I Built 4 AI Agents in 2 Weeks – No Code!

I built my first AI agent in one week. We processed thousands and thousands of LinkedIn profiles through this tool. By the end of the second week, I had 4 agents. We have a tool right now that's like vacuuming up direct messages in real time. The agent that I built that does the mutual intro thing, there's no code. There's, there's literally no code in it. I couldn't imagine how long it would've taken if an actual person had gone and done it. Hey, I am Adam Brekane. I am the CTO at GatewayX. Based in St. Louis, Missouri.

I'm Andrew Warner from Bootstrap Giants, and we're going to go through Adam's story of how he's learned and started using AI in his businesses.

I was a pretty early adopter of like GenAI. First of all, AI is this like, you know, machine learning. It's been happening for decades. Like we're using it. It's how ad algorithms work. Like that's AI. But this like GenAI, like new world of ChatGPT and, you know, chatbots and stuff like that. I've been using that for years and using code generation tools like Cursor and GitHub Copilot. I was in the beta of GitHub Copilot, whatever it was, 4 years ago when that came out, and it was wild. It was like this amazing tool that I can start writing code and then it suggests a bunch of code and all you have to do is hit tab and now you've written 10 lines of code just because you write one comment. Even then, there was a language that we had to learn of how do you talk to this thing. Something I noticed really quickly was when you're writing code, you're writing functions and logic and telling the application, do this, then do this and move this data. There's always this debate among programmers of how much should you comment. Comments in code are just for the humans. It's just a thing, it's not executed by the program. It's just a note for whoever comes next to understand why did you make this weird choice? What are you doing? It can help you organize your thoughts. And something we found really early on with Copilot was it really likes comments because that's your little instruction to the model to say, hey, I'm trying to do this in this way with this information. And then the suggestions it would make were a lot better. So we learned some stuff pretty early on. That's about what we learned. Years pass, literal years pass. ChatGPT-2, ChatGPT, you know, there was kind of a 3-ish, then 4, and like these reasoning models. I've spent the last several years, uh, not a naysayer, but kind of dismissive of AI. And last year specifically, like, they came out with, with 4, uh, 4.0 at the beginning of last year, at the beginning of 2024. And it's like, it's cool, it's pretty cheap, but is it useful? Or is it like this question, like, is it, you know, is it actually— is this a thing that's going to change businesses in our lives? Or is it like the same hype of NFTs that came and went, you know? Oh, NFTs are going to control the world. And like, are they though? No, they didn't. They probably won't. So I was kind of dismissive. And honestly, it's just because I didn't get it. It, uh, I— all of last year, severely below the line, severely like, I don't understand it, but I'm also afraid of it because if I don't understand it and it does happen, you know, what's that gonna mean for me? Will I be valuable? Uh, am I at the point in my life and career where like I'm not gonna learn the next thing? You know, where, where like, like, you know, if you're 22 years old, like, yeah, you're gonna learn the next thing and learn the next thing. Okay, if you're 59 years old, like, are you going to learn the next big technology or are you going to accept the forced retirement? Like, nah, I'll accept the forced retirement. Well, I'm like in the middle. So I was like, well, okay, is this like, am I going to figure out this AI thing? Am I going to program AI? And the reality was I just hadn't seen anyone do it well. There's a lot of hype. Most of what you read, 80% is hype. It might even be more than 80%. is hype, it's not really true. You know, a question we've started to ask when we look at a post of something that looks really cool, it's like, well, what are they selling? 90% of the time, they're selling the thing that they're showing you, that they're demoing and telling you how amazing it is. Okay, well, so it's not fully real. Like, they're just saying this is amazing and they got a lot of venture money. So, I saw in January, A YouTube video, one very specific YouTube video. I could send you the link to this video. And it's just like YouTube content creator, young dude who's like, here's how you can run rag agents on your computer. And for whatever reason, I listened to it on the way home in my car, driving my son home from school. Listen to it. Got to the end. It's like a 30-minute video. Got to the end, listened to it. It's like, you know what? Like, that was actually pretty interesting. I think I want to try that. And so I watched the whole thing again at my computer and I followed along and I like built this AI agent following exactly his steps, copy and paste. I was really dumb. I was really bad. Like, Andrew, have you— what's the most recent thing you've been like bad at as an adult? You know, often I think people like pick up an instrument or something and like—

Playing the guitar for sure.

