Transcription
Well, OpenAI just did it again. Their latest releases just disrupted a ton of businesses. ChatGPT just rolled out a new feature that makes it so you never need to leave their app. And OpenAI also shipped Agent Kit, a tool that makes it simple for anyone to build agents, and they made them all available to everyone. So, in this video, I'm going to break it all down. I'm going to share how to use them, and I'm going to give my thoughts on what all of this means. So, let's go ahead and dive right in.
Monday, October 6th. The day I'm recording this video was OpenAI's DevDay. And since we got the new Sora 2 last week, everyone was wondering what the heck is OpenAI going to announce this week. Well, here's the quick TL;DR of what was announced. It was mostly features for developers like GPT-5 Pro, Realtime Mini, and the new Sora 2. It was announced were all made available inside the API. Codex got some improvements and a Slack integration. They added some small bits of polish to their platform and enterprise controls and a handful of other things that will make developers' lives easier.
But the two things that really impressed me and the things that most people are out there talking about right now are the abilities to call upon apps and use them directly inside of ChatGPT and the new AI agent builder called Agent Kit. So let's focus on those two things.
Starting with apps inside of ChatGPT. So the new ability to use apps inside of ChatGPT is really simple. You can just call upon the apps. So let's say for instance I want to create a Spotify playlist. I can say something inside of ChatGPT like, "Use @Spotify." And you can see it automatically found Spotify and it added it as a tool that I'm going to use. But let's go ahead and type, "Use Spotify to create a playlist of music similar to Blink 182."
Now, if I submit this, it'll say connecting to the app. And then it'll ask me to actually connect over to Spotify. So, if I click continue, agree to their terms here. Agree once again. And now Spotify is connected. This is assuming you're already logged in to Spotify. If you weren't logged in, you would probably need to log in. A few seconds later, I've got a playlist called "Rock Songs Like Blink 182." It starts with a Blink 182 song and then goes into other similar bands. And if I click on "Open in Spotify," I can jump straight to this playlist that it created for me with all of the songs that are similar to Blink 182 songs, including a few actually from Blink 182.
If we take a peek over at the press release on the OpenAI website where they introduced these new features, we can see the tools that they have built in by default from the beginning. We've got Booking.com, so you can actually book hotel rooms straight from ChatGPT. You've got Canva where it can generate images directly inside of ChatGPT, Coursera, Expedia, Figma, Spotify, which we just demoed, and Zillow. And then here's some of the apps that are coming soon that you'll be able to basically chat, interact with directly from within ChatGPT. AllTrails, DoorDash, Khan Academy, Instacart, Peloton, OpenTable, Target, The Fork, TripAdvisor, Thumbtack, and Uber, and even more coming soon.
In fact, ChatGPT has an SDK, or software development kit, which is designed to help people develop apps that will work inside of ChatGPT. This to me seems like a much, much better, more powerful, more integrated version of what the plugins was that they announced a couple years ago that never really caught on, but I think that was because you had to go and make extra connections to these plugins and it was kind of a pain in the butt. Now you just talk to the tool you want to use and it will just try to access that tool. Much lower barrier for anybody to actually use these tools.
Let's have it generate something in Figma for us. "Create an org chart in Figma for a fictional AI company that generates a billion dollars per year." Let's see what it does. I didn't even add Figma. And you can see basically it just described what the org chart would look like. But notice at the bottom here, it gave me a little prompt that says, "Use Figma for the answer." So because I didn't add Figma, it didn't do it with Figma, but it suggested that I use Figma. So now if I click this, you can see that it added Figma as the tool, but it also started to just go ahead and use Figma. Now it's asking me to connect to my Figma account. So I'll continue. Approve. Figma is now connected. And a few seconds later, it generated an org chart in Figma and I can go full screen without ever leaving ChatGPT. I could scroll around. I could zoom in on it. We can see here's our CEO, Alex Chen, our CPO, CFO, Cello, etc. It, it built out the org chart and all I did was prompt it to do this. All right, cool. I could either open this in Figma now, like so, and it's in my Figma account. Or I could make changes to it directly inside of ChatGPT. "Add another layer of employees to the org chart." Going to think for a moment. Now it's generating a FigJam diagram. And as we can see, it added another layer of employees underneath. I didn't do anything other than just tell it to with chat.
