Transcription
Today, we are replacing our marketing team with AI agents that we created on a no-code tool called NAN. And we even built an app that lets us communicate with this chief marketing officer that can delegate tasks to other AI agents. By the end of the video, you will get a glimpse into how AI will replace many jobs in the economy. But you will also understand how to effectively use AI and automate tasks, and how you can communicate with those agents to be more productive. Let's dive in.
Over the past two days, we created an AI agent that controls other AI agents. So I can tell it to do a thing, and the AI agent will decide which AI agent of which we created to use for the task. We're trying to replace the chief marketing officer. Let me show you how this works.
"Please do some research on the latest about Google's uh new CLI tool. It's a competitor to Cloud Code. I want to create some content around this tool and I want to make some tweets about this. So please scrape the internet, find some tweets about it, and create three posts um on Buffer for this. Actually, for the third one, instead of scraping it, generate an image of someone using a developer tool that is called Google CLI. So generate an image for that one. So we should have three total posts."
So, we're going to send that into our app that we vibe-coded, and it's going to send that information into this agent. Now, this agent is reasoning. It's like, "Okay, which AI agent should we use right now?" It's using the research agent. It's not going to show where it is in the research agent, but this is what it's using, right? We have access to the news. We have access to um Appify. Ampify has about over 4,000 different APIs you can hit in order to scrape the web, scrape the internet, and most importantly, scrape social media. That traditionally is super hard to do on your own. And with just one API key, you can do all of that. So, like, basically, this is one integration that we set up, and if you were to go to a uh Appify or Appi, however it's pronounced, you can have it set it up to Reddit, I'm pretty sure, Instagram, TikTok, pull a bunch of useful data, and um, yeah, that's it's able to gather relevant context for the post that it's about to post. You can get one API, but we went ahead and connected the MCP, which now has access to all 4,000 different APIs. So it may spend a lot of time on this research step, which is good, right? We're already saving a ton of time. We want this to upload it to Buffer. So Buffer is the tool that we use to manage our social media. And in Buffer, you can create ideas, right? And this is like the CMO group. And this is where the agent puts all of the the posts. So it'll go there once it's done. We have no idea which direction it's going to go. Like, this is, I mean, it makes its own plan and then ends that plan. Yeah. Literally has a step in here. It's called "think." And so it will literally spend time thinking after each step.
"Does it hit this every time or does it hit it sometimes?"
"If it needs whenever it needs to think, it'll think."
"Yeah, that's great. H also the research agent also has think. So now there's two thinkers."
"Wait, is there think?"
"Oh wow, you can think."
"Okay, so..."
"Oh, okay. So here it's now thinking again. I swear if it starts researching again."
"There we go. Now we're on to the media step. So it's either um hopefully it's both generating an image using FAL workflow."
"Okay, it's done with the link. Now it should go to poster."
"It might go back to media to make the other ones."
"Okay, let's see what it does. It's thinking. Oh no, back to poster. Must have gotten all the information it needed. And now it is posting them as an idea on Buffer. And this is where you can schedule posts on social media. Makes it very easy."
As you can see here, it just completed. Wow. The execution took seven minutes. Have you seen one that long before? No, that's the longest one I've seen. That's crazy. If we go to Buffer, we should see those posts.
So, as we can see here, and it used emojis. I have not set up the examples that well, but it created a post. This is what the future of coding looks like. It did get this, I think, from the Notion database. Yeah. Um, as an example, because "this is what the future of coding looks like" is one of the hooks that I have literally used before, word for word. Yeah. Um, and here it has this post that it generated from on FAO. So I feel like we maybe want to upgrade the model, but I mean, that's fine. It's, you know, someone using a Google CLI. You know, it looks like a not bad. Yeah. Um, and then it, for these ones, it actually scraped them. Scraped the image off the internet. No idea where it got this. That's okay. Um, yeah. And so it created these three social posts.
Okay. So now, what I want to do is I actually really want to break down what's going on here, right? This is a chief marketing officer agent, right? And obviously, it's not fully complete, but it will be. I guarantee you by the end of the year, we are going to create an AI agent that replaces, you know, 20 workers that we can do um a lot of really interesting um interesting things like generate full content videos, generate ads, like we're seeing all of that technology come on the scene right now. What I want to do is break down kind of how we're doing this now.
So this is the master agent. Whenever we send data from the app that we're creating, right? We have the VAP that we vibe-coded that sends the data from the app into the webhook, right? That is what makes it a VAP. It is a vibe-coded agentic application, right? And it's just any app that you create that has an agent. Doesn't need to be a mobile app, but in this case, it is. So, this agent has access to these tools, right? It can do it can uh, you know, send information. It can do research. It can generate or find media. It can post things on social media. That's really cool. And we're going to break down these individual tools. And so the first one that I want to break down is actually the biggest one, which is Helper 2.0. And Muhammad, why don't you break down this behemoth of a tool?
