at best, but when you connect this to Zapier, this makes this absolutely amazing. In fact, this is the only way that you can then connect the agents that you're building with OpenAI with ChatGPT to thousands of other apps. This is literally the only way.
And if you wanted to actually explore those apps, there's several different ways to do it. For example, you could come over here and you can type in what app you want to see if this integrates with. For example, I want to see if it integrates with Snowflake. We could type that in right here, and it will pull up whether or not it actually does. Or you could scroll through the more than 8,700 different apps. Or you could tell it what kind of use case you want to do. And guess what? It will pull up what apps can actually get that use case done.
Now, my favorite part about Zapier is that you could get started with it today for free. In fact, if you go to the pin comment below, you could get access to it right now so that you can hook up your ChatGPT agents to more than 8,000 different apps. In fact, this is the only way to hook them up to thousands of other apps. So, what are you waiting for? Run to that pin comment below right now so you could get started with this.
Now, in addition to that, there are a few other things that we can add into these agent workflows. We have "if and else," so we have logic right here. We have "while," we have user approval. And then we also have different data things. For example, if we need to reshape data, we could do that with this. If we need to set a state, we could do that with this. And again, like I shared with you before, we could come over here and we can actually preview this at any time up here.
Now, the other part of this that makes this incredibly useful is you could see that they have a bunch of different templates. For example, they have ones around data enrichment. They have ones around planning help. They have ones around customer service, internal knowledge assistant, document comparison, and they even have ones around structured data Q&A. Essentially, what these are is we could click into them, and they're templates that are already built out. So you can see exactly how these are actually going to work.
So we could see that we have a start right here. We have this, this goes to a classification agent, or it ends. We have different conditions. We have a return agent. We have a retention agent. We have an information agent. And we also have "end" if it doesn't actually fit here. And we could see this goes through all the way to the end. So this is what happens right here. Somebody comes over here and asks a question. This determines if the user input is reasonable. If so, it identifies the user's intent. If it fails this and it doesn't actually give a good input, it just won't answer them. And then this will route to the correct agent right here. It will handle each customer's intent differently. Then it will provide them with an answer right here and then actually give the reply over here. And you could see you can add in a hallucination guardrail here with vector store in order to actually check accuracy.
Now I want you to think about what can make this way more powerful right here. You guessed it, using something like Zapier. So, if we come over here, let's say that we have all these different agents doing something. For example, this information agent. Guess what? We can actually get this to do. If we come over into tools right here, we can actually give this access to specific things. For example, maybe we want to give this MCP access to Zapier right here. And then you just enter in your Zapier API key. You give it a description, and it will then be able to hit a bunch of different things. For example, let's say that we have all the different knowledge that we want this to look at inside of Notion. Well, guess what? Notion wasn't one of these tools right here. So, we could add it in right there, and then it'll be able to access our Notion. Or maybe we have a bunch of information that we wanted it to be able to access on something like Slack. Well, guess what? You'd be able to do that same exact thing. You'd then be able to look at Zapier inside of this agent right here and not just have MCP access as a tool, have it as a tool that your agents are actually using.
Okay, so let's say that you've actually gone through and you've built one of these out. For example, I built out this very simple research agent right here that will go through and actually look at what I'm trying to research. It will then do a web search for that. It will then summarize and display that thing exactly. So, we're going to click on "Publish" right here. We're going to call this "Research Agent." And then you will see that we will be able to publish this. We could see that this has gone through. This has been published. And there are two ways for me to be able to actually show this off. Agent SDK. We could actually copy this code and then put it on your website, and it will then actually make the agent pop up. Or if you come over here to Chatkit, we could follow the Chatkit quick start right here, which basically shows us exactly how to set this up on our website. And again, you don't have to have any code in order to actually do this. And we could see over here that we could then add this to any domain that we want.
Now, if we come back over here, we could see that my agent is going to be right here. We could see any drafts that we're working on in here. And then again, we could find all these different templates inside of here. So, let's say that we come back over into this agent. We could come over here. We could click on "Preview," and we can say, "Please do research on Apple stock." Now, what this is going to do is we can see this actually go through and hit each of these right here. We could see "Start" right here. We can see how this web research agent is actually behaving over here. We could see how "Summary and Display" is actually behaving. And then we could see what is actually going to get displayed over here. And we could view the details if we wanted to. We could close this. We can open up a new chat. We could close this preview.
And what's even cooler is you could see that we could come over here and see all the different versions that exist for right here. And guess what? Inside of these tools, we can actually add things like files and stuff like that. So you can let your users do that with their tool. In addition to that, if we come over to "Evaluate" over here, this now actually goes through and we can create graders. For example, we could describe the pass/fail criteria, i.e., is the agent polite? And we can actually go through and we can add in several of these. For example, I'm going to say, "Is the agent accurate?" We're going to add this in here. I'm going to say, "Is the agent thorough?" I'm going to add this in here. I'm going to say, "Is the agent lengthy in its response?" And I'm going to add this in here. And then, guess what? We can actually go through for all the different chats that it's had, and we could go through and grade all of these with literally just one click of a button, which is mind-blowing.
This allows you basically soup to nuts, be able to build out AI agents. And again, with Zapier, you are then able to connect this to thousands of different tools. And guess what? We can now go through here. We could see that this test is in progress. And you can see that you can literally do everything straight from here. We could build out agents. We could monitor their usage. We can check the logs. We could check storage. We could evaluate them. And we can fine-tune them all from inside ChatGPT now. And this is going to kill so many agent builders. And then we can see from here that we actually go through and this will actually look at each of the chats that have occurred and determine whether or not it has passed all of these different emails. And again, this is an absolute game-changer.
So, in terms of actually using this and making this useful for you, I would recommend that you start with two things. One, I would recommend that all the different projects you have in ChatGPT and all the different custom GPTs that you have, begin to build out AI agents and AI workflows for these so that you can put them on a website. Why? Because this will allow you to put them behind a paywall, and you could begin to charge for all the different knowledge and all the different use cases for ChatGPT. And again, hooking this up to Zapier makes your life 100 times easier because it allows you to then give ChatGPT access to thousands of other apps. And the use cases that you can build this out for are absolutely insane. But essentially, what you need to know is now you can build ChatGPT and all of ChatGPT's capabilities into everything that you do on your own website and don't have to rely on just custom GPTs or projects in order to get it done.
Now, if you enjoyed this video, I'd strongly suggest you check out this video right here that walks you through a bunch of other ChatGPT updates that just dropped and new features that you're not going to want to miss. I'll see you over there.