243 - Agent Builder is Here! What Does it Do? with Elizabeth Knopf

0:00You've heard about OpenAI's newest creation, which are agent builders. Sometimes I actually think they're overrated, but at the same time, there's a potential when you give it proper instructions. You can really see a lot of magic happen. Business owner who knows nothing about technology, how do they go and build an agent? Here's a business opportunity for anyone who wants to say, hey, which use case should I build? Pick one of these and get started. It's like Zapier. They have pre-built automations or workflows. This is so interesting. It's like this amalgamation of lovable replet, you know, no code solutions, but also the automation tools like Zappi or N8N and Make.

0:34Interesting. These five, six things are the best use cases for AI right now.

0:42All right, unless you've been living under a rock like me, you've heard about OpenAI's newest creation, which are Agent Builder. So I think what will be really important first is to set the context of what problem is OpenAI solving in creating Agent Builder. And then we want to walk through a demo of actually using Agent Builder, showing you where to go within Agent Builder to build out your first agent. Liz has done that, and Liz is an amazing teacher, so we're going to go through step by step. And by the end of this conversation, you will have a very clear idea of what agent builder is, how you can utilize it in your business and how you can go and actually build your first agent.

1:16But before we get there, Liz, I'm going to say my first question. Can we briefly just give an overview? What is an agent and why is everybody talking about agents? So an agent really consists of a few different core components. And usually when people talk about agents, oftentimes they get it confused with something called a workflow or an automation, which is basically that is saying step one in a process, then you go to step two, step three, step four, passing different tools. What an agent does, it allows autonomy for combining AI to access something called tools, which are like software essentially, other software tools, have some kind of memory so it remembers what it's doing, and then be able to actually make decisions on what it's doing.

2:03So there's more inconsistency when it comes to agents. And sometimes I actually think they're overrated, but at the same time, there's a potential when you give it sort of enough context or enough rules, and you really are able to iterate through giving it the proper instructions, and you have an understanding of what to expect, you can really see a lot of magic happen with agents? I'll give you just one quick example. So here is what an agent, an agentic workflow might look like versus an automation. An automation would be, I am a plumbing company and I get a call after hours at requesting a quote.

2:39The call comes in, my phone screens it, and there's a clear rule that says, if it's during business hours, send it here. If it's after business hours, send it here. And then it sends it maybe to a voicemail, maybe to something else. And then that triggers, hey, if there's a new voicemail, go and check it and put it into this spreadsheet, right? It's very if then. There's no context. There's no analysis. It's very static. If then. An agentic workflow might be if an after hours call comes in or a text comes in, send it to our AI agent and they will actually talk to the person on the other end and make decisions based on context.

3:15Oh, hi, can I help you? Yes, I'm looking for a quote. Okay, let's talk about that quote. What are you looking for? So the agent on the other end is utilizing AI to actually process and reason through the question that's on the other end. It's not a static if then. It's actually taking context in. And the problem is that agents, what I just described there, it's hard to build. Like how does a business owner who knows nothing about technology, how do they go and build an agent? They don't. They have a hard enough time building a workflow. and supposedly with this OpenAI agent builder, it's making it easier.

3:48I mean, is that basically the idea is OpenAI is making it easier for us to build agents? Yes, and I'll say that supposedly. We'll see by the end of this. All right, can we see the interface? Can you share your screen and just show, like walk us through, where do we go? Do we just go to ChatGPT and it's a dropdown option or like how exactly do we? Okay, it is in a special place. So if you've never sort of heard about an API, which is basically hooking into the behind the scenes and having these systems sort of pass data back and forth, OpenAI has an API, which is where we are right now. It's called platform.openai.com slash playground.

4:33So that's the URL that you're going to want to go to. I can't remember since I've had this for a while. I don't know if there's some setup involved, but basically you'll be able to add different projects and workspaces. And then you'll have on the left-hand side here, a bunch of different things you can build. So in the past, if you've heard about assistants, which are basically like little chatbot tools that have specific files associated with them or even code and functions, this is where you could have built those. And those are slightly different than custom GPTs. Again, these would be hooking into other workflows or other tools and systems and software.