Yeah, this thing is huge because this is just another layer on the onion because right under this, it calls in three other agents. So let's break it down. The helper has access to the email agent, which basically controls your entire Gmail. So it can mark unread, create a draft, send, label, reply, get labels, and get all the emails. Can also update, delete, get, create, and invite people to your calendars. And then the other tool that has the other tool that it has, the contact agent controls your entire Slack. Um, so it can get your users, get Slack channels, open messages, message user, and get channel history. That's insane. So like, it can basically fully control your email, fully control your calendar, and fully control your Slack. Yeah.
And so if we go over to my screen, what we can do here is we can actually test this out. So we're going to do a simple request in this giant thing right here, right? So we can hit test on N after you create any sort of workflow. If you're using the webhook URL that's right here, make sure that you put it on test mode. Um, and like make sure you press this button, or else it won't send it through. So now we're going to send it through on our VAP. And by the way, by the way, in order to message this master CMO, all you have to do is add in new webhooks and we named them, right? So this is master test, and we can uh, we can very easily hit this um URL um by adding new ones. And you can do that by pressing settings, and we're actually going to create this app on the App Store that you can use, but you can also create your own VAPs. But anyway, um, let's go ahead and test this.
"So, I'm going to say, 'Please schedule an appointment tomorrow morning at 9:00 a.m. to film the uh short-form content with Muhammad, and also send Muhammad a message on Slack reminding him of this.'"
And you're going to see it start here. Now the agent is thinking, right? It sent it through the webhook. Enthropic is thinking. Now Helper is working. It's not going to light up here, but it is going through this entire process. So we have basically an agent handed it off to another agent that is now thinking, and it's going to delegate it to two other agents. And that can go on in theory, like for as long as you want. Yeah. Yeah. That's insane. And um, yeah. And so this one won't take as long. Um, but if we go back to Muhammad's screen, we can await his message.
"It said hi."
"Wait, why did it just message me hi?"
"It just messaged him hi."
"But it's still working."
"It's still working though."
"But is that just mean it's doing the calendar event?"
"Um, maybe. I don't know. Maybe it wants to see uh what I'm up to."
"So in the instructions, wait, so why did it message hi?"
"Like so in the user message, I actually just told itself to message myself hi."
"Oh, you were doing testing. Okay, you forgot to fix. Okay."
"Yeah."
"And so when you actually make these, you want them to be JSON variables so that other AI agents can talk to each other."
"And it's okay."
"Yeah."
"Yeah."
"Yeah."
"That makes sense."
"Yeah."
"And we actually did that a little bit in the last video. So if you want to learn more about that, you can check that one out. But right, um, should we test it again?"
"Yeah."
"Yeah, let's try it again."
Okay, so we're going to test this workflow again on my screen here. And what we're going to do is, "All right, tomorrow uh at 12:00 p.m. I want to film some short-form content with Muhammad. So please schedule that on my calendar and then um message Muhammad um letting him know on Slack."
And you can see here our agent was sent this information from our app that we've created, the VAP that we created. And now the Helper is now going to work. So it's doing all these things. And we can actually go ahead and watch Muhammad's screen. Hopefully he gets the correct Slack message this time.
"If you go to Slack."
"All right, I got the message. 'Hi Muhammad, just wanted to remind you about our scheduled meeting tomorrow at 12:00 p.m. for filming short-form content together. Looking forward to collaborating with you. Let me know if you have any questions or need to make any adjustments to the schedule.'"
"Lovely."
"And if I did, you can also go back to the master CMO and and ask it, 'What did Muhammad say about my meeting?' and it would know because it gets the messages."
"Wait, what do you mean?"
"Like on on here?"
"Yeah."
"Wait, it gets previous messages?"
"Yeah."
"Well, you can tell it, 'Hey, go to Slack. Check what Muhammad said.'"
"Oh, I see. I see. I see. Oh, amazing. Okay."
So, as we can see here on the app, it said, "Perfect. I've taken care of both of those tasks for you." Um, and that's really cool, right? And um, yeah, so that's kind of the basic essence. Oh yeah, I forgot to check the calendar. So if we go to calendar, uh, we can see here on calendar, it should be scheduled for 12:00 p.m. Look at that. Short-form content filming with Muhammad. And it even spelled your name correctly because it did the Slack part first, and it even invited me. Oh, and it invited uh it invited Muhammad uh to the event. That's awesome.
