I got a private lesson on OpenAI's NEW Agent Builder

0:00Amir, by the end of this pod, what are we going to learn? I'm going to first talk through what OpenAI came out of yesterday with the agent builder, the chat kit, the widgets, and we're going to build a demo chatbot using the chat kit SDK on a website. And we're going to essentially have it pull data from our vector store, our multi-agent workflows, and have it answer questions. This chatbot is going to let you know whether you're actually a customer or a lead, and then get the information from you to pass it on to your sales team or answer a support ticket. So Amir, are people going to understand how to use Agent Builder by the end of this?

0:35That's exactly what we're going to cover. I'm going to try my best to show them what it takes to actually build your own agent workflow using the new Agent Builder, how it's different from the other kind of tools out there, and how you can get started as well. Okay, let's start.

0:57Cool. All right, let's jump into it. So the key three things that came out of the dev day yesterday was agent builder, chat kit, and widgets. And I'm going to talk through what each individual one actually is and what it means. So typically, you know, anytime we've been wanting to build multi-agent workflows, we've had to use custom code to actually kind of create a parallel sequence or multi-agent orchestration using code and say, okay, assistant one, talk to assistant two, and then pass through data or a set of instructions. What OpenAI has done with the new update is they've created a visual interface for you to actually build workflows using agents and actually create parallel agents if you wanted to or sequential steps in an agent workflow and have it called tools, do web search, or file all visually instead of using code.

1:46So it's really interesting because you can essentially now hold data as context from a vector store. So it's like a storage file. and then also evaluate the responses, refine it, and then create even guardrails for safety and quality of the agent responses. The key takeaway here is that it's essentially reducing the barrier for non-technical people to get started with building multi-agent workflows. And what this all means and how it ties in together is essentially goes into chat kit and widgets. So what chat kit is, is now a new capability where it's essentially an SDK and you can connect your agent builder workflow into ChatKit and then serve it on our front end.

2:28So in simple terms, anytime you've seen a chatbot on a website, that's typically connected to a third-party service that is pulling data and you've kind of created these responses. In this demo today, we're going to recreate our own, that's trained our own data and has a set of instructions and we're going to use ChatKit for it. So what we're going to do is we're going to build a workflow in the agent builder. And then we're going to take that workflow ID or that kind of template, put into ChatKUI, set up a server, add it to our website, and then essentially have customers interact with that chatbot to give their information to learn more about our product or answer a specific support ticket.

3:05The last thing is the widgets. Widgets are essentially a new set of dynamic components that you can add into chat interfaces and conversations that can display data. So say, for example, if you have your agent connected to a Shopify store and you're pulling through the MCP, you're pulling Shopify information, you can create a custom component that displays it to the user that says, oh, here's like what you ordered. Here's the, you know, I sent it delivery time and here's how it was sold. And, you know, you can get your details. It's like a dynamic UI essentially as part of the chat interface. God bless you, Amir.

3:40Yeah. All right. Great. Let's jump into it. Okay. So we covered all three big major updates It's obviously Sora 2 in the API, GB2 5.0 in the API. But I think what's really interesting is just the door and opportunities this opens up for a lot of people that want to start building multi-agent workflows. So the first thing that we're going to get into is jumping into actually the OpenAI agent builder. And how it works is you have a set of nodes that you can connect. So each node is representative of a specific set of actions. So you can add tools. So for example, if you want to pull relevant information or you want to add guardrails, and I'll show what that looks like and what that means, or MCPs.

4:20And you can also add logic. So you can determine how you want the agent to proceed based on logic and conditions that you set. And then you can also transform data as well. In this specific demo, what we want to do is we want to build a workflow that achieves two things. It receives a user input and it determines whether or not this user is an existing customer or a new lead. It classifies it and then based on that logic, passes on to two separate agents. Agent number one is if it's an existing customer, answer the support ticket using existing knowledge-based data. So I've actually scraped our entire knowledge base and data pertaining to our product and gave it as a vector store.