Now, let's call upon Canva. I'm going to say, "Use Canva." All right, so now we've got Canva attached. And we'll use the prompt, "Use Canva to create a flyer for a rock band called NORC that has a concert on Halloween at the North Park Observatory in San Diego at 8:00 p.m. The flyer should include someone playing the guitar and spooky Halloween elements." Let's see what flyer it makes for us. And about a minute later, we've got four different flyers here that we can take a peek at. Look at this. Somebody playing the guitar with, you know, some Halloween elements here. Here's variation two that it created. NORC concert Halloween night rock show at North Park Observatory. Here's another one. And here's another one.
To me, this is really fascinating because it almost turns ChatGPT into like what feels like a team member that you can just pass stuff over to take care of for you. If you need something done in Figma or something done in Canva, you could basically treat it like a team member and say, "Hey, go make this for me real quick." And it will do it. And then if you need tweaks to the thing that it made, you just tell your team member, "Hey, now go change this." and it will go and do it. And more and more tools are going to be rolled out into ChatGPT, including apps that people like you watching this could build. They even mentioned that they're probably going to roll out some sort of monetization in the future where if your app gets integrated and gets used a lot, you'll probably get paid for that usage, which was one of the things they originally planned on doing with the custom GPT store that never totally panned out very well. But again, because this is more native and this is happening right inside of the chat, it's much more likely to get used, especially if ChatGPT is going to recommend tools to you. If I tell it to go and create me a flyer, and then it says, "Hey, why don't you make this flyer in Canva?" And I say, "Okay, make it in Canva." It makes it a lot easier to just fire up these apps and use them inside of ChatGPT. And I'm super excited to see what the like developer community builds for this because, you know, they're going to build some really, really crazy integrated apps directly into ChatGPT that do all sorts of cool workflows for you.
Now, if you want to use this yourself, as far as I can tell, it's rolled out for everyone. According to OpenAI's own press release here, it does say, "We worked with a small group of early partners to launch the first set of apps available in ChatGPT today." I don't see any restrictions on countries or specific plans that that can or can't use this feature.
What if I told you there's an AI video agent that combines the creativity of an entire advertising team with the power of V03 and NanoBanana combined? And what if I told you that Starbucks spent nearly $600 million on advertising worldwide in 2024? So, let's combine those two thoughts. I want to see if I can recreate a Starbucks-level ad for under a hundred bucks using only this one AI tool. So, for this experiment, I'm using Nvidia's new AI video generation platform. It's specifically designed for marketers who want professional-looking ads but don't have the massive budgets of big brands. So here's the prompt: "Create a 25-second Starbucks ad for Strato Frappuccino." I also uploaded some reference images directly into Invido to help guide the visuals. Once the script and assets were added, I selected ultra quality to give it that professional look and clicked GENERATE VIDEO.
"WE BACK AGAIN, Y'ALL. HOW Y'ALL BEEN? IT'S GETTING HOT." >> Pretty impressive for one-shotting it. So, I'm just going to type into the edit command box. "Speed up the playback to 1.2x." And within minutes, the ad regenerated with those changes applied. Now, here's the finished Starbucks ad created entirely with AI for under 100 bucks without leaving MY ROOM. >> "WE BACK AGAIN, Y'ALL. HOW Y'ALL BEEN? IT'S GETTING HOT." Hey, good ter [Music] [Applause] [Music] >> And honestly, it's kind of making me crave some Starbucks right now. So, if you're thinking about making high-quality ads without the huge budget, I definitely recommend checking out Nvidia's generative plan. The ultra quality option is what really sets it apart from other AI video generation tools. So, check it out at the link in the description box below.