So, okay. So, that is the basic essence of like the communication side of it, right? That is the first tool, right? Based on whatever we put in, it has access to our calendar, it has access to email, and it also has access to Slack. That's pretty cool. But, you know, we might not just want to communicate with people. We might also want to do some research, right? We might also want access to news. We might want to scrape the internet. We might want some of that. Um, we might want to use some of that in a longer workflow. Uh, you want to break down this tool real quick?
Yeah. So, yeah, the research tool is just another AI agent that has um access to the news API. Um, so it can just gather all the news for today or a month or the last year. um SER API just for basic Google searches. But what I'm really excited about and honestly, which I found out about today and I'm gonna deep dive on it all day today um is the Appify MCP. So Ampify is they host over 4,000 different APIs you can use and mostly for scraping the web and maybe social media. Um, and they just released released an MCP server. And so what's crazy about this tool is like MCP is like really hard to set up. And so this is like and when you create uh when you create these workflows, even on N8N, it's really annoying because you have to put in your own credentials and API keys to use all of it. You have to set up billing on all of them. There's only one portal for billing. So you can use all these different tools with one singular key, which is like your personal password or account. Yeah. Yeah. And that makes it super easy. Um, and if I go ahead and open the node, you can see how super simple it is. It's just the SSE endpoint, my O, and you get that on the Appify site. Okay. And it's you just sign up and you get one. And they also have a free tier as well. Amazing. And so like this this agent that our agent has access to is kind of like it gathers context. You can kind of think of it as like a little MCP or it serves that kind of function of gathering relevant context because for marketing again, context is so important in marketing. It's actually the most important thing. A lot of people don't don't focus on it enough. And so why don't we try and give this a go on our on our um VAP here.
"And so what we can do here is we can say uh please search the internet um and find the latest news regarding Cluey and their recent fundraise. And then um yeah, please use the SER API for that. And then I want you to um message Muhammad and let him know what research you found, please."
And so what we're doing here, right? We sent that data from our VAP to our master agent, which will then do some research. And then after it's done researching, it should indeed go to the Helper agent to then send the message to Muhammad. And again, this is just a glimpse into the future of AI agents, right? Right now, you do have to do a lot of manual work, and that that is just kind of how it works. Now, if you want the edge in anything in AI, like you do have to do a lot of extra work to get things to be like to be able to automate things. And I think the best people I've seen at creating apps that allow for, you know, high-quality automations or high-quality agents are the people who have like a deep attention to detail. They're willing to fail over and over again so that they can just like automate a bunch of these boring tasks. Yeah.
"How long do you think it'll take before someone automates an entire company?"
"I mean, I bet you it's already happening. I bet you it's already happening at like a small scale, like little e-commerce companies. Yeah. Um, but like until someone were to automate like if it was one person doing like raising money in a startup. I mean, there's literally a startup that raised money to all they do is buy businesses and then replace all the employees with AI agents. I forget what it's called."
"What?"
"Okay, wait. So, it's done researching here. And now that it's done researching, it should send that information back to Helper."
"Yep. So now it is in Helper. And if we go to Muhammad's screen, you can see that he's on Slack. Let's see if he gets it."
"Oh, okay. All right. So, what does it say?"
"Muhammad just got a message. 'Complete research on Cluey with a K. Recent fundraising activities. Here are the key findings.'"
"But it did raise um $15 million. That's true."
"All right. So, it did research on Cluey, use the SER API, and it sent it over. And so, that what I'm trying to show here is you can specify what you want it to do. And then the agent will decide that if it needs to use multiple tools, it can do them in different orders. And so, like what kind of like allows for all of this to run smoothly, it's actually something in the master CMO agent that has access to all these other agents. And that is the system instructions. And I think we've we've done a pretty simple system instructions, but do you want to break this down real quick?"
For sure. Yeah. So, one of the most important parts of a system prompt is how you set up the agent's mindset. So, that would be right in the beginning. And so, I said, "You're a master CMO agent that will have control over other agents, right? You're current you currently have access to four agents." And then, so we break it down for for the um the orchestrator agent, the master agent. Yeah. Um, and so we basically just say, you know, what the agent does at a high level and sort of the tools it uses because this isn't really going to do any of the work. It just has to know the context of what that agent can do. And so let's say that um the user asks, you know, "I want something in Notion or I want to message someone on Slack," it can point to the right agent that has access to Slack. It doesn't actually have to do any of that. It just...
"Totally."