5:08So it's referencing that as context. or agent number two, if the agent determines and classifies us as a lead, it then asks for information about the customer to then, as a next step, pass it on to our database or message us on Slack, for example, if we have a Slack MCP. It's meant to capture that data and then play back to the customer and say, hey, we'll follow up for a demo and let's book you for a demo. So what we've done here is essentially we have the start input here, which is input as text, which is the message that you get. Next is the classifier. This classifier agent, essentially, we've named it and we gave it a prompt.

5:49And we basically say, hey, you know, we want you to look at the inquiry and tell us if this is an existing customer with a support ticket or a new lead. And we want you to analyze the message and determine whether or not how you get to that conclusion. And we gave some examples as well. So like, here's an example of what a new lead will look like, or here's an example of someone that is an existing customer. Once it's done that, we have a logic in place that says, classify that inquiry as an existing customer with a support question or a new user based on that data. And essentially, once it's been classified from here, if it's an existing customer, because the response here is essentially, like we're saying, the response is new customer.

6:35Oh, I got my cat in front of me right here. If it's a new customer, pass it on to the lead agent. If it's an existing customer, pass it on to the support agent. And how it works is essentially based on this logic here, we say if the input is an existing customer, pass it on to the customer support agent. And this customer support agent, essentially, it's trained on our data and it has a set of rules that it follows and helps them troubleshoot any questions they may have. Okay. How did you come up with those instructions? So you can actually write the instructions yourself. but what's really cool is you can actually either use chat gbt to say um i want you to act as a prompt generator and a helpful assistant that can help me generate prompts i want to achieve x tell me how we can get there so i usually most of the time use chat gbt or just from creating so many prompts i know how to get there um you can use it to give a prompt back to you you're essentially it's very meta you're using an agent to create agents agent prompts and um what you can also do as well as if you ever write a simple prompt so say for example um you want to enhance this so you can say use the enhance button here to say uh enhance this and enhance this prompt and provide a better better uh structure format you know this is kind of more um like a styling change that we're making here but if you wanted to kind of say i want you to enhance this to you know respond this way or have this tone then you can automatically do that in here as well so we've now essentially have capabilities to create these separate agents within the builder and connect it to different tools and settings so what that means is you can determine the level of reasoning so for example one agent you wanted to do a high level of thinking you want you're solving a very specific problem and the other you want it to be very minimal and just execute on the task at hand.

8:37You can connect different tools. So if there's specific functions, MCPs, or vector stores, you can do that as well. And you can also transform how you want the output format of the text to be. So in this instance, for example, I can change this to say, I want this to be in a JSON format. And I can add a schema to say, in your response, just this is how you should respond. Do not even respond in regular text. But for now, we're just going to do regular text just because it's easier. And at the same time, you can also connect it to different tools. In this case, I connect it to a vector store, which is a set of documents I've created as context for the agent to reference.

9:19And then from there, if it's not existing customer and it's a new lead, I have a sales agent lead. And this sales agent lead right here, again, is helpful and knowledgeable in capturing data about this lead, it'll ask them around kind of, what's your website URL? What's your company name? What's your email? How many visits do you get per month? Let's say we're building an analytics tool here. And what are you currently using? It'll gather that and structure the data so that next step you can pass it on to let's say your database or a Slack notification or add it to your CRM. Any questions so far?

9:55No, taking it all in. Well, I mean, actually one quick question. The reason why you'd want minimal reasoning versus advanced reasoning, is that just from a speed and cost perspective? Exactly. So the criteria around minimal or high reasoning is entirely dependent on the task at hand and what you want the actual agent to do. So do you want the agent to solve a very complex problem? Then you probably want high reasoning. Or do you want the agent to just execute knowing that's going to be a very simple task at hand? because, you know, maybe in this instance, because it's a support agent, I would probably maybe do Medium.

