You’re Using ChatGPT Wrong - AI Agent Builder Changes Everything

OpenAI have just released AI Agent Builder, and that is a drag and drop workflow builder that allows you to build extremely complex workflows with ease, with no technical background whatsoever. And it also includes an extremely cool, probably the most exciting thing of customizable widgets, where you can change color schemes, you can change its aesthetics, you can change the roundness, the shape, you can change the different options that it has. So it gives you a fantastic output at the other end, allowing you to deploy really complex flows that also include widgets as you head down there. Well, it's my job today, Michael Elliot, to go through this, starting with the individual canvas and brief, run you through a few templates so you understand exactly what is happening inside of OpenAI's agent builder. And then from there I'm going to go and build a live demo of some flow in action so that you can see its power. Let's jump in. So this is OpenAI's agent builder. I have it in dark mode. You can have it in whatever mode that you want it to be in. But when it comes down to simplicity, I don't think you would argue with me when it. When I tell you that it's as simple as it could physically possibly be in terms of just the structure you've got. Create a new workflow button. Here we've got workflows, drafts and templates. Drafts are a few drafts I was working on this morning, workflows and workflows that I've published. And then templates are obviously templates that have been created by them and that is where we are going to start. So if we look at. Let's take the planning helper section, you can see this is exactly what the canvas will look like for you. Whether you're in dark mode or light mode, this is what you will have. Now these are just, if you're familiar with nan, these are just notes. So this is what explaining what is actually happening. I strongly advise that whenever you create any automation, you have notes alongside them just for future reference, but they're not necessarily necessary. You could literally just delete them all. And here you can see that there's obviously the start, which is the input as text. So that's us literally adding the text to the. The flow. Then you have the triage. And if you look at the instruction section here you are an assistant that gathers the key details needed to create a business initiative plan. And its response format is schema. But down the side here we have include chat history. Yes. So it has a memory so that it can continue to back and forward with us. We can select the reasoning effort low, medium or high to see how much effort the the AI actually puts in. This template has medium. But what is the most amazing piece of kit inside of here is you can change the output text widget which is the key chat kit, which I'm going to go into a little bit later on. But when it comes down to the output, the JSON JavaScript output gives it then the condition of being true or false. So we basically if input parsed has all details equals true then it will follow the true route which is the launch helper. If otherwise it will be, it will be false. But each individual node you can see, so this just has the variables in it, the most basic variables possible. But these individual, these beautiful things here, these are actually agents. So as you can see down the left hand side we've got a list of individual core widgets and there aren't that many, which is beautiful because it makes it simple for you. You can manipulate these, you can add guardrails. So like if you don't want offensive language to be used and things like that, you could set them in place or a jailbreak one. So to avoid it breaking and falling over and giving information that it shouldn't, you can set those up. But the point I'm making is There is what, 11 individual nodes and you can build pretty much anything on the planet with it on a canvas that looks as beautiful as this one. So let's give it a go. Let's jump into this. Just before I do, it's probably a good idea that I showcase this to you. This has like a very brief instruction. Collect the missing data from the user. Look through the conversation to extract the following Goal team or organization what the team or organization wants to achieve Target completion date or time frame Available resource or currency capacity Headcount budget tool access. So if we hit preview on this and we say I'm looking for a business plan to contact thousand people and I have a budget of $500, what is the best approach I could use? And once we fire that off we can start and you can see it literally in live in the information I'm missing a couple of details called target completion date and time frame of reaching a thousand people. Target completion date within 30 days and email is the preferred. Oops, let's remove the microphone. You can see how beautiful this is. So completion date and time frame. Target completion date is 30 days and I prefer cold email. So now it's got the information to finalize the plan, well it has more information, team size and headcount available to do the work. 10 team members. I don't have the leads yet, but I do have instantly. Let's see what this does. Where it tells me to get the leads from. Here's a step. So now it's gone from getting the data of actually getting the data into showcasing me the actual solution. So this is its tailored approach. Step by step plan, contact lead generation, list building. Divide 10 team members into pairs, each pair responsible for finding 100 leads. Use three free and low cost tools like LinkedIn, LinkedIn Sales Trial, Google search industry directories to build qualified contacts, validate emails. Cold email strategy draft 2 to 3 short personalized email templates. Email sending set up to 8 to 12. It gives you a perfect layout. But what I've just showcased to you there was actually like error handling which is fantastic. Like I gave it information but it didn't have the correct information. So if it failed and I didn't provide it everything, it didn't just die and stop like an LLM would. Like historically an LLM would just be like oh dear, like what? You know, like might ask you a question, it might actually lead you down this path. Then it gets roped up and hallucinates. Well