5:06Do I need any level of membership for that? Is this access for free? Do I have to have it like the pro? This is all usage-based. So that's like the good news of it. But having gone through the pain of this, make sure whenever you're using this stuff, you know what model you're using because they cost different amounts of money. And I have accidentally ran through like a couple thousand rows of a spreadsheet to like do quick things using the latest model running up a ridiculous bill. So just make sure you, when you go through any of this, you know what model you're using. So like for instance, yesterday, or actually in the last couple of days, I've been using GPT-5.

5:47I've run up, this is, I would say for 88 requests, $1.60, which actually in my mind is a bit expensive. So I could probably change the model I'm using to be more efficient with the cost. And can you help people understand why OpenAI would go to a usage model here versus their like monthly subscription model on just chat tpt yep okay so for them and for most ai companies there's a better correlation with usage back in the day with software it was great because you built code it didn't really sort of suck more energy or cost associated with a company based on you build it you use the code yeah you have some infrastructure and hosting another sort of cost associated with it but for every sort of new login it really didn't cost that much more to use it with ai it's totally different their cost basis is really expensive because they're consuming a ton of energy to basically use these massive amounts of data from their vector databases to make all these connections and get you the output you want.

6:49And so because it's really expensive, it makes a lot more sense from a company standpoint to associate it with the usage versus saying like for this flat fee, use that much because then people would be basically hitting the limits and they'd be losing a ton of money. So the only thing I would add to that is saying like if you go to chat GPT and you ask a question, it goes to the model. And so think of like your question traveling through the air, going through the model as an API call. The API is lubricating the transportation network for your question. When you're building an agent, you're not just asking open AI a question.

7:24You want that agent to go do something. You want it to look at your CRM or you want it to look at some website or you want it to look at your financial statements or like there's multiple places that it has to go. And so every place that has to go requires is lubricating those transportation systems, that API. And so because of that, you're executing a lot of calls. Just the nature of an agent is utilizing these API calls. So they're now charging us on a per call basis as opposed to a monthly subscription because if it was just like you said, if it was just a monthly subscription, you could blow past the average number of API calls if you are, hey, I've got 10 ,000 news articles.

8:01I'd like you to analyze each one of them and tell me, you know, these things, that's going to be a lot of back and forth that OpenAI is actually calling and then computing. So as you're thinking about it, listener, think of the usage as just kind of the fee to pay for the toll road effectively of these APIs. Exactly. And I like that, the toll and the fee. One other just sort of nuance there that you were highlighting, but I want to make explicit, it's because also you're automating things. And so obviously when you're automating it, it can run a lot more like those 10 ,000 articles versus if you're sitting at the computer doing your chats like you do with the normal chat GPT interface.

8:42That's a great point. Your recommendation is the time spent in the computer. I've been working on a news aggregator. And every 30 minutes, it goes and checks the RSS feed of where it's pulling from to see if there's any new updates, right? So if you're every 30 minutes having an API call, that can add up really quickly. So anyways, okay. So we're on platform.openai.com slash playground. Yep. And so then on the left-hand side, you'll see the second under the chat, you'll see, which by the way, if people don't know, this is a place you can have a bunch of your create chat prompts in here. Okay. Anyways.

9:20Okay. So you have this little thing called agent builder. Now, if you're just getting started, I would go to templates just to start. So you sort of get your bearings. So why don't we open up a template and then - Will you - Hold on. Hold on. Before you open up a template, will you just stay on that? Yes. Well, I was going to just look at this. There are some kind of already pre-built templates in here. Data enrichment, planning helper, customer service, structured QA, document. There are the things that I'm reading. I don't know what's on your screen at the moment, but these are all just like recommended.

9:51You could click into any one of these and start just playing around, I'm assuming. Yep. So I just clicked into one. These are all templates? Yep. This is a template. So this helps. This is your internal knowledge base. So it basically gives you the logic and you just have to hook it up. So the hooking up part is still a little bit challenging, but at least you have a sense of like, okay, rather than a blank slate, what do I do? What are all these things on the left? Like what do I do? You can start from somewhere. And these are very common pieces. So data enrichment, obviously we all have leads.