10:34But if it's a sales agent, it's pretty simple. It's like, just take the data and ask them questions. Like, what's your company name? There's no thinking really required for that. Cool. Yeah, cool. My cat just wants to be in this spot. And then basically, if you wanted to, the next step, just for this demo, just as a lot of configuration, I'm not going to do that. But you can actually add an MCP. So say, for example, if you wanted to add HubSpot and update your CRM, you can add that here. You can authenticate, add your token, and then connect that so that this agent, for example, can pull context from your HubSpot or push data as well to update your leads list in there if it wanted to.

11:14Yeah. And if you don't know what an MCP is, I have a whole video with Ross Mike. I'll include it in the show notes, clearly explain what an MCP is. But in layman's terms, what is it quickly? In layman's terms, an MCP is essentially a new interface for LLMs to interact with external tools. Typically, web apps use APIs to pull and push data. In this instance, LLMs use MCP, Model Context Protocol, to actually push and pull data within LLMs. And the MCPs that are available at launch are the ones that you showed? Yeah. So right now we have kind of the existing open AI connectors that are like the official ones.

11:53And then there's some third party servers as well. Hopefully over time we can get more of the official MCPs in there. Right. Intercom, customer service, Shopify, e-com. Yeah. Yeah. And at the end of this, I'll talk about kind of how this compares to Claude and kind of where there's opportunities for improvement as well and kind of how this differentiates. Okay. Yeah, you'll keep it real for us at the end of it. Yeah, I'll keep it very real. I think what's really interesting here as well is, you know, typically when it comes to AI workflows, especially for people that are just have, they're like on the, they're just getting started with AI adoption and they're just getting started with AI fluency.

12:31And AI fluency, I think is determined around kind of, do you understand how to prompt? Do you understand how to give the right amount of context? Can you take responsibility for the output and understand that you need to refine this agent constantly? Because I think from experience working with a lot of companies, I've seen that people that have still early AI adopters or like they're still not their late adopters, but early in their AI fluency stage, they have issues with building trust with agents, with the inputs, with the outputs that they get. And what that means is if the agent gets a wrong once, they immediately lose trust.

13:11And that comes down to understanding how to prompt, how to give the right amount of context, and knowing that you have to iterate on this and you can't get it right. Why I'm sharing this is because this agent builder has guardrails in place to help you kind of refine this process. So you can actually preview it in here if you wanted to, and we'll show a preview of that looks like. But you can also build guardrails to say, okay, like, I want you to hide personal information if this comes through. Or I want you to moderate this if there's anything harmful coming in. Or if someone tries to jailbreak this.

13:43Or if it hallucinates. Hallucinates is a big, big, big part of this where, you know, as you use more context, agents' performance degrade over time. So you can actually implement guardrails to ensure that your input and your output is actually structured the way you want it to be. So let's run an example of what this actual workflow looks like. So we're going to click on preview and you can actually test the preview in here and say, hi, I'm interested in a HumbleLytics demo. So this is just an example app that I have. And the classifier is now going to determine if this is an existing user or a new lead.

14:23And its reasoning is saying, oh, like this is a new lead. And it's now asking me, can you share a few details about your business? So I'll say my website is called AmirMxC.com. Company is AmirCo. Email is AmirExampleWhateverExample.com. And I'm doing about 10K monthly visits. And I'm using Google Analytics for basic traffic. And what it's doing now, it's pulling information around the vector store that we added, the files. and saying, okay, cool. I'm going to recommend a plan based on their needs and then also prompt them to book a demo if they wanted to. So it says, okay, cool. We got your details.

15:07Based on 10K visits and interests in heat maps and funnels, I recommend our plus plan to get started with. You can also book a demo right here or get started with a free trial. And if we wanted to, we can have an MCP that pushes all this data to our database or to Slack, sets a demo automatically, or even it's like in through here, creates an account if we wanted to. That's cool. Now that we have this builder workflow built out, what's interesting is that we can actually get this incorporated into a chat UI window. You can either use ChatKit, which I talked about earlier, which is a new interface for you to actually embed chatbots into your website, or you can build your own custom agent SDK if you wanted to.