Playing the guitar. So like, it's as adults, I feel like we kind of forget what it's like to be terrible at something. And, you know, because we do the stuff we're good at, we hang out with the people that like the things that we like. We like the things because we're good at them. I was like bad at AI. I was slow. I didn't understand. I didn't know how to use these tools. This tool was a video on N8N, which is a platform like Make or Zappy or whatever. It's a workflow, online workflow tool. And I followed along. I did it really slowly. And I built a dumb, useless agent. Because like, you know, what's the first song you learn on guitar? It's like Mary Had a Little Lamb or whatever, right? Like, that's, that's the level you're at. Well, that's the level I was at in, in literally January.

What did the agent do? What does it mean that you, that you built?

So the agent, there's, there's one of the most popular kinds of agents is called a RAG agent, Retrieval Augmented Generation. What it means is you could give it a bunch of documents and then ask questions that those documents could answer. So if I gave it, you know, if I gave it all of the newsletters from Bootstrap Giants, I could say like, what do these Bootstrap Giants guys believe? Like, what do they like? What do they not like? Like, what are their top 10 tips? Like, it would, it would go into that, that knowledge base and, and answer questions. And this was one where you could hook it up to a Google Drive folder, but you're like setting it all up yourself. So it's not just a tool where anyone could hook it up to their Google Drive folder and start talking to it. It's like, here's how you would custom build one of these agents for yourself, meaning you get to control how the data comes in, where it comes in. You control how the agent thinks about the data, how it's crafting the prompts, all of that. So it's customizable. But still easy enough using this tool, N8n, easy enough that in the span of a week, I had an agent that I could chat with that I had built myself that could look at and understand documents in a Google Drive folder. One week from when I watched this video the second time to I have an agent. I built my first AI agent. And this, like, this has been a concept for years. But again, it was like this one video unlocked it for me. And then I had an agent. And by the end of the second week, I had 4 agents. And they were like slightly more useful for business, not production ready. By the end of that first month, we have 2 agents like running that we're actively using in workflows every day. Like they're useful, they're production. That's happening.

Doing what?

What? What'd you say?

Doing what? What do the agents do that are actually productive, or what did they do?

Yeah, so the one that is the most interesting is we use it for the Mutual Intro Playbook, and the agent, when it's given 5 or 10 or 20 LinkedIn profile URLs, it looks at Jesse's communication history with those people, And it ranks them in order to say, you should reach out to these, this one or these 3, and you should definitely not reach out to the, you know, to these ones for, for these reasons, like because they've been very unhelpful or negative in the past, or sometimes they've been very helpful and very interested. And so it ranks all of that.

So Mutual Intro Playbook is the process that Jesse and all the GatewayX companies use to get an introduction to the ideal customer. The problem with that is you don't know who do I ask for an intro, and what you're saying is you fed the agent Jesse's what, email or LinkedIn contacts, and let the agent figure it out?

LinkedIn contacts, LinkedIn DMs. So yeah, so the problem, it's an opportunity too, is often if, you know, if I'm looking to talk to the CEO of this agency, I might know 10 people that, I don't know this guy, but I know 10 people that know this guy. And of those 10 people, you could imagine there's, there's a pretty wide array. Maybe some of them work in his agency. Well, that's interesting. But maybe it's a junior customer support sales rep, you know, whatever. Maybe there's one of their investors that I know. Maybe I know his co-founder. Like, there would be lots of ways. So, so the question is, who should I reach out to? And, you know, if you're doing this one-off, like, I want to meet one person one time, It's possible to go and look at those 10 profiles and understand it. If you want to do tens or even hundreds of these intro requests per week, you need a system. Not only like do we have a full-time employee that's running this, but this employee is using these AI agents to help on these like really time-consuming— he's processed thousands and thousands of profiles through this tool.