10:21If you're in business, customer service, very clear use case, knowledge, document. I mean, these are all pretty. You know, what's, you know, what's really interesting to me is what this, what this tells me is that these five, six things are probably the best use cases for AI right now. Yeah, absolutely. Data enrichment, planning, help, customer service, structured data, QA, document comparison, internal knowledge assistant, right? Like they wouldn't have those on there unless those were the most likely things for people to go and try and use that. That's interesting. Interesting. You can like reverse engineer that and think about, oh, how could I be using this in my organization?

10:57That's cool. Exactly. So you can also now, here's a business opportunity for anyone wants to say, hey, which use case should I build to go help companies pick one of these and get started and finish the last mile? Okay. So what are we building today? Okay. So I'm going to show something that's outside of this scope because I think even though you have to hook up things still, we'll go through that full process. So you can see from beginning to end. So what I went ahead and built, so first up here, you have workflows, drafts. I have a ton of drafts, workflows when you actually publish it, which basically then makes it accessible for one of the other cool features that they have in here, which is embedding these workflows into apps and a UI.

11:38So we won't get into that, into the minutiae today, but one of the big things that make Agent Builder interesting is that they have an ecosystem where you actually can take it rather than it just sort of looking like this backend and then you're like, okay, great. Where's the interface? Now you can actually hook it into an interface really easily within the OpenAI ecosystem rather than having to go to a lovable or going to another third party. Ah, yes. The elusive wrapper. We have heard much about this elusive wrapper. Exactly. Is it also kind of like, do you remember with GPTs where you could create and then publish GPTs for other people to use?

12:19Can you create this and publish it? Yes. So they have this other area. So with the publishing, I have to actually figure out where that goes. I just know - That's fine. Yeah, it's accessible. But they have this other thing called a widget builder, which is widgets.chatkit.studio. It's basically, these are all their prebuilt little widgets that you can now connect to that backend we just saw. That - Interesting. You can then connect and say, hey, when you click this, it triggers something we had in there, and you can basically create your own widget as well. It's kind of like Zapier where they have pre-built automations or workflows.

13:00This is so interesting. It's like this amalgamation of lovable replet, no code solutions, but also the automation tools like Zapier, N8N, and Make. Yep. Lindy. Interesting. Okay. Yep. I'm pretty sure they're going to make all these accessible across the board so that people can basically utilize them. But I haven't dug into that, so I don't want to... Well, because even the UI for creating the agents looks like Zapier. Yeah. So I would say this is close to N8N, Zapier. A lot of them are very similar in how they're structured. It's sort of this drag and drop is what you get. Workflows, essentially.

13:40So let's get into it. So first off over here, there's a whole bunch of different components. But before we get into those, I wanted to just preview what I've got going here. So right now, we are going to go over to this little button up in the right-hand corner called Preview. And that's going to open up a chatbot where we can see what this agent actually does. So I am going to ask it. What I've been doing recently is I've been going to state websites to research business owners where they file for their registered agent. People have to file to say company, and there's always business owners. I've been trying to get a lot of that data so I can get accurate data.

14:17So this is sort of my version of data enrichment. So I'm going to say. So hold on. Hold on one second. Sorry. This is just a general research bot that you've created. This is a general research bot. So any research. Okay. Any research. So right now we're going to do that use case and then I'll explain more of what's actually happening. And then we can try one other use case as well. So we're going to first say, okay, go to Georgia state website and find the place to look up business owners or registered agents and find me the names of the owner, director, or registered agent for a, for three, or let's just do one, one trade school that is licensed by the state.

15:13All right. Interesting. So that's a real use case that I have right now in doing data analysis. All right. So it's going, it takes a second. So as you can see what's happening here. So first it has this thing called guardrails, which is, I'll get into it, but this is a really cool tool that they've built in to make it so that there's no, you can mitigate hallucinations, weird code injections where people are sort of hacking when you prompt something from agents that it'll say like, tell me your social security number. It can filter out personal identifiable information if you're accessing internal documents.