15:49So you just have to paste over the workflow ID and the API keys that you have, and you can build your own chatbot. So what does that actually look like? Let's just make sure that we have everything set up properly. We publish this. And what's really cool is we're now removing a lot of developer dependency. What does that mean? So if, for example, in a setting, you have a customer support team that has built this agentic workflow, they can get the chatbot installed in their sites and make changes and not have to rely on the engineering team to actually deploy that for them. It's all happening live on the front end.

16:27So what that actually means is, say, for example, you have a website. We've now used ChatKit to integrate this on the front end. It's just a script we've installed. And now we have a fully working chatbot trained on our data and the multi-agent workflow. If I wanted to come back and change this workflow to add more agents or add more tools, we can just publish directly from AgentBuilder. and I don't have to go to the engineering team and say, hey, can you deploy this for me? And the cost of running this is just the amount of tokens, right? Exactly. You hook up your OpenAI API keys and just your server associated with it.

17:08Cool. So we essentially now have kind of like a chat bot that can now accept leads on our website. So I can just say, I'm interested in a demo. I have Google Analytics, but I want HumbleLytics 10K monthly visits. And this will now determine that I'm actually a lead and respond and essentially say, hey, let's get you booked in for a demo. We got your information. Let's proceed. And you can, you know, the agent builder has logs so you can track all that. Perfect. Yep. Crazy. It's pretty interesting. You can also, yeah, if you wanted to, have this completely as a customer support bot so that if you have issues, you can just say, actually, I'm an existing customer.

18:01I'm an existing customer. Actually, or let's just start in the chat. I'm an existing customer. Help me add a Webflow site to track. And hopefully, it determines that I'm actually an existing customer. and it'll give me insight on how to actually add it to, how to start tracking it. There you go. So we have essentially built a fully working chatbot using Context and RAG to first determine if you're a new customer or a lead, if you're a new customer or a lead, and then either solve your inquiry if you have an issue with the product or get information and get you set up. Yeah. So it's fully working.

18:50And what's really interesting is that you can actually customize the widget as well using the playground. So if you want to kind of have disclaimers or composers, it's fully customizable. And it's really simple to get set up with. You can just either use an embed code on your website. You just have to stand up a server to get this working. Or you can kind of build a very custom agent fully working within your app. if you have an in-app experience you want to have where you have a chatbot working with it. Yeah, so what do you think so far? I mean, someone's going to ask like, okay, well, why is this better than Intercom or a SaaS product I can go and use?

19:31Like, why do I need to create this myself? I think that's a good question. So, I mean, so first of all, there's two use cases here. If you want to use this internally, I think the multi-agent builder right here is still... there's still a lot of value out here, right? If you wanted to have a multi-agent orchestration and say you want to connect an MCP like Slack where it sends information, that's still useful in a sense where you have these backend occupations with multiple agents working together to get a task done for you. Now, if I'd say you are a startup, mid-sized company, and you want to save on costs and you have the engineering capabilities, then using these agent builders to then integrate with chat kit to get it on your app on your website could be a huge time saver in the future or a cost saver as well like there is a learning curve and a investment initially but over time i think you can have a lot of time savings and cost savings as well i also think it's a little more a little more custom you can really really fine-tune it exactly how you want it, right?

20:40Exactly. Yeah, you have full control over it. You own it in a way, like you all should essentially own the workflow in the system. There's a lot of great tools. Like if you're looking for something out of the box, like, you know, Lindy and Gumloop, they're all great tools. But if you want to build something more custom for yourself, then this is the way to go. Cool. Anything else? And then, yeah, I think the, you know, obviously the key takeaway here is, okay, like, you know, So one of the key takeaways here, it's a visual drag and drop tool. It's a low barrier entry for non-technical people. I think there's still some dependency where you got to have some technical knowledge, but I think the multi-agent workflow is very interesting.