Right.

I couldn't imagine how long it would have taken if an actual person had gone and done it. Plus, the quality is better. Like, multiple times I've looked at the list, and I'll recognize someone that it ranked as like 3rd or 4th in the list. And I'm like, oh, but that's probably the person, right? And I'll go in and spot check because there's the kind of the engineering side. It's like, well, but the agent was probably wrong. It hallucinates, whatever. I go in and I look at the conversation history. from the people it ranked in places 1 and 2. It's like, oh yeah, no, those, those are better. Like, that's— yeah, that's right. I wouldn't, I wouldn't have expected that. I wouldn't have known that. Like, uh, but, but because it had all this context, because it could search, uh, so much more effectively than I could, like, it finds stuff I wouldn't have noticed.

That's super exciting that it works to that degree already. And you're still in the— okay, at this point in the story, early stage of figuring out what this could be. What did you do next?

Yeah, at this point, like, I think we, you know, I'd look at January as kind of a failure of agents, like learning what agents don't do. Because we got into our minds, well, let's build an agent to run this whole mutual intro playbook. The mutual intro playbook, you say, I want to meet this person. And that means you have to find the mutual connections, you have to rank them, you have to reach out to them. You have to respond when they respond back to you. If that doesn't work, you make another. You could think, and a lot of the AI hype is like, yeah, build AI agents. They'll replace your employees. They can run all this process. Maybe in theory, but they don't do that well. What we found, we built that agent, by the way, and we got to the end of the sprint. And there's a tool that it's kind of similar to called Happenstance. We got to the end of the sprint and in the retro meeting, I said, hey, like, you know, we built a 4 out of 10 version of this tool that these other guys built. And Jesse's response was, why did we invest in a 4 out of 10 tool when these other guys are doing it better? So, well, hold on a minute. First of all, we think they're only doing it 7 out of 10. And we've talked to that team directly. We've met with their founders. Like, they don't want to make the changes that would make it for us a 10 out of 10. So that's why we want to build something. We also wanted to build a 10 out of 10, but we didn't know what the heck we were doing. So we tried, we thought we could do it, we thought it would be easy. We bought the hype, even though I'm super skeptical, bought the hype and didn't work. And realistically, what we learned is the agent just— you think of an agent as an employee. That's what LinkedIn tells you, all these creators on LinkedIn tell you. But if you think of them as employee, you really discount like how much agency employees have and how much they can understand and how they ask for help when they need help. And they can figure out and understand this context. And, you know, if I show you like 3 different faces, you could immediately tell me, oh, that's happy, sad, that's this. And the AI like might get it wrong part of the time. And when you have these really complex agents with lots and lots of steps, You, you introduce a little bit of, of kind of error at each step. You know, if you said, AI agent, like, or ChatGPT, make me a recipe for lasagna. I don't know if you remember, but some years ago people were asking it to make recipes and it was terrible at it because they were like plausible looking. But then you'd look, it's like, wait a minute, you want me to use 6 sticks of butter in this? And it's, it's just like not quite right, right? It's like all the right pieces but the wrong, the wrong proportions. Well, now it's better at those things, but it's wrong in little ways. Oh, it's 5 minutes too short here, 5 minutes too long there. If you're only just talking to it once and it gives you a response, there's a little bit of error. Usually that's okay. If I'm going to have it rank these mutual intros and I've got 10 profiles and it swaps who's in place 1 and 2, do I really care? Probably not really. I bet the process will still run fine enough. And it's just as good as a human do. Human would make some mistakes and it's subjective. Now imagine there's an agent that has 20 steps from the input that I gave it, the request, and the output, the action it actually did. And at each of those steps, there's a 5 to 10%, you know, muddling of what it was that I truly wanted and how to think about it. And so then you have 5%, and then it calls one of its tools and it adds 5% more error, and it calls another tool which adds 5% more error, and it calls another. By the time you get to the end, you're like, well, okay, I can like look at it and see that it understood at the beginning. And man, partway through it just went off the rails. Like the agent that we built, it would return list of people and sometimes it'd be just the wrong list of people. You're like, okay, that was just a garbage one. And you'd run the same tool again with the exact same input and it would give you a great list of people. And you'd go in and look and it would say like, okay, yeah, Andrew Warner, who has a lot of software engineering experience because he lived in San Francisco and blah, blah, blah. Whoa, whoa, whoa, hold on a minute. Does Andrew Warner have a lot of software engineering experience because he lived in San Francisco?