And if you go to my screen real quick, one other thing like when you're thinking about this, right? One thing that's really important is access to Notion. So if you can see here on the poster agent, right? It has access to Notion. Here is where we have our SOPs, like get content SOPs. These are basically like I have a list of all the good hooks I've used. I have a list of how I do my podcast research. I have a list of how I do long-form, short-form tweets. It has access to all of that information. And so if I ask for a tweet, I want it to look at this and be like, "Okay, how does this guy write?" So as the the AI in the posting step, right? It should know to do this. And you could actually do this in two places, right? Um, you could put that in the master instructions or you could put that in the instructions that are part of this, right? You can you have a system message for each individual agent. And I'm sure that's actually where you put a lot of the information. Um, and so you don't have to put everything in this main master agent instructions. It just needs to know to send it to that agent because once you send it to that information, it needs to know what information to give it. Yeah. Right. It needs to give it all the information, but you're basically delegating it to this new agent that has a more um in-depth and uh a more specific agent instructions. Yeah. Yeah.
Yeah, I mean, if you think about it, it's kind of exactly how a company works. You know, the CMO, his main job is to delegate labor to other more skilled workers. So, and if you delegate it wrong, that's when things can go bad. So, it's really important to make sure that you tell the master agent exactly what does what and what information comes back. Yeah. Yeah. It's like it's more you have to be able to send full instructions, right? If you're like if we're sending information down to the engineering team like of a new feature we want, we don't have to know how to implement it. We just know have to like, hey, I want this feature that does this because this user said that we should add it and these 10 other users said that we needed to have it and I think it should look like this, this, and this, right? But I don't need to send them full instructions because that's what they know how to do. Yeah. And that is how this is becoming this like agent army that you can create. Yeah.
And we just created the front-end interface, and you watched the last video that we created. Um, we're actually adding on the Vibe Code app, we're adding the ability to integrate directly uh into N8N. So when you create an app, you can just press the N integration and plug in your master webhook, and you can immediately hit that agent, and it will actually generate your app for you in a in a format that will work with N. Yeah. Yeah. So that's really cool. You should do that. Um, like you can sign up. I'll put the link in the description for that. It's a really easy way to do it.
Let's go ahead and move on to the third agent here, which um this is the one I think we've talked about the least, the media agent. Yeah, the media agent. Let's talk about that.
All right, so the media agent um it has four separate FAL workflows to generate media. Um, so, and it also has two different inputs as well. It can take in just text or image and text. And when we add the the on the Vibe Code app, you it'll automatically work out of the box. So you'll be able to input images and text. Yeah. So it'll just send you an image URL, and then the agent knows what to do and where to send it. Amazing. So it'll just work. That's awesome. Um, so with those four, we can go text to video, text to image, image and text to video, or image. And then we also have again the Aify MCP, but specifically we're using this to scrape images off the web. So if we wanted like a real specific image, um it would scrape it from Google Images and send it back to the agent. Amazing. Yes. And that that I think is really important. I actually really like the ability to like get specific images. I think it makes any text output more fun if you have an image there. I think Perplexity was one of the first ones to render images in its response. Obviously, OpenAI renders images in its responses now. So it's like that's why Perplexity has to kind of pivot right now. They're in a tough spot. But I think it's really important in your outputs to have images, and why not just like pull them from the internet. Yeah.
So what we can do now, now let's do a mixture of step three, which is this um the media agent, and then let's go back to Helper. So what we can do here is I am going to give this a test on my phone here and we can say um uh please generate an image of a man staring at the camera holding his phone and have some uh text on the screen saying, "Oh my god, this is insane." And um I want you to generate that image and I want you to do that three times, right? I want to have multiple um attempts at this. And uh please make sure you send these to Muhammad after. I want you to send these image to Muhammad in Slack. So please do that.
