Master ChatGPT Agent Builder Before It's Too Late: Dev Day Breakdown + Full Tutorial

While you were sleeping, OpenAI just launched a new app store inside Chat GPT. And guess what? It has 800 million active users on day one. At dev day, Sam Alman revealed agent builder, a tool that could replace tools like N8N or Zapier. Codeex, which makes coding 10 times faster. And with apps like Spotify and Figma built right into ChatGpt, it's now a full app platform. In this video, I'm breaking down all the big announcements you might have missed. Plus, I'll show you how to use OpenAI's new agent builder to build your own AI agent from scratch. No coding needed. Watch till the end or you'll miss out on what could be our generation's biggest opportunity. Let's dive in. Okay, so

here's where it gets wild. Forget everything you know about app stores. Apps now just live inside Chat GPT. Now look, OpenAI tried this before with the GPT store. Remember that this is completely different. Apps aren't in some separate store anymore. They're right there in your conversation. Let me show you what I mean. You're in chat GPT, right? You type Spotify, make me a playlist for my party. Boom. Spotify just appears right there. Interactive, personalized, instant, no switching apps, no copying links, nothing. Or you're like, Corsera, teach me about machine learning. And suddenly you've got courses, videos playing right there in the chat while you keep talking to chat GPT. The apps SDK is built on this

thing called MCP, model context protocol. Basically means you get full control. Your back end, your data, your UI design, all of it. But here's the actually crazy part. They have this feature called talking to apps. Chat GPT can literally see what you're looking at. You don't have to explain context. So you're watching a Corsera video and something doesn't make sense. Just ask Chat GPT right then and there. It sees the timestamp, knows exactly what's being discussed, explains it to you. The

launch partners that are already live include Figma. You can literally sketch something and have it turn into a working diagram. Using Canva, you can start creating posters, slide decks just from conversation. Okay, this one's actually wild. You're chatting and say, "Show me threebedroom homes in Pittsburgh under 500,000 with a yard. " Boom. Map appears. Full Zillow experience inside chat GPD. You see a house you like, click it, goes full screen with all the details. Then here's where it gets insane. You can ask chat GPT, "How close is this house to a dog park? " And it pulls that context from the Zillow map you're looking at and searches for you. Here's the business

side. Monetization is coming. They announced this agentic commerce protocol, which is just fancy talk for instant checkout inside ChatGpt. You can sell directly to 800 million people. If you're a developer, you need to get what just happened here. This is legitimately 2008. Again, back then, the App Store gave you maybe 500 million iPhone users over time. The app's SDK gives you 800 million Chat GPT users. Right now, today, now it's in preview, so not everyone can build yet. But later this year, when the directory launches, that's when it's open season. And for everyone watching who's not a developer, this means your Chat GPT experience is about to get 100 times more useful. Everything you used to have to switch apps for, shopping, learning, planning trips, finding homes, it's all going to be right there in one conversation. The gold rush literally just started. All

right, next thing. Agent kit. So, until now, if you wanted to build an AI agent, it was a nightmare. OpenAI just said, "Forget all that and build everything in one place. " It's called agent kit. Think of it like a complete toolbox for building AI agents. Everything you need is right there. Here's what's inside. Agent Builder. This is the visual part. You literally just drag boxes around and connect them like you're drawing a flowchart. No coding needed. If you've used those workflow tools before, same idea, but this one's made specifically for AI. Chatkit. This is your chat window. You know how every AI agent needs that chat interface where people type stuff? This is that. Just drop it into your app. Make it look however you want. Your colors, your style. Done. Evils. Okay. So, this is the testing part. And this is huge because with regular automation tools, you can't really test if your AI is doing a good job. This lets you see exactly where your agent messes up, then fix it automatically. That stuff just doesn't exist in normal workflow tools. Connector registry, fancy name for connect your data securely, link it to your company's stuff, third party apps, whatever. It handles all the security headaches for you. So basically, OpenAI looked at tools like Zapier and said, "What if we made this specifically for AI with all the AI specific stuff already built in? " Now, here's the part that blew my mind. This engineer at OpenAI, Christina, built a complete working agent on stage in front of everyone, 8 minutes total. She made an agent that figures out what kind of question you're asking, an agent that searches through files to find answers, an agent that plans out schedules, nicel looking result cards, security stuff to protect private information. 8 minutes. That's faster than most people make breakfast. And real companies are already using this stuff. HubSpot, they're using Agent Kit for their customer service AI. This is basically what website builders did for making websites. You used to need to know code. Now you just drag and drop. Agent Kit is doing that for AI agents. All right. So