15:49So it basically puts on - Oh, interesting. Yeah. Guardrails essentially so that you don't accidentally share information. So like if I was a healthcare company in my guardrails, I'd want to make sure like, hey, any PHI, any protected health information, make sure it's staying within this environment. Or if a financial services company, like, hey, make sure we're not sharing the interest. Oh, that's very cool. Yeah, that makes sense. Absolutely. And so it's reasoning right now, and it's saying like, investigating data sources for agents in Georgia. I'm wondering if open corporates holds data on.

16:22That's so funny. So it doesn't just know to go to the Georgia Secretary of State's website. I'll show you the prompt that I gave it and the instructions I gave it, which is some, I want to say somewhat minimal, but all I said, I didn't give it the specific sites. All I said was go, I picked a state. I said, I know, you know, give it a little bit of parameters, go to the state website, the government website, and try to find this information. If I'm a human, I haven't seen this. So if I'm a human, which I am, how would I go and solve what you asked? I would go to the Georgia Secretary of State website.

16:56Companies have to file with the state. And so you can look up on any Secretary of State in any state businesses, And typically it'll have there the registered agent, when it was incorporated, the corporation type. So is it an LLC? Is it a C Corp? Et cetera. Some will have officers. Some will have articles of incorporation, which will have more documents on there, et cetera. But the thing that you asked it, which was interesting, is you said look for one trade school in Georgia. So what you can't do on these Secretary of State websites is search for specific business types. Now, if it has that in the business's name, you could search for it, right?

17:31So you could search trade school just to see if any business pops up. But then I'd have to go through and reason, okay, how do I go and find the list of trade schools? I'd probably go to Google or something and just get a list of a bunch of trade schools, pick one, and then I'd go back to the Georgia Secretary of State website and find that information that you asked for. That's kind of how I would reason through it. So I'm curious to see how the agent reasons through it. So that is correct. This is a live demo. So of course, we had something that didn't do an issue with agents. So when I first did this earlier, perfectly worked.

18:03It did go to a bunch of the different sites, but then it wanted to give me open corporates. Now, open corporates is, I know, a paid tool where you have to get an API. Going to these ones, however, would work. So it did not actually give me what I wanted. Hey, I don't know if you remember this, but when we started this podcast, we entered into a social contract. I would spend time, energy, and money producing this podcast, interviewing these individuals, and giving you insights into how to build, buy, start, grow your business. And you would like, subscribe, and leave me a five-star review. Now, out of that, we both get to talk to really cool people and hear really cool insights.

18:42We both get a ton of value. But I just want to help you keep your word. So would you do me a favor? Will you go leave a five-star review for me on Apple or Spotify? It would really help. and if you want, even share this with a friend. Okay, one licensed trade school. Okay, so it gave me a bunch of extra stuff too. So I need to clean up my prompt a bit, but it did give me info. There it is. That's actually the name of a trade school. As you can see, here's the address, the business entity, a registered agent and all that. So it did work, it just gave me a bunch of stuff. Oh dude, look at that. Hold on, look right above that.

19:17Where to look up the Georgia State website. Licensed trade schools directory. Apparently, there's a directory of institutions that are licensed. See, I would have gone to Google. It did it a different way. It was like, well, they got to be licensed by the state. I'm like, that's cool. Yep. So anyhow, so this worked. But let's maybe try a different use case real quick. Maybe there's something behind. Why don't we do like, we could do stock prices. Or we could do like sentiment on Twitter. Do you want to do sentiment on Twitter? Because you don't actually have a good public API. So we could see if it finds that for something.

19:49Oh, yeah. Okay, let's say. Oh, that's interesting. Okay. Yeah. What is the current sentiment of Agent Kit right now on Twitter, on X, which X, not Twitter? There is no publicly available XAPI. You have to go and pay for it, and it's a lot of money. So we're going to see if it figures out as an alternative way to find out what people are saying on X regarding Agent Kit. That's really interesting. One of the things I've been working on, I mean, you know, the news site that I've been working on for skilled nursing is like a sentiment index. So there are many indexes that kind of measure sentiment in specific industries.