Okay. So we're sending this workflow. Can't wait to see those. Hopefully it spelled my name right. Hopefully it... Yeah, we got it. No, that's good. Amazing. All right. Okay. So now, as you can see, the master agent has passed it to the media agent. Now, the media agent is going to work. Media agent. And one thing I want to do in the future for the media agent is upload really good uh examples of um like if we we could actually add a glyph workflow that would be really useful because I've mastered my um I have a um YouTube thumbnail workflow that's really good and it uses my own image. Yeah. Every time. That would be really cool to add. I mean, anything with an API can be hit with these. Yeah, that's really cool. And so, we can we can really get granular about YouTube thumbnail ideas, Instagram story thumbnail uh ideas, and we can put all of our brand assets in there so it keeps the colors correct. Yeah. Um, because right now it's using a pretty simple workflow on FAL. We haven't really gotten into that yet, but as you can see, it's done actually no, it's done two of the image generations. Now, it's going in for the third. Um, and then it should go to the Helper agent to message you. Yeah. And every time it generates an image, it stores it in that memory. And then once it has those three images, it knows it's done generating. And then it'll go to the Helper and send it to me. Amazing. And so it's done three. Hopefully that's the last one it hits. Hopefully the next one is Helper. Come on, baby. Oh, it's go. Amazing. Agentic. And so now it is on Helper. And if we go to Muhammad's screen here, boom, look at that. On Muhammad's screen, you have three links here. Oh, and they rendered in here. Yeah. Amazing. And so we created these images. "Oh my god, this is insane." Yeah, this is insane. And you know what? We're using the uh flux context Yeah. model here. And Flux Context is actually much better at editing than it is um than it is generating images from scratch. But that's not bad. And um I think if we um wanted to dive into the details here, we could make sure that if we say thumbnail, it would do horizontal orientation. Yeah. But that's something we could include in the system instructions in the media agent. Yeah. Yeah. And so yeah, that is tool number three. Generates images and you can create. This is one that we, you know, we've probably put in 30 minutes of work creating these workflows. We could easily put in a full day, get all the brand assets, and that's one thing we might do in a future video. Yeah.
Okay. And so the final one that we can dive into, final tool, um, is the poster agent, right? And so let's go ahead and let's try to do one. Um, I have an idea here. And one thing I built into the VAP on Vibe Code is you can actually just press this text button here and you can actually just type into. And so what I want to do is I want to generate an image and then I want to turn that image into a video. So we only need one agent for that. If you remember here, we have this image and text to video. Yeah, which is super cool. And then I want to take that whatever that output is. I want to post it on social media using the poster agent, which will actually just post it to Buffer. It'll basically allow us to schedule it with one click, and then I want to let you know when I've done that. It'll send you a message on Slack. So, let's see if we can do that all in one little message here.
"So, I'm going to say, 'Please create an image of a man on a unicycle with an orange shirt that says download Vibe Code.' Um, then turn it into a video. Then post to Buffer. Then uh message Muhammad when you are done. So we can use text to type that in."
And we're going to hit send. And as you can see here, it gets sent to our master agent. It should be sent here. Now it is likely generating an image and then generating a video. Then it will go here. So let's see. And this is using three separate agents. Three separate agents. It's going to use this agent, this agent, and this agent. Well, four technically, because this five. Yeah. Wait, six if you include Zapier. Oh yeah. Yeah. With workflows here. Oh man, this is insane. There's so many layers of this onion. Yeah. And it's only We're It's We're gonna This is only the beginning. This is only the beginning, too. Like, we're gonna make this a lot more refined, too.
"Did it already go to the poster agent?"
"Okay. So, it's hit this twice. So, it hit the media agent twice. And now, it's going to the poster agent. Perfect. Okay. It is done with the poster agent here. And so, as you can see, it is now going up to Helper. So, it's in theory, it should have posted it to Buffer. And now, it's going to tell Muhammad all about it."
"Wait, did I tell it to include the video?"
"I think you just said, 'Well, let's see what it does.' You said, 'Tell me about it.' I I don't think I specified to tell uh to send you the video, but we'll see if it does because it should just be a link to the video, right?"
"That'd be awesome."
"Yeah."
"Yeah. So, if we go to Muhammad's screen, we'll see what the agent says, and then we'll check and see if it was posted on social."
"It came in. 'Hi Muhammad, have completed your tasks that you requested. An image of a man on a unicycle with an orange shirt that says download Vibe Code converted to a video and posted on Buffer and social media. Everything is ready to go.'"
"That is incredible."
"Um, and Helper still running for whatever reason. Maybe it's just checking its work or something like that."
"Workflow successfully executed."
"If we go to Buffer, we see that we got this. Uh oh. What is this?"
"We even watch this. Let's see. What is this?"
Okay, so I see the Vibe Code lo or the Vibe Code colors. Download Vibe Code. And it's using V2. It's not using VO3 because V2 is just insanely expensive to run. But we have this little video of a guy riding a unicycle. And as you can see, it generated the video. It did all the steps, and it messaged Muhammad. I could have updated it to message the link to the video because it is represented with a link, but that is just some things that we could refine with the agent. And not going to lie, we made this today. We made basically most of this flow today. Yeah. And in the next video, like we're going to really I think we should make a really refined workflow for something that we do because we I use a lot of AI um workflows for like YouTube thumbnails, uh research reports. I think we could include some of the actual workflows I use and like really build it out. Yeah. Yeah. Anyway, so if you guys have any ideas, let us know. Hell yeah. Well, we'll see you in the next one. [Music]