Zero To Your First AI Agent In 26 Minutes (no code)

This is your quick start guide to build a fully functional and deployed AI agent today with zero lines of code. We're going to start off with the basics of what makes up an AI agent and then implement it using NAV. All of this is doable with no code. As per usual, it's not enough for you just to listen to me talk about stuff. So throughout this video, there's going to be little assessment which if you can answer this question, then you are well on your way to building your first AI agent. A portion of this video is sponsored by Lovable. Now, without further ado, let's

get going. All right, let's first start off with a crash course on what is an AI agent from a practical perspective. An AI agent is defined as a software system that uses AI to pursue goals and complete tasks on behalf of users. For example, a customer service AI agent would be able to take user queries and help them solve their problems. Or a sales assistant AI agent would be able to qualify leads, book meetings, and follow up with sales prospects. There are lots and lots of different types of AI agents, but each AI agent is made up of six core components. The first one is model. This is the brain that powers the AI agent. And this can be Chhatabi, it can be Claude, it can be Gemini, it can be small models, it can be big models. Next up, AI agents need tools to be able to perform their respective task. For example, a personal assistant AI agent would need to be able to have access to things like your calendar in order to book appointments. Then there's knowledge and memory. A therapy AI agent needs to remember the sessions that they had with a patient over multiple sessions. And a legal agent may need to have access to a knowledge base of specific cases that it's meant to be analyzing. Audio and speech. Many AI agents will have language capabilities to be able to more naturally communicate with humans. Guardrails are safety mechanisms to ensure proper behavior. You don't want your customer service AI agent to be swearing at people, for example. And finally, there's orchestration. These are systems to deploy and monitor and evaluate your AI agents. You don't want to just make your AI agent and release them into the wild and not care about what happens afterwards. You can have all these different components there, but if you don't know how to assemble them properly, then it's also not going to work out. You can give your AI agent like the best tools out there, but if you don't tell it that it has these tools and it doesn't know how to use it, then it's completely useless. That's why people spend a significant amount of time working on the prompts. All right, that is our little crash course today on the theory of building AI agents. If you do want to have a little bit more in-depth um explanations for things, I did make like a full video over here that you can check out and it goes into a lot more depth, but this is enough for us to build our first AI agent and we're going to implement all of these different components and the prompt

using NA10. But first, let's do this quick little assessment which I will put on screen now. Please answer these questions to make sure that you fully understand what it is that we just talked about. Okay, so this is NAN which is a flexible AI workflow automation tool and it's going to be what we're using to build our first AI agent. Okay, so after you