you want to see how this agent builder thing actually works? Let me walk you through it. First step, getting to agent builder. Super simple. Just open Chrome, type agent builder open AI and you'll see the announcement page introducing agent kit. That's the one. Scroll down a bit and you'll see agent builder with this nice interface preview. Click on that. Now it shows you all the documentation how to use the platform, all that stuff. But let's just jump straight into the platform itself. And boom, this is what you get. Clean interface says create a workflow. Basically build a chat agent with custom logic and tools. Let's create one from scratch. I'll hit create. Okay, so this is the interface. Let me break down what you're looking at. On the left, you've got all your building blocks. Agent nodes, your actual AI agents, MCP, model, context, protocol, connections, guardrails, security stuff, PII protection, file search, that's basically R A, searching through documents, if else, conditional logic, branching paths, while loops for repetitive tasks, user approval, this one's cool, makes the agent ask you before doing anything. So if you want human in the loop, you just drop this in and the agent will pause and ask, "Hey, should I do this before executing? " Pretty straightforward, right? All right. So what are we building? I wanted

to make something fun. How about an agent that suggests music based on your mood? Like if you're having a rough day, it suggests chill music. If you're pumped up, it gives you energetic stuff. If you're stressed, maybe some meditation tracks. Let's do it. See this

default agent node? I'm going to rename it. Let's call it classifier agent. Now the instructions. This is what tells the agent what to do. You are a helpful assistant. Your goal is to understand the user's mood and based on that classify if they are happy, sad, or stressed and suggest a nice playlist for them. Simple enough. Now, here's where it gets specific. I'm going to set the output format to JSON so the next step knows exactly what mood we detected. Click add schema, then add properties. Property one, happy. When the user is happy, this is valid. Property two, sad. When the user is sad, this is valid. Property three, stressed. When the user is stressed, this is valid. Hit update and save. Now, our classifier agent knows exactly how to structure its

response. Next up, we need branching logic. Drag over the if/ else node and connect it to the classifier agent. Now, we set up the conditions. If the mood is happy, go this path. Else sad, stressed, go this path. Perfect. Now we've got three different paths depending on what mood the user is in. All right. Now the fun part. We're going

to create three different DJ agents. First one, happy DJ agent. Name happy DJ. Instructions. You are a mood DJ. Your job is to suggest a nice playlist on Spotify when the user's mood is happy. Make sure you choose the best and most liked playlist. Now this needs to actually search Spotify, right? So we add a tool web search. Click that and configure it. Search only in spotify. com. Users location India Bengaluru. Hit add. Same process for the sad DJ agent. Name it sad DJ. Instructions are basically the same. Just change happy to sad. Add web search again. Spotify. com. Same location settings. Last one. Stress DJ agent. Yeah, the names are kind of funny but whatever they work. Same instructions just for stressed mood. Add web search spotify. com same location. Done. Now we need to close this workflow. Grab the end node and connect all three DJ agents to it. Beautiful. We've got start classifier agent if else logic. Three DJ agents and end. That's a complete workflow. Let's publish this thing. I'll

name it mood DJ. If you want, you can add chatkit here to embed this in your own website with your domain. Pretty cool. Hit publish. All right. Moment of truth. Let's test it. Click preview to open the chat interface. I'll type, "Hey, I'm having a productive day at work. " To make it more productive, suggest me a nice playlist that can keep me going. Watch what happens. The classifier agent reads this, figures out I'm in a good mood, classifies it as happy. The if else node sees happy and roots it to the Happy DJ agent. Happy DJ uses web search on Spotify, finds playlists, and boom, it suggests Loi Beats by Spotify, a super popular instrumental playlist for productive work, even gives me the link, click it, opens right up in Spotify. Perfect. So, that's it. That's how you build an actual working agent in Agent Builder. We went from zero to a fully functional mood-based music recommendation agent in what, like 10 minutes. And this is just a simple example. You can build way more complex stuff with file search guardrails, multiple agents working together. Pretty wild, right? All right.

So, we've got our Mood DJ working, but let's make it even better. Right now, it just gives us text responses with links, but what if we could make it more interactive, like have actual buttons you can click? That's where widgets come in. Let me show you how to add that. So, I'm going to click on the Happy DJ agent here. Opens up the side panel. See where it says output format and it's set to text. Let's change that to widget. Boom. Now it asks me to add a widget. I can either upload one or create a new one in widget studio. Let's create one. I'll click create. Okay, this opens up the

widget builder in Chatkit Studio. Pretty cool interface. Now I just need to tell it what kind of widget I want. So I'll type build a widget that helps me with the list of Spotify playlists that I should hear. Hit enter and watch this. It's generating the widget right in front of me. And look at this. It's created this nice interactive list with playlist names like deep focus, lowfi workday, morning run, Sunday chill. Each one has its own card. Looks clean, super clickable. Perfect. This is exactly what we need. Now, I'm going to download this. Hit the download button. It saves as a JSON file. Spotify playlist. widget. All right, back to the workflow. Now, I need to actually add this widget to the Happy DJ agent. Click on it again. Go to output format widget. And this time I'll click upload. Let me grab that file. Here it is. Upload it. And nice. It's now listed as Spotify playlist in the widget options. Select that. And we're good. But wait, I have three DJ agents, right? Happy, sad, and stressed. They all need this widget. So, let me do the same thing for sad DJ. Click on it. Change output format to widget. Upload. Select the same file. Apply it. Done. Now, stressed DJ. Same process. Widget. Upload. Apply. Now, let's test this thing and see how it looks. I'll click preview to open the chat. This time, let me try a different prompt. I'll type, "Hey, I'm working out and need some great music to pump me up and also refresh me. " Send it. Watch what happens. The classifier agent kicks in, reads my message, figures out I'm pumped up. So, it classifies my mood as happy, roots it to Happy DJ. Happy DJ searches Spotify. And look at this output. It says pumped and refreshed coming right up. And then shows me this beautiful widget with multiple playlists, beast mode, chill hits, each one with its own open button. This is so much better than just text links. So now our Mood DJ

