The Effective Way To Build An AI Agent In 2026

0:00I just spent the past couple of weeks building an AI agent into the app that I'm building. And what I thought was originally only going to take one week to build out actually ended up taking me like an entire month to build. And even right now, I'm still actively improving it because it's not perfect. And I was able to get that first version working within the first week, but the results were just really bad. And then I spent the next couple of weeks creating a completely brand new architecture and implementation for how the agent should work. And in this video, I wanted to talk about some of the big changes that

0:27I made from building a really bad agent to and building an actually pretty effective agent and useful agent. For full context, I'm not an expert. I'm just a random dude on the internet that likes to build way too many apps and is currently having a little bit of success with the current app that he is building. So, just take what I say with a grain of salt. I'm not an expert, nor am I trying to pretend to be one. I'm just sharing what I've implemented and what I find working for me, and maybe you'll find it useful for your end as well. So, first let's get into the big

0:51mistake that I made when building out my first agent. All right, so the first thing that I want to talk about is I don't think that what framework you use to build your agent matters. Like there's so many agentic building frameworks like there's this one, Eve, that's released by Verscell. But then, funny enough, I actually just built my agent using Verscell's AI SDK. I didn't find any issues with it. I found it pretty good. I also know that Claude has their agent SDK. I think Open AI has an equivalent one as well. And honestly, I think this is just over cooking. It's so unnecessary. At the end of the day, I

1:20think an agent is just a chatbased experience that has tool calling enabled to take actions for you or for your users. That's it. And any modern chatbased framework, chat framework, whatever you want to use, has support for tool calling. So, really, I don't think it's that complicated. So, use whatever framework you want. I built out my agent using the Versell AI SDK and

1:40I'm having no issues with it right now. Now, let's look at some of the big differences and some of the big mistakes that I made when I built out my agent inside of my application. For me, the big issue that I found was that I built out my agentic experience by treating the model as if it was a 2024 era model, not a 2026 era model. What do I mean by that? Well, 2024 models compared to 2026 are significantly less smart. They don't think or reason nearly as much. And I think the biggest one is that they required really explicit instructions about what to do and what actions to

2:09take in certain scenarios. Basically, they did not excel whenever they faced any type of ambiguous situations. they would come to the wrong answer or do the wrong action or do nothing at all. Now, compare that to 2026 models. They are incredibly smart. I dare I say, they're borderline creative as well. And most importantly, they're smart enough to make judgment calls on their own and handle ambiguous situations and still arrive at the correct answer or the desired output. Now, what exactly does that mean tactically in terms of building out an agent? Let's get into that. So, I'm going to talk about some of these more tactical things by

2:41actually walking you through the agent that I built. And just for a little bit of context, the app that I'm building is called Yorby. It's an AI social media marketing tool primarily right now focusing on helping you generate AI social media content to market your startup on social media. My co-founder and I have a lot of experience doing social media both as individual creators as well as running a social media marketing agency. So, we're very opinionated about how to make good content and we kind of dumped all that into Yori and right now we are working on a brand new agent to build out effective carousels to market your

3:08product on social media that gets a lot of views. So this is the agent that we ended up building and and I'll walk you through this agentic experience about how we created this carousel at the very end. And this is how the carousel agent currently works right now that we built out. Essentially, you can pick any source carousel from whatever post that you want to remix. We have a database of carousels that we've been scraping and collecting that you can just remix with the one click of a button. It creates a plan. It asks clarifying questions, shows you the plan, you can revise it back and forth, and then it spits out a

3:36carousel at the very end. And then the final generated carousel still maintains like that original viral format of the original source carousel, but then it's just remixed both from like the photo perspective as well as a text content perspective to fit your brand for your startup. Instead, that's a quick overview of the agent that we built. Let's get a little tactical about what we did to actually turn this into a actually useful agent using all the learnings that we went through right here about using the models like it is a 2026 model and not like a 2024 model. So when I first implemented this carousel, what we wanted to do, what I tried to

4:07do, I don't have it available anymore just cuz it's like a month, two months old at this point. But I tried to oneshot it with a user just saying, "Hey, this is my product. Remix the entire thing." And then I tried to prompt it in the background with a system prompt with super explicit instructions and how to make certain judgment calls on what to do to make this effective carousel. But honestly,

4:25no matter what, it just was not great. The output was not great. And the main reason for that was because the input wasn't good. it didn't have enough context about what it wanted to do. And I think you can even see this in your behavior in terms of how you use AI. I know I'm guilty of this as well, but I people chronically underprompt whatever they want their models to do. They don't give enough information because humans are lazy. So then a big change that I made in this agent experience to yield better outputs is really really emphasize on input gathering, context gathering. So how do we do that? Well,