Mm-hmm.

No, no, no, no. A lot of other people lived in San Francisco and had software engineering experience. And because there was kind of this muddling going on, it took all these traits from other people in this retrieval augmented generation thing. It took these traits from other people that were true and it assigned them to you because, yeah, sure, same ballpark, right? You live there, you probably have this too. In our database, if you looked at all the people that live in the Bay Area, probably 90% of them have software engineering. Well, so the LLM is going to kind of assume that the other 10% do too, because that's like the word cloud that it's made for itself. And so it's these like hallucinations that we just like couldn't— honestly, we abandoned that agent. Lots of like bits of the agent are useful, and we've used them to make these like smaller, more single-purpose agents that do one thing really well. And, you know, they're only 2 or 3 layers deep of LLMs. So it's 5%, 3 times, not— 30 times. But any of these queries we were running that took like 10 minutes to return, it was all garbage.

So what's next?

Yeah, I think we probably want 10 to 20 of these single-purpose tools. There's still AI agents, like some of the stuff the AI agents do really well. is they can understand the concept. And so, like, our, our, the Slack bot that runs this Mutual Intro, uh, tool, it's really good at understanding any possible way that you give it the LinkedIn profiles. If you paste it as a list, if you do it a bulleted list, if you use a different URL format, if, you know, whatever, you could add in some text instruction too. Like, it's really good. at just pulling that out. That would have been like a whole job. Someone would have, you know, a year ago, I would have built a web page. It would have taken me a, you know, a couple days to like get the input right, get the thing hosted somewhere. So I think what we've figured out is, is these small tools that do one job, they can do it really well. And so we're going to dive into more SOPs that we have internally and replace more and more like little bits of the work. Like, oh, this is a thing a person is doing that they truly don't need to do and might even go better with the AI. Like, let's drop in. Like what? Give me an example. More example intros. So then one of the next ones is in order to make these lists of people that we're searching, someone's still going into LinkedIn Sales Nav. They're searching for people. They're finding, okay, they're sorting that list and saying, okay, these are the These are the 20 that I think I want intros, you know, from. Okay, then I'll take those, drop those in a spreadsheet. All of that, like this, these tools can, can use LinkedIn Sales Nav, uh, pretty effectively.

Because they're clicking or they're getting API access? There's no API access.

There's no API access. Uh, the clicking ones is like a whole other story, uh, of like mostly hype, but the, the API access, like part of the magic on LinkedIn is, is there's a lot of scrapers and tools built on top of it, kind of in the gray zone of like getting access.

Yeah.

And so if you build on top of those, they handle the like, how do you get data in and out without breaking something? Like we have a tool right now that's like vacuuming up direct messages in real time as they happen. Well, that's great, 'cause now those direct messages are in our database. And so when we wanna ask about the direct messages, we don't have to deal with LinkedIn anymore. I just go to the database. So there's one little tool that says get the DMs. There's one little tool that says get all the connections. There's one little tool that says, you know, grab Sales Nav, like search filters. And then we, we have all of those so we can query them endlessly at that point.