sign in, you can create a new workflow. What we're going to be building today is a hybrid AI research assistant and learning assistant. This is actually one of my favorite workflows. In my line of work, I have to learn things like a lot of different things really, really quickly and keep up with, you know, all the trends and things that are happening in the AI world. So, what I do is that I have this AI agent that collects all the information surrounding a specific topic, summarizes it, converts that into audio format, and I would actually listen to these condensed summaries to learn about a specific topic really, really quickly. I am very much an audio learner, so this works really well for me. And it's especially helpful if it's surrounding a topic where there's not a lot of like YouTube videos and courses and resources that's already been created on that topic. Okay, so coming back to NATO here, the first step we're going to do is we need something that triggers the entire workflow. So in this case, we want to create a form submission where the user is able to input the query that they want to search. So the title of this form, we can call it search form. description is input your search query to create an audio version of a specific topic to learn and the elements that we want in this form is topic so we can put a placeholder like I don't know like live coding for example and then we want to add another element so we can call that time period because we want the user to specify like what time period they want to be drawn find your resources from this can also be test. We can say like past 6 months something like that. We will make both these required fields as well and we will execute step to see if it works. So this is what it's going to look like. The search form that we have here the topic. So we can say like vibe coding and then time period is past 6 months submit. And we see that it was able to submit as a task. Great. So this is going to be what triggers it. And after the user um goes in and inputs what they want to submit, the next thing we want to do is have the AI agent. This is where we're going to start building the AI agent. All right. So with this AI agent here, the first thing I want to do, remember the first component of an AI agent is the model. So I'm going to connect a chat model. In this case, I'm just going to use OpenAI's Chat GBT. Let's do that. So here you can create a new credential. And it's super easy. It literally prompts you exactly what to do. Um you can ask the assistant as well. So how do I set up credentials for OpenAI? This is the NA10 assistant. So it'll tell you exactly how to do that. So we go here, sign in, go to the API keys, and create a new secret key. So NA10 project, create new secret key. Copy that, paste it over here, and there you go. That credential created. Wonderful. Next up, we're going to write the prompt here. And to do this, I'm actually going to go into chatbt and I'm going to copy paste this meta prompt here. So, this is a prompt where you can specify what your use case is and it would generate a prompt for your AI agent. Um, and I'll actually put this prompt in the description as well. So, you can use this to get started very quickly. You're basically telling CH2BT to create a complete self-contained NATO ready agent prompt for the following use case. So, this prompt is going to produce a good starting prompt for your AI agent. So, I'm going to say create a research/arning AI agent that takes the input of a specific user query and time period to search for the information to produce a summary about that topic. So that covers the role, the inputs and the task. We also want to add this summary will be translated to audio format at the end, but this agent will only create the text summary first, but make sure that it's optimized for audio. So that covers the role, the input, the task, and the output. So constraints is going to include make sure that sources are reputable sources and base the sources on as many primary sources as possible. So in terms of the tools that the agent will need in this case we'll use perplexity as the way to gather the information to produce the research surrounding that topic. So we need to tell it you have access to perplexity API in order to search up the information to produce the summary. This is good enough to get started for a lot of the other parts of this. This prompt should be able to take care of filling in most of the gaps. You will also store the information in just simple memory for now. So just have some storage of that information. So this is good enough to get started. you know, don't worry too much about it. For any additional information, this prompt will fill out most of it for you. So, press enter. All right. So, it has this prompt over here, which we will copy paste over here. So, here's the prompt and there are a few small tweaks that we do want to make. So, click expression in order to allow you to use variables. So, for example, here you have the research topic and it just shows like research topic, right? But here what you can do is actually go to the schema of the input from the previous node which is the form and you can drag this variable. So to the variable that the user submits. So this is going to be the topic and then on the time span it has time window and we can just replace this with the time period from the user in the form. This has other stuff like word limit, audience regions and things like that and the focus. Yeah, we can just leave that because you don't have that information provided here as well. interpret input and normalize time. So we'll just very quickly do the same thing. So just drag the topic and the time window is going to be the time period. Great. So we're searching with perplexity blah blah. You know, we will fix this in case it's not good, but for now this is fine. One more here. I'm just changing the topic in a time window because I know that these are things that have been already submitted by the user. So might as well include them. Okay, great. Now, before we can actually execute this step, we need to provide it with the memory and the tools that we said we're going to provide. So, starting off with tools. So, under tool, we're going to give it the perplexity tool. Super easy. All you have to do is search it up on NAN and then it will show you the tool over here. So, for the credentials, very simple. You can also just click create new credentials. You can ask the assistant um for the exact way of setting this up. It's very, very similar to OpenAI. So, in the interest of time, I'm just going to use the one that I already set up here. The operation is going to be message a model. Model that we want to use from perplexity is the sonar model. Um and the text that we want to get. So what are we going to actually input into the model? Right? We would actually like to click here that says let the model define this parameter. And in terms of simplifying the output, we're also going to let the model define this as well. Great. So now we have this tool set up. And in terms of memory, let's include a simple memory that can just store the specific sessions in here. Again, like all of these things we can change like the important thing is just to get the components there first and then we can optimize it later. Okay. So, and for here the session ID, we'll just write define below and we'll just call it summary. So, we're just giving a variable name to save information as. All right. Now, we have the moment of truth and let's actually try running this AI agent. What it should produce is a summary. Okay, node executed have a bunch of check marks. So that is a good sign. Let's actually see what happened. Okay, if you go over here and look at the output. So it looks like we do have an output. Okay, this is looking promising. So vibe coding the practice of guiding to write code. What it is vibe coding blah blah. So findings 3 to seven items. Okay, so we might want to change the format of this a little bit. But it looks like we do have the output here. So that is good. So if you click on the logs, you can actually see what exactly the AI agent was doing. So if over here the AI agent came first is the simple memory and it inputed the prompt that it's executing. So it started with that. Then it went to the open AI chat model, gave it the prompt that's over here. The model decided that it was going to message the model in perplexity to do the research and use the perplexity tool in order to gather the information that's there. Then that information is passed back to the OpenAI model where it compiled everything together into a summary and then stored it again into simple memory. Yeah. So this is a great way to just see like what your agent is actually doing just to make sure. You can also look at um perplexity if you're like being paranoid like I am to see okay like it actually does have the content that is coming through. All that information is there. Great. And then you can also look into simple memory to double check as well. Oh, look. It looks like it did. It did save all this information into simple memory as well in the chat history. Wonderful. Great. So, at this point, what you want to do is actually click save because if you don't click save, then you're going to lose your entire workflow and feel very sad. So, this is great. Now, we have a summary that's here. It's not perfect, but it's pretty good. So, the agent itself has done its job. Wonderful. But, I do want to have this translated to audio format. So, what I'm going to do is add another node here and call it. This is going to be another OpenAI node. And under OpenAI, there's a lot of different actions. And one of the actions that you can take is generating audio. So, I'm going to use the same credentials that I had from OpenAI audio source. And the text input that I want to generate here is going to be the output from the AI agent. So, I'm going to drag the output variable that is here. Now I'm going to execute this step to see if it actually works. You always want to execute things one step at a time by the way. So you're able to catch any errors. And it looks like node is executed successfully. Let us see. Ooh, stall it. Vibe coding. The practice of guiding AI to write code has surged in popularity and capability over the past 6 months. This rise is reshaping how developers work and how software companies view AI assisted development. Now what it is vibe coding involves using artificial intelligence models to generate, explain, test and refactor code based on user prompts. Create million dollars in cash signaling strong market confidence. Source TechCrunch June 2025. Hands-on experiments show Vibe coding can quickly produce usable code for production features when combined with human validation, although results vary by domain and data quality. source YouTube 2025. Okay, so this is our first try and it's honestly not bad, right? We managed to get it to work with just the initial prompt that we had. Obviously, there's a few things that we do want to clean up here. Like for example, you don't want to have the citations like embedded into the audio and you don't need to like say like what the title is perhaps and maybe you know there's like little things that we can tweak. And this is what we would do with the prompt. I changed the prompt in order to get the format, to get the summary to be the way that we want it to be. And but that is really not bad at all. Now, to finish off this workflow, it would be a pain in the ass if I just had to like go and download it every time, right? So, what I'm actually going to do is I'm just going to ask it to email it to me through my email. So, we can add another item. We can call it email. Wonderful. Gmail. And it has, let's see, let's see. Send a message. Wonderful. Same thing over here. You can create a new credential in Gmail. Super easy. You can sign in with Google and then it would directly allow you to connect it um with NA10. All right. So the sort resource here is going to be the message. The operation is that we want to do is going to be send. So the email I'm going to send is so tina lonely octopus. com topic summary. Email type html is fine. And the message I'm just going to say here's the audio file. Under options we can do attachments. So the audio file that's here, we can use that as the attachment. And let's now try executing this step. And now let's actually check our email. Wow, amazing. Look at that topic summary. And it has the audio file here. It's sent as an attachment — title by coding summary for past 6 months. Audience generate. — If you want to take your AI agent to the next level and build a customized and aesthetic web app without even needing to write a single line of code, you should check out Lovable. Lovable lets you build full stack apps just by describing what you want. You can take your AI agent/AI app into a functioning working product with a backend, a front end, and a database all integrated for you. It works with tools that you're already using like N8 end. And it also has built-in integrations with other tools like Superbase and Stripe. You get clean editable code that you can edit within Lovable itself or you can export it to wherever you want. There is a free tier with five build credits per day that you can get started for free. You can also use my code TINA20YT in the next 30 days to get 20% off your first purchase of the Lovable Pro plan. The link is in the description. Thank you so much Lovable for sponsoring this portion of the video. Now, back to the video.