4:57we have this plan mode. you know, you see all the tools like cursor, codeex, cloud, all the AI coding tools have a plan mode that users can then review the plan and add more text input and more feedback about what they want changed. We also have like ask clarifying questions tool. I think this is the greatest tool that has ever been created. I believe Claude first created

5:15it with Claude code and it's incredible. And this is basically invoked whenever the LLM has any ambiguity or doesn't have full confidence on certain pieces of information or in certain context. And it prompts the user to provide more context until that model builds up more confidence to take that action on behalf of that user. And then you can see we have this tool right here as well, carousel context ready. This is a tool that actually does like fetches the original source carousel. It analyzes all of the photos there and all of the text there. And then it also looks into the user's reference photos that it has

5:47generated to see, hey, are there any relevant photos that we can use from the user's photo library to actually create this remix prompt or this remix carousel? And actually, if you look over in the actual code for this and you look at all the tools that we have, you can see so much of this is really just about showing the user information, getting feedback from the user, and then iterating back and forth, back and forth until it has so much context about what the user truly wants the final output to be. Then it finally executes that action. Like if we go over here, carousel clarification tool asking for

6:16more clarification from the user. Draft planning tool, draft tool, editing, photo resources, fetching additional photo resources, planning base, creating the base context from the original source carousel that the user is trying to remix. Remix plan tool. That's the plan mode that we just showed you right here. Let's see what the plan is that the LLM has cooked up for any text changes that the LLM is going to make as well as any visual changes that the LLM is going to make, allowing the user to review, edit, review, edit before finally giving it the green light to create. And at the very end, the carousel show remix tool, media library

6:48tool. This is the tool that actually fetch the reference photos from the user's library. And then there's really only one final tool right here, the text on screen tool, which is the single tool that handles generating the final carousel based off of all the input that the user provides. So really, I have found in my usage both personally with AI in AI coding as well as other AI apps

7:07that I've built, input matters so much. You cannot outprompt bad input. So I think if you're trying to build an agent to execute some type of action, you need to really really invest into the tooling available to gather that information to execute that end goal. And I guess kind of building on top of that, just to go one level higher tactically of what this would look like, is you can have an opinionated system prompt of what should be done, but the tools should be pretty general purposes. It should really just be like CRUD operations like reading, updating, deleting, creating. That's

7:36just what we do with all of our tools. And the reason for that is because once again these models are getting intelligent enough, able to handle ambiguous situations enough and make judgment calls such that if you just provide these general purpose tools for the models to gather the information and then execute and take action on that information, it can kind of stitch together that plan, especially if you provide it with really rich instructions and really rich context from the user's behalf. And lastly, I think something really tactical that you can really do to build a really great agentic experience is to just look at what the

8:04AI coding tools are constantly adding as features in their apps. I use chat GPT and cursor a lot. And like I mentioned earlier, this ask clarifying questions tool that we have right here, this came directly from the AI coding tools because they've been asking the clarifying questions for a very, very long time. And I think moving forward always like these AI coding tools, they're always going to be the bleeding edge of what is possible of what can be done with agents. So really, when you're building an agent experience, you can just rip features and copy features from these existing AI code editors and coding agents. There's no need for you

8:33to reinvent the wheel because you can already see what's working from the bigger labs and just import it into your app or your own agents that you're building out. So, as a quick summary, the changes that I made to make a pretty effective agent are number one, not caring about what agent framework to use. Use whatever you want. Agents are really simple. It's just a chatbased experience with tool calling available

8:51to it. You can do that in any framework. Number two, the models are really smart. So, you just need to provide general purpose tools that let the model read data, update data, and create stuff. And I think the vast majority of your workload like 80% of the work that the agent should be doing and that you should be investing in to the user experience of is gathering information and gathering context from to provide it with really high quality input to execute whatever output that it's trying to generate because I personally believe that you cannot outprompt really bad input data. So the input data that's going in is really really important and

9:22the input data that you put in is going to overwhelmingly influence the quality of the output that the agent is going to have. It's a quick little video about how I think you should build an effective agent in 2026. Let me know if you have any questions in the comments down below. Going to try to respond to as many of them as I can, especially in the first couple of days. Do you agree with me? Do you disagree with me? You think I'm a stupid? Do you think I'm a trash developer? Let me know in the comments down below. That's all I got

9:44for today. I'll see you in the next one. See you.