But like I said before, that's just one of two announcements that I thought was really cool from DevDay. The other one was the introduction of Agent Kit. And this is the one that I feel like is really sort of disruptive to a lot of companies. To me, this seems similar to what NADN does, MindStudio, Lindi, Make.com, Zapier. A lot of these tools have this sort of like visual workflow builder where you can connect tools and have it call different AI agents and things like that. And well, that's what this Agent Kit does. It's a visual canvas for creating envisioning multi-agent workflows. And it's got what's called ChatKit, which allows you to embed it within your own apps and websites as well. So, you can build agents inside of Agent Kit, connect them up to your app or your website or something else, and let other people run those agents. It, it's pretty wild. And you can see it's this like visual workflow builder. If you've used tools like NADN or MindStudio, you'll probably recognize a very similar kind of workflow here.
It also leverages MCPs, or Model Context Protocols, basically making it so any other app on the internet can sort of create their own MCP server and make it so that that app could essentially tie into this agent workflow builder. So, right now you've got your tools like Zapier and Make and NADN and MindStudio and all these tools and they integrate with a ton of platforms, right? You're not limited to just OpenAI's models. You can basically connect them to anything with an API. Well, this Agent Kit is going to be very similar over time. More and more companies are going to develop MCPs for their product, meaning you'll be able to tie them into tools like this Agent Kit, but you'll probably always be stuck using the OpenAI models. So, if you really want to build these workflows and use other models outside of OpenAI's models, those other platforms are probably still a bit better for that.
Now, if you want to use Agent Kit, this one isn't actually in ChatGPT. This one is inside of their developer playground. If you go to platform.openai.com/agentbuilder, that's where you'll see this new agent builder that you can play with. And there are a few templates that you can use right off the bat. So, you've got Data Enrichment, pull together data to answer user questions; a Planning Helper; a Customer Service Bot; Structured Data Q&A; Document Comparison; Internal Knowledge Assistant; or you can create one from scratch. And we can see here that we've got a very similar node-based workflow just like all these other agent builders that you've probably seen online. I don't know about you, but this feels pretty disruptive to a lot of those tools with the main exception being that this leverages only OpenAI models. Those other tools, you can basically grab and use any models you want.
So, we can click on the start node here and we can see that, uh, it's got an input variable. So, if I was to preview this right now, it probably asks for, yeah. So, this is where you give your input variable here. And then we click on "My Agent" and you can see the only instructions it has is, "You're a helpful assistant." It uses GPT-5, but you can select any of the OpenAI models that are available. You can choose how much you want it to reason. You can give it access to various tools. You can change how verbose it is. And again, if I was to preview this right now, it would probably act just like ChatGPT with nothing else. "How are you today?" And that actually gives me nothing. It actually threw up an error a second ago. So maybe we do have to build this out a little bit more. And right now it feels like it's sort of designed for making like chat-based workflows. It doesn't seem like it's going to automate a whole bunch of stuff behind the scenes yet. There's not enough like triggers for it to start with yet. It's like input variables or, you know, feeding it variables behind the scenes are literally like the only two things you can do right now for your input like trigger.
We could add a file search here where we connect our nodes and then it will actually query any files. So if we upload documents like PDFs or doc files or things like that, it can actually query that document for an answer in the chat that you output. You can add guardrails. Tell it what you do and don't want people to be able to do with your bot. So, you have the ability to make it so it won't share any personally identifiable information. And we can change the settings on that, like don't, don't show credit card numbers or IP addresses or people's names or whatever. And you can have it sort of moderate that before it even gets to the chat. You can have it detect potential hallucinations. You can have it detect jailbreaks, all sorts of guardrails before it even gets to the chat. You can connect it up to various MCP servers. This is where it starts to get like infinitely scalable. Any product that has an MCP server, you should be able to connect it up. So, right now there's all these built-in ones, you know, HubSpot, Shopify, Gmail, Google Calendar, etc. But if we click on "plus server," we can add MCP server details, which would be provided by the company you're trying to connect, and connect this into almost any tool you can imagine.