And yeah, there you go. Pretty cool, right? Okay, so this is a functioning workflow at this point, right? And I think a lot of people on YouTube at this point will be like, "Yay, wonderful, amazing, great. " You know, and maybe they'll just tell you to, oh, all you have to do is deploy it now and then you're good to go. But, but remember the six components of these AI agents. We have not done all six components yet. We do have the model, we have the tool, we have the memory. We decided to have audio and speech functionality. So there's two more. Pop quiz right in the description. What two things are we missing right now? Yes, guard rails and orchestration. These are the two things that people always skip out on. And then when they actually deploy their workflow and actually use it in real life, they're going to end up with a lot of problems because they don't have these components to make sure that their AI agent is functioning properly and doing what it's supposed to be doing in the long run. So that's why I'm actually going to add in these two components now. So, when it comes to guardrails, the two minimum things that you should think about doesn't contain things like foul language, abuse, you know, like racist stuff. Um, it probably won't because it's coming directly from perplexity, but you can imagine if we're combining a lot of different sources together and not going through something like perplexity, you do need to screen for these kind of things. So, we need to make sure we have something in place for that. And the second component that we want to make sure of is some sort of error handling ability. So, what happens if it comes up with something and perplexity fails for some reason, right? and it doesn't have the information it's supposed to have. You don't want your entire workflow to just break. So you need to come up with error handling to think about anticipate these cases when it happens. What should your workflow do? So let's actually implement these first. Let's put in a mechanism to make sure that the summary that's coming out is not containing bad languages. And we want this to be done right after the AI agent. So we're going to add another node here. This is also going to be from OpenAI and they have a action that is classify text for violations. How convenient, right? So yeah, using the same OpenAI credentials and it's just going to make sure it doesn't violate any standard safeguards. So we will drag the output variable here from the AI agent which is the summary and we just want to make sure that it doesn't violate anything. And let's execute the step and see what it looks like. So it says here that it's flagged as false. So and it's flagging for these different categories like sexual hate, harassment, self harm, sexual such minors, etc., etc. And it actually gives a score for all of this as well. And just to test this out, like say for example, we're going to write here like I don't know, I hate you. You suck. Technically, you should flag. Yes, it flagged as true. And the category of flagged that is harassment. Yeah, that's not good. So, we do know that this works. Wonderful. Let's put the output back. Now, what we need to think about is say like if there is no um flag and there was no issue with the summary, we probably want to just go through with this entire workflow. But what if there is? Well, there's a lot of things that you can do, right? Maybe you wanted to redo it again, like ask the AI agent to redo it again. You can ask it to send a warning message. You can go ahead and still do it, but then just have like a flag when you're sending the email saying, "Hey, there's a violation within the flag. " And maybe in the body of the email, write that, oh, like um here is what it got flagged for, just FYI. So, there's a lot of different ways that you can deal with this, and there's no right or wrong answer. It's about how you want this AI agent to behave. So in this case, what I want is if it does classify something as a violation, I want it just to directly cut this workflow and just send a warning message. So to do this, I'm going to add another node after the violations one. It's called a switch node. And what we want it to do, so if the flag value is equal to false, then we want it to continue on the workflow. While if the flag value is equal to true, then we want it to do something else. and just toggle convert type when required just to make sure that these errors disappear. All right, so we have if it's false it would continue on and if it's true we want to add another node that is still going to be like a email node and just send a message summary error. There was a text violation flag please check workflow for details. All right. Now, to test to see if this actually works, what we're going to do is over here. I hate you. This should flag as harassment. And if we click the switch. Okay, maybe that didn't work. Let me try that. Let's try again. So, detect input here. Um, we can do something like you are terrible. I hate you. Bad. Execute step. This is flagged as true. So in the node it should also go here and it should have sent an email. Now final last component is