this flow here can tell you how to strategize to reach out to a thousand people. It's not the best. You change the prompt. These are templates of course, but what I'm saying to you is you can see how easy that is to connect. Like this can give you dedicated business plans on just about anything. Using five individual, individual nodes that have the most simple and basic instructions added to them. Wow. Like I was blown away when I first looked at this. You can change a few things. Display response in chat, show search sources, continue on error writing, conversation history. You want those on. There's nothing that you would really change there. Um, so so far so good. Like absolutely amazing. Now let's go back to templates again. Document comparison and here this is a fantastic workflow. So what I'm going to do this once. This needs to. This needs me to upload two documents that are similar. So like two lease agreements. I'm going to go and grab them from chat GPT and I'll be back. So I now have two lease agreements but let's try and break it again just to showcase to you how powerful they are to pass it back. Please give me information on both docs retry agent. I don't have the documents. Please upload both. Okay. No problem. So let's upload the lease agreement and let's upload this lease agreement and let's fire them off. Now that's going to go back to the triage request but now it's going to go to propose reconciliation. So it's gonna go through this. Here's a clear summary of each document followed by key differences. It gives me the locate where they are 12 month term, the rent, the deposit. So it's given us all of the information in real time that we need in order to be able to make best in class judgment based off of these two PDFs. It gives you the key differences. DOC 1 is residential tenancy. DOT 2 is an official office lease Sheffield flat versus Manchester office suite. If you would like I could now process a single reconciliation option that aligns the second documents changes with the first baseline or tailor a hybrid lease if you're trying to standardize terms across residential. So let's approve it. This is a binary approval approve or reject. So this is now the approval agent and I think you can see pretty clearly how this works now when it comes down to the approach to, to building these. You can see real business use cases here. Like you can see how this would be deployable as a flow which is always key whenever you look at any tool. I'm big on not looking at every shiny new tool. This isn't a shiny new tool. This is a groundbreaking tool and this is something that you need to learn how to use. And it is going to be the future. Eventually NAM will have the same exact process. So when it comes down to next steps, if we just go back and finally look at this, look at what it, what it, what it provides us, it allows us to be able to send the emails. Okay, I've reviewed the lease documents. I propose concise comparison. I can draft suggested revisions. If you want to harmonize terms please let me know which deliverable you prefer and you can fire off an email. Pretty cool stuff. For Google's new. Sorry for, for OpenAI's new AI agent builders. That is incredible. And these are done by simple, simple agents with simple context information. The only thing I had to do was upload document documents inside. I don't know where they was. They would have been saved. Maybe they weren't safe because I come out of the preview. But when you upload the two documents when prompted by the retry agent which simply has that basic instruction which is, which is a, which is brilliant, right? It's absolutely amazing like that the builder is fantastic. And because it's built on the open AI platform with the accesses that it has. It is something that you should 100 be be using. So the one of the things that you definitely should be using when you think about these workflows is just this. It's guardrails. So guardrails are give you an abundance of options. I'm not going to deploy one. I don't think it's necessary because it's quite. It's basically what it says on the tin. But you have options inside of your guardrail node which are hallucinations. So it detects hallucinations confidence threshold. You can set it to minimum confidence threshold of like 30. So like if it thinks that you know it's something's going to hallucinate then it will. It won't proceed. You can turn that on jailbreak. When it comes down to jailbreaking you can like have it if detects in prompts injection and jailbreak attempts in order to be able to basically take the model off of its task. That is something that you don't want to happen. Moderation to make sure that any aggressive nature or harassment or something that you don't want to come through your flow happens. You can do that. Or personal identif identifiable information if you under strict guidelines to keep that confidential etc. You might want to use those. But that's definitely an honorable mention when it comes to mcp. MCP is crazy. This is where you can get nuts because when it comes down to like your MCP sort of structure I can. I'm going to save this for the. For the build flow which is what we're going to hit in next. But MCPS allow you to be able to connect with Stripe, Intercompact, Dream Deep Wiki, Shopify, Gmail, Calendar Drive which are the main three that I would use it for. Dropbox, also Intercom. You can have all of these things that are actually available to you. Let me just go back into that a minute. Let me get rid of that seems to be breaking it. When you go. When you go into. Never mind. I just didn't click the button. By open AI by other developers server you can add your own MCP server which then gives you access to just about anything that you physically possibly want. You can see the direct connections. I'm going to build a flow in a moment that uses one of my direct connections in order to do it. But I think those are important. One of the key things that obviously we that you have is user approval if statements. So inside of here you'd use a common expression to create a custom expression. So basically this is a loop and this is what you'll have seen. Loop. They have the loop node in nan. You'll be able to see it. It'll wait while something is happening. If something in your flow takes time, it