We could add if-else statements. So when somebody gives their input, let's say if they ask a question, let's see, let's say, "User asks a question," and we'll call the case name just "question." Then we can have it actually branch off and go to one agent and have this agent do something very specific. If it's not a question, it's more of a statement, we can create a different agent response and have it branch off to this agent. So if they ask the bot a question, we can have it answer the question. And this is a very, very basic example. If they don't ask a question, just say, "Reply and tell the user we only answer questions. Please ask a question." So again, this is very early days for this Agent Kit. I imagine we'll see these nodes build out a lot and we'll see a lot of tools add MCP just so that they can get connected to tools like ChatGPT. And we're going to see a whole ecosystem build out where you can build these insane workflows.
Now, right now they're not quite up to par with MindStudio and NADN and tools like that. Again, those just integrate with so much more and you can choose which AI models you want to use. But man, this is getting really, really, really close to that. And as this builds out, I don't know, some of those companies are going to have to figure out how do we stand out among ChatGPT who can just potentially build our features into their product.
But here's some examples that they showed during DevDay to kind of give you an idea of what you can do with it. So they had a website that explained the agenda for DevDay for everybody that was at DevDay. And they built in real-time on the stage a chatbot that anybody can use inside of that little agenda app that they had. They uploaded all of the details into this agent about the DevDay sessions and who was talking about what, when, and where. And then if somebody asked a question about what they should go see, the agent would query that information and suggest what they should go and see at DevDay. I think it was a fairly simple and fairly safe example that they knew would probably work live. But you can see you enter your question, it goes through some guardrails. I believe the guardrails that they passed it through is they didn't want it to share any personally identifiable information if somebody asked for it. And so that was the sort of pass-fail. And then it would categorize it based on the type of question, pass it through an if-else statement. If they were asking about session information, it would go to one agent. And if it was something else, it would go to a different agent. A really, really simple agent that if you opened up their little DevDay app and asked about what sessions, it would make sure not to share any information it shouldn't and then it would give you the information about the various sessions and the events. Very, very simple workflow.
Another example they shared on their YouTube channel was this travel agent that they made. So, if they asked about flight info, it went to a flight agent that would share details about the flight. If it was something other than flight info, it went to an itinerary agent. And just for fun, she had it like color the background based on where they were traveling to. So, it decided that yellow was the ideal color to show San Francisco to HND. I think she said that was Tokyo. So, some pretty simple workflows right now. But the bigger implication is where these workflows are headed. Once more and more companies build out MCP servers that can connect to these, you're going to be able to build insane automations that somebody gives it a prompt and it sends it to this tool and then sends it to another tool and sends it to another tool and then messages somebody on Slack and then notifies team members and then sends an email. And we're going to be building all these crazy automations that you might do with like an NADN now just directly inside of ChatGPT. And I imagine if I had to guess, this is a prediction, but if I had to guess, they're going to allow you to build these with natural language pretty soon, too. "Build me an agent that does X," and it will go and build out the node-based workflow based on the text prompt that you gave it. So, really, really cool stuff. Really fun stuff out of OpenAI. I'm excited to see how this evolves.
Again, if you're not following me over on Instagram, make sure you do that. I'm giving away an iPhone 17 just for following me on Instagram. Uh, we're doing a competition giveaway, so head over there. The link to Instagram's in the description. Thank you so much to Invido for sponsoring today. Really, really appreciate you guys. If you like videos like this, make sure you like this one, subscribe to this channel, and hopefully I'll see you in the next one. Thank you so much for nerding out with me today. If you like videos like this, make sure to give it a thumbs up and subscribe to this channel. I'll make sure more videos like this show up in your YouTube feed. And if you haven't already, check out futuretools.io where I share all the coolest AI tools and all the latest AI news. And there's an awesome free newsletter. Thanks again. Really appreciate you. See you in the next one.