orchestration. So this includes things like deployment, includes things like monitoring, evaluating things and improving the agent over time. So the easiest thing that you can do is just to deploy it and hope that it keeps performing the way that you want it to be performing. But for most production ready workflows, you do want to include something called evaluations. And this is where you have a lot of different test cases that you want to run through your agent. So you're able to see the agents response to all the different test cases. And depending on um what the results are, you can choose to change your prompt and tweak it so that you can keep improving the results of this. Here's a saying that what doesn't get measured doesn't get done. So only by measuring your agents behavior will you be able to improve that behavior over time. By the way, if you do want to know like more details about evaluations, things like that, I do have a video that I'll link over here that does dive deeper into this. But uh for this video, I'm just going to show you how to do that. Okay. So, here is the evaluation spreadsheet that we're going to input. And here we have different topics like climate change, AI agents, elephants, carrots, um and different time periods that we're going to test out. And the way that we're going to pass this through NA10. Uh we're going to come here to NA10 first. And the first thing we're going to do is actually add another trigger node. And it's going to be called when running evaluation. This is the evaluation trigger. And you want to connect the Google Sheets, which is the one that we have over here. You can have it by creating new credential. I already have it linked over here, but it's super easy to link to your Google Sheets. I just go through the authorization and then from the data set, choose evaluations and you want to choose sheet one. Great. Now, next up, we want to add a node that is literally called the do nothing node. So, this one is really just for like aesthetics kind of practical purposes that you can connect two different triggers to the to this node going to the agent. Then coming over here, we have this branch that is going to be classifying the text violations, generating audio, etc., right? That we already have. But we want to get another branch that's able to evaluate all the test criteria. So we want to add another do nothing node and then add another node, the evaluation node. So this one we want to be the set output node, so we're able to get the outputs and capture the outputs. So again, we're going to connect that to the Google Sheets and we're going to choose evaluations and choose sheet one. And now we're going to execute previous nodes. Add the name. We can just call it output. And we want to add the value that is coming in over here. And this will allow it to actually write um the output on this column here. And finally, we want to add another evaluation node. This is the set metrics. There's a lot of different types of metrics like correctness, how correct it is, how helpful it is, how good this string similarity is, how it's categorized. You can define your custom metrics as well here to evaluate your tests. In this case, I'm just going to pick the helpfulness one. Super simple one. It comes with a prompt that tells the model is an AI model to um act as an expert evaluator that assesses the helpfulness of the responses and gives you a score from 1 to five which we can capture. The model that we're going to use is the OpenAI model. Again, I just connect that GPD 4. 1 mini. It's good. And configure this. But the user query as query is fine. execute this step and we can see that it gives us a helpful score of five. All right, let's clean this up a little bit. Tidy up workflow and let's actually try running this. So to run this, we can click save here, go to evaluations and we can run a new test five. We can click into this. We saw that there are four total cases and each of these different cases has passed. We can also see over here that it wrote down the output for the information that's here. So you can try this out and you can see that there are have additional use cases that you add here to test this out with. Um right now we see that the helpful scores have all been pretty high with this one is the lowest it's out of four. You can also add obviously like other types of evaluations like some other things that I would recommend adding would be some sort of metric that will allow you to see if there's like certain keywords that are being contained within the summary. You might also want to test like the overall um structure of it, overall length of it, a lot of different types of evaluations that you can do. Okay, so for this simple example now that we do have the complete workflow. Okay, so the next thing we're going to do is to deploy it. So to do that, it is