will wait. User approval means a human step needs approval, which is important if you're auto quoting and things like that. You'd want to use the user approval step. If. Or else if statements common sense. Exactly what it says on the tin. Transform data and set stay. These are individual modules that you can use to transform data objects, add schema, add individual properties. You can add them as strings, numbers, booleans, objects, arrays. It's very similar to all the other no code platforms but you have the ability to manipulate the data in the same way. Sticking on task of this actually being a basic overview. The final thing I've got to show you probably is the note. Sticky note. Now one thing I would say, I'm going to say this because this is just a bit weird. When it comes down to AI, when it comes down to open AIs, agent builders and how to use them and like this tutorial, when it comes down to the. The sticky notes, just be careful with the sticky notes because one thing I've noticed is it doesn't give a clear signal that they are, that they are small. So like if I was to just smash a keyboard like that, right, there's no indicator that there's anything else there. Unless I hover over it and see a scroll bar, I can't see it. So very minute detail. But I just thought I'd let you know because it caught me out in one of the templates. I'd shrunk it and I was like, what? Like where is it? And then I found out that it was there. So that's just something for you to, to keep, to keep note of when you're doing this. But it all starts with this beautiful start node. So now we have an overview of the canvas, we have an overview of the individual tools it can use. We've run through live example. It's my job now to go and build a custom flow and I'm going to connect, I'm going to connect my Gmail and make it maybe send some emails or something. I'm going to do it, do a flow which is actually workable in a business use case where you might be able to see it. So yeah, let's jump in. So now let's jump into a live build and what we're going to do, we're going to build something really simple. I'm going to use a couple of nodes that we haven't seen previously. The input variable is text. You can actually change this to boolean, object, list, whatever. But we want it just to be a string input. And then I'm going to actually add a guardrail. And this guardrail is going to be based off of personal identity verification, moderation, jailbreak and hallucination. And we're going to set it to a tolerance of 40%. So that should give it pretty good confidence of a pass and end rate. We then drag these nodes out. If it fails, we want it to end, we don't want it to, we don't want it to continue. Because with some of, well, depending on some of those things like the hate speech and things like that, you wouldn't want it to continue. And then to continue, we're going to add an agent. So top left we'll hit agent. Now inside of the instructions for this agent, please give me feedback on my top five performing videos. And then this is where it gets really cool reasoning effort. Medium add tools. We can add web searches, code interpreters, functions, custom file search or an MCP server or indeed a client kit tool itself. Like where you put this, where you put this in. So you can obviously go deeper with what would, with what you're trying to do. But to keep this on a basic level, let's do a file search and I'm going to drag in my YouTube because I downloaded the YouTube documentation. Let me just grab it. Because I downloaded the YouTube information earlier, that makes it just a little bit easier to do to have that. So we've attacked advanced, advanced options, chunk size tokens. We don't need to worry about any of that for a file that size. And then this agent is going to give me feedback on my top five performing videos. So let's preview what are my all time best performing performing videos. So it passed the guardrail test then if we look at what's happening right now, because I don't want to, I don't think I'm blocking your screen. Quiet. But I did move that over so that you could see it. If we just shrink it a little bit so you can see it and then zoom in, you'll be able to see that it's reasoning with itself. It's searching the file, it's compiling data, analyzing videos for views and reviewing the views for several videos. For example, Nana Banana's breakthrough Graphic motion will shock you has only 214 views, which isn't impressive. However, the Truth About AI agent I want to tell you has 3000 views views. There are also videos like the no BS guide to your first 1500 with just 77 views. That's a shame guys go and watch that video because that is a cracking video. I don't know how people like people that are out there making no money and don't want to learn how very very strange like and subscribe because I put that sort of content out there all the time and that is something that you really want to watch. So here, let me just change it back to normal screen so it feels like it's normal to you guys and let me zoom out here the output. Here are the all time 5 videos by lifetime views from your analytics report. This 6300, 38003600 and just under 2k for the forget chat GPT. 12 AI tools are about to take over ranked by watch, watch time or subscribers gained. For example the Gemini walkthrough leads by far on watch time 530 hours. I could go backwards and forwards with this and be able to ask it questions so it deep dives into the data. But I just wanted to show you how you would put a basic flow together and by putting a basic flow together and going through that you can feel free to expand this. These can get really detailed. There's no point in me trying to showcase a basic no code platform that go look at some of my NAN flows. Really technical. There is no point me trying to do that in here because it's meant to be simple and if you treat it how I've just treated it in terms of the build of what I've just done, you start off with something basic like that, upload some of your files, ask it some questions, start there and then manipulate the data after that. That will give you the best possible output so that you are comfortable using the system and then you can advance from there. My name is Michael Elliot. Like subscribe. I'll catch you in the next one.