doesn't just suggest playlists. It gives you an interactive, clickable interface. One click and you're listening. Way more userfriendly, way more professional. And the best part, we built this entire thing. classifier logic, three specialized agents, custom widgets in like what 15 minutes, no coding. Just drag, drop, configure, and publish. This is the power of agent builder. You can take this concept and build way more complex stuff. Add guardrails, file search, multiple approval steps, whatever you need. The possibilities are pretty wild. All right, that's it. That's how you build a complete agent workflow with custom widgets. Pretty

cool, right? Okay, Codeex, let's talk about this. Earlier this year, OpenAI launched Codeex in preview mode. It's their AI software engineering agent that actually writes code for you. Today, Codeex is officially generally available, meaning everyone can use it. Now, it runs on GPT5 Codeex, which is GPT5 specifically trained for coding work. What makes GPT5 Codeex different? It adjusts thinking time based on how complex the task is. Trained on actual real world engineering work, can work independently for over 7 hours straight. optimized for refactoring and code reviews. The numbers on this are genuinely crazy. Sam Alman mentioned at a launch event that within OpenAI itself, almost every engineer now uses Codeex. It was only half of them back in July. They're merging 70% more pull requests every week. Codex automatically reviews basically every PR. It's catching hundreds of bugs every single day before humans even see the code. Daily usage has grown 10 times since early August. Don't just take OpenAI's word for it. Here's who's actually using this in production Cisco. Their code

review times are cut in half, 50% faster. Engineers spend way less time on manual checks, more time on actual transformative work. Instacart, they integrated the Codex SDK into their platform. Engineers can spin up environments and complete full tasks with literally one click. Codex even cleans up technical debt automatically. In the live demo at Devday, this guy Roma built a camera control interface from just a sketch live on stage. Codex suggested the VISCA protocol, read the old documentation, wired up an Xbox controller as a joystick, then added voice control using the real-time API. The ending was nuts. He used the Codeex SDK inside the voice agent to reprogram the app in real time. Created movie credits rolling on the fly. Zero code written by hand, just talking and codeex executing. Software engineering just fundamentally changed. Quick model

update GPT5 Pro is in the API now. This is the smartest model OpenAI's ever shipped. Available right now to everyone. Perfect for stuff where you need absolute precision. Legal work, financial modeling, healthcare, complex reasoning chains. Next, voice AI just

became default infrastructure. Open AI launched GPT Realtime Mini. Smaller, cheaper version of their advanced voice model. Same voice quality, 70% cheaper, natural speech to speech. No weird robotic delays, just real conversations. This is basically the end of voice AI being special. Pretty soon, not having voice in your app is going to feel like not having a search bar. Customer support, productivity tools, cars, education, healthcare, voice everywhere. And at 70% less cost, every single developer can now afford to experiment with it. Last big announcement, Sora 2

in the API. OpenAI actually dropped Sora 2 a week before dev day, but they went hard showcasing it at the event. And they're calling this the GPT 3. 5 moment for video, meaning it's the first time AI video is actually good enough for real work, not just demos. Mattel, the toy company, they used Sora 2 API to turn sketches into video concepts in minutes. Designer uploads a sketch of a new toy, adds a description. Sora 2 generates a video showing how the toy moves, how it sounds when kids play with it, how it feels in action. That's how you validate ideas 100 times faster. No prototype building, no video production, no sound design, just sketch to video. It's an API preview. You get full control over video length, aspect ratio, resolution remixing, and variations. You can take a still image and expand it into video or start with one frame and let Sora extend the whole scene. Sam

Alman closed dev day with this line. Software used to take months or years to build. You saw today it takes minutes now. You don't need a huge team. You don't need massive infrastructure. You just need a good idea. This is literally 2008 happening again, but way faster and way bigger. The app store created Instagram, Uber, WhatsApp. What's the Chat GPT apps SDK going to create? The developers who start building now won't just ride this wave. They're going to define what this whole platform becomes. So, which one do you think hits first? Distribution, automation, or creation? Drop your take in the comments. I am genuinely curious what you think. If you want daily AI updates like this, join our WhatsApp channel, links in the description. Hit subscribe, ring the bell, all that. And seriously, if you're a developer or you've been thinking about building something, now's the time. Apps SDK is in preview. Agent Kit is live. Codeex is fully available, the platform's open, the distribution's there, the tools are ready. So, what are you going to build?