really easy on any all you have to do is go toggle this from inactive to active. And then to actually see it, go to on form submission and here we have the test URL. Just toggle this to production URL. Copy this and there you go. Amazing. So let's just try something out. Say it's called like building AI agents time period. Let's say it's 2 months. Submit. There you go. Here is the

summary. So moment of truth — title building AI agent summary for past two months. An AI agent is a software system that autonomously performs tasks by combining artificial intelligence with tool use and data access. Building AI agents involves designing workflows, managing security, and ensuring ongoing monitoring and updates. Key findings bulleted three to seven items. OpenAI's 2025 releases include APIs and SDKs, simplifying agent workflows, integrating tools like web search, and enhancing observability for production reliability. source open AAI 2020 20 — is a fully built and deployed AI agent that has all the six different components and the prompt done. Of course, there are some tweaks to the prompts that you want to do and based upon your evaluations, you might want to go back and tweak the prompt even more to be able to come up with your perfect AI agent. But in our goal of getting an AI agent up and running, we have done it. So, at this point, there are a lot of other things that you can do to improve this AI agent. Like for example, this form that we have to submit your topic is not very aesthetically pleasing that you can use a vibe coding tool like lovable for example to create a more aesthetically pleasing UI like this. Similarly, the workflow right now just sends an email, right? But instead you can vibe code using lovable, a UI component that allows you to create the summary, create the audio file and actually just download it directly from the UI as opposed to just having it sent to your email. You can also add other components to this as well like a dashboard for example that showcases all the different summaries that you've generated. many other things that you can do. Now that you've built your first complete AI agent, I hope this was a helpful video for you. I have a final

little assessment. Please answer these questions on screen to make sure you've retained all this information that we have covered. And let me know in the comments what AI agent that you want to build yourself. Now, thank you so much for watching until the end of this video. And best of luck building your first AI agent. I will see you guys in the next video or live stream.