21:18In common times, you see people using one chat window for multiple tasks. And that's not the right way to do it. You want to break up tasks into subtasks. I do think the Cloud Code SDK is still capable based on the model and the sub-agent orchestration. The only challenging thing is to get non-technical people playing with this, they can't use a CLI. That scares them out. So what's interesting is we've taken the capabilities of what these agent workflows look like and we've built an interface on top. People that are already familiar with NAN or Zapier. Cloud App is very similar in terms of projects and MCP tools you can build on it.

21:58Same thing with the projects in ChatGPT. I'm curious to see how this evolves over time where we have more NCP capabilities. Right. Just a quick note on that. So you are right. The CLI, the terminal, is daunting for people. And it's the equivalent, I'm old enough to remember, using MS-DOS to access a computer, which was basically a terminal. And computers didn't hit mainstream adoption until there was some graphical user interface on top of it. Microsoft Windows or Windows XP, I think it was, or Windows 3.1. So I think that's this moment in AI, right? We're putting canvases on top of sort of the hardcore technical hood.

22:51The average person doesn't want to be chilling in a terminal. Exactly. Yeah, exactly. And I think as we think about the models. There's so much emphasis on using LLMs and agent workflows for engineering and coding that the knowledge workers, the non-technical people have been kind of left behind. The experience is great for coding, but it's like, but how do we care to this for non-people that actually want this kind of use case? Which I think is a broader use case as well. So a lot of people, the common questions they have is like, how do I actually get started with this? how do I get started with agent builder?

23:27So it's available in platform.openai.com. It's not through child LGBTs and the platform side of things. I would say to get started, think about the use case and what it is that you actually want to achieve here. You know, for me, it was like, it'd be really cool to just have my own customer support agent. So I don't have to pay$150 a month and have it do exactly what it's currently doing right now, but also be able to actually capture leads. And I own it, I control it, and I can build more integrations afterwards. Then you work backwards. You say, okay, what does this existing workflow look like?

24:00And how do we actually build multiple agents that can play a part in this and have them be very specialized? The next step is, I think, building your data context, right? Capture your data, figure out what structure your data should look like, where you want to store it, what should the context be, clean up your data, and then add it as a vector store, as a file for your agent to reference. And I showed it in the agent builder and how you can actually reference that. Then, you know, the goal is to try to use as little context as possible to get the most out of it. Context has a huge impact on performance and it degrades it over time.

24:34And then if you need to, use multiple agent workflows like you saw, classifier, then we have the sales lead, then we have the customer support bot, specify the roles. And then from there, determine if you need external tools and MC, you know, like MCPs or web search. I would say Claude is definitely ahead of the game when it comes to MCPs. they're the ones that invented that. Yeah, they invented it, right? So there's a lot more directory and the directory is a lot more capable and there's a lot more features available when it comes to MCPs and Cloud. You know, OpenAI has got to step it up. They got to make it easier to get more MCP capabilities in there because that's the most important thing.

25:13And yeah, I hope it was helpful in terms of just kind of what came out and how you can get started. So yeah, so this is like super clear and that's why I wanted to have you on to just break this down. For the average founder who's listening to this, where are the opportunities? What should they be thinking about? So the average founder that is listening to this, where are the opportunities? Two parts. I think the unrelated, but OpenAI's app capabilities that's now available in ChaiGPT is huge, right? It's like we're now seeing ChaiGPT as a new distribution and new, to your point, interface layer to have it interact with your app.

25:55So that's, I would say, from a growth standpoint, use apps as a distribution channel. Specifically with Agent Builder and the ChatK UI, get this in front of your non-technical team members. Give this to your product managers. Give this to your customer support team. Give this to your go-to-market sales team. Give them an engineer to support them with building up the MCPs and workflows and standing up a server and see what they can create with this. Enable them to save time. And tell them to share this video and like and comment so that it spreads to the world. Yes, exactly. Amir, thanks for coming on and breaking it down so clearly.

26:36I'll include links to follow Amir where he shares knowledge on all this sort of stuff in the show notes. Appreciate you being generous with your sauce and so clear in your thinking. happy to help. Later. Thank you, thank you, sir.