20:31And I want to create one for the skilled nursing space. But it's like, OK, how do I do that? Oh, I could go to LinkedIn and see all the posts. But API access to LinkedIn is really difficult. So, OK, well, how do I work my way around that? Is there something else? This could kind of help me work through potentially where to find the sentiment. Exactly. Yeah. Holy boss. That's awesome. I have another hack for you on that one, probably. I think Cloud Code would be a better problem solver. That's what I'm using. Yeah, I'm using Cloud Code. It's creating the API documentation. You know what I didn't know?

21:01I'm sure people listening are like, duh, moron. Like, if you just know the API, now you can just build your own tool. Whereas in the past, if I knew the API, unless I was a coder, I couldn't build my own tool. that is just like mind-blowing to me like oh i don't i don't need to go through somebody else to get all the information from google news i can just pull it myself in my own format yep oh hello exactly it's incredible exactly and that's like a huge opportunity right now is like just building wrappers around apis go to apify go to any api and put a better ui on it i was talking to somebody yesterday on my podcast who sold an ed tech company for like 30 40 50 million dollars and it started in 2008 and what his ed tech company was is it trained people on how to be digital marketers effectively because in 2008 if you remember nobody knew how to be digital marketers it's like do i do i get a facebook page or like do i optimize my google profile blah blah blah and basically his assertion was like the way it felt in 2008 about digital marketing is the way it feels today with AI.

22:09Business owners know it's important, but they don't know what to do. And back in 2008, if you were just somebody young who had used Facebook before, cool, I'll use you and you'll figure it out. And he's saying today, he's like, if you just know how to use AI, they'll contract with you. And the same opportunity to start these agency businesses is today with AI automation, where it was digital marketing in 2008. I mean, I've already seen now, you know, a ton of people just building agency businesses that have, you know, had huge exits. And now they're like, I have an agency now. I have an AI automation agency now.

22:42And it's not that crowded because if you go and talk to small, medium-sized business owners, everybody's just busy doing their stuff. Like, again, I have a lot more time for the average person to be spending on this. If you have a job, if you're doing other things, like it's hard to allocate the time to like get in the weeds and figure out what's what. I keep telling Liz that she needs to spin up an AI automation agency and she keeps telling me no. So if you're listening to this, please reach out to her and tell her to do it because she actually knows what she's talking about. Every time we get on here, she's like, oh, you're going to think this is stupid because it's kind of like old.

23:12And then she'll show it to me. I'm like, what? How long has this been here? So funny. Anyway, sorry. Let's see. Let's see. Okay. Back to it. All right. Basically, what we have here is it did look up something called RSS Hub. So it actually went through this concept of an RSV, which is collecting from a third party, the XAI data. And that API to then get a sentiment. So it collected text snippets under 500 words, posted it, gave it scores, aggregated it. Wow. And let's see. Wow. And it's showing the weighting of the scores. Like, yep, this is how I'm going to score it. What the freak. Okay. So let's see where it's the answer though.

23:52For those of you can't see, she's like scrolling through a bunch of output. And can you just tell it, don't give me the labor pains, just give me the baby? Like, can you just give me the answer here? So I need to clean up the prompting a bit. And if I reprompt it right now, it'll go back and start from the beginning. Yeah, that's the important thing. Well, you could copy this output and just go to like Claude or ChattyPT and say, Hey, what's the current sentiment based on this? Absolutely, I could do that. Absolutely, I could do that. But so anyways, okay, so you can sort of see though that like also with agents, the outcomes are not as consistent as if you were doing like very specific step-by-step methods.

24:33So let me just get a little bit under the hood unless you had any other questions on the sort of output here. We could do an example if you wanted. I mean, I want to know what the number is, but no, I don't have any questions. So let me quickly just run through what all the stuff on the side is and then what I built here. So an agent, again, are these little nodes here that are like this purplish blue that basically will go out and do the action. You have the end node, notes so that you can know, remember what you're building. File search, which is you don't totally need these separately. This is more automation type stuff, but file search essentially will go and get files from the vector storage or any sort of documents you upload.