What are you building all this on?

Yeah, there's a tool called N8N.

So it's still N8N, the one that you created the original RAG agent on?

Yeah. And 8N is, is pretty, it's pretty cool. Are you familiar, like how familiar are you with like Make and Zapier?

I know Make and Zapier really well. I did a search to see what this one was. This is an open source workflow automation tool that allows people to do the same thing that Zapier would do.

Yeah. It's, yeah. When I, when I frame it like, you know, Make is a little newer than, than, than Zapier. So it's like slightly more new age, a little easier to use, more relevant, like current integrations. And it's a little more customizable. So if Zapier and Make are here, N8N is like here. It's not, it's not code. You're not sitting on your computer writing Python. You don't need an engineer, but you're building workflows and super natively. They have blocks for LLMs. They have blocks for agents where you can connect tools. So it's, it's, you know, in any workflow that I would build in N8N, there's 2 to 5 like LLM prompts. somewhere in the workflow. And maybe the path that the workflow takes depends on the result of a prompt. So it's a little more flexible around AI than those other ones that are much more like A to B with these 5 steps in between.

All right, let's close this out with you dreaming with me a little bit.

Yeah.

Where could this go? Where would you want it to go?

I mean, I don't know, I have so many dreams. Like, one that pops to my mind is, well, wait a minute, like, why can't the agent build these agents? Like, come on, like, I just want to tell it what I want. And the agents are all defined in code, in theory, you know, under the hood. So that's one thing. I think the bigger thing is more like, if we get to understand this how to think about these tools, how to use them effectively, like how to build these products as well that are useful, at least to us. I think we're already probably in the top 1% of understanding this based on my conversations with other people. And we're really low in how we understand it. So my dream honestly is like, well, if we're that great at it, if we get— and we get better and better and better and better, and we're getting better more aggressively. than most other people, we're going to be ahead on so much. I'd love to put out content for people to teach them how to do it. I really enjoy that. I'd love to buy businesses with the explicit intent of teaching the employees how to use all these AI tools better. And also, you know, we have like AI products we can build in. I think the cool part about AI to me is it's all super custom. Like, I'm, I'm, you know, to me, SaaS is dead. That's the, that's the AI hype for you, which is, which is you don't— I don't want to use the same product that, that you want to use. You know, why do we both use HubSpot for something? Well, because it's kind of the only option and it's somewhat customizable and whatever. But why isn't your CRM like the thing you dictated into the world that only you care about because it works perfectly for your business? I think this like mutual intro tool I don't think this is a thing that we SaaSify and make available to other people, but the process to create it could be SaaSified or run it as an agency and say, well, I'll make a custom one for you too, because I know how to build a custom one and I know how to build all these little tools. So there's something magical in knowing how to talk this whole new language, which is LLM prompt engineering, that whole space. And if you're really good at that, you can build really cool custom stuff that like doesn't have to scale to, you know, SaaS because it's something that only runs on Andrew Warner's computer or Adam's computer.

So one thing I would love to come out of this is to have somebody say, I, I would pay to have Adam teach me XYZ because I think that would be pretty interesting for Bootstrap Giants to add a course on this. Do you think that Someone who is not technical like you, who is maybe like me, having lived in San Francisco, that's the most tech that I have. Only an AI would think I could code.

Yeah.

I think someone like me with time from lessons from you would be able to code up my own agent.

Yeah, I think there's something to, you know, from my, from my kind of stint in Buddhism, part of the reincarnation journey is is you're, you're, you're building up kind of karma and understanding over not just a lifetime but many lifetimes. And you'll be presented with what you need when you need it, when you're ready. Like, enlightenment is already there waiting for you. You're not ready for it yet, so you're on your journey. I think people are on some version of an AI journey. You know, January 2024, I wasn't ready. January 2025, I was ready. Right. And then I saw this video. If I had seen the video in November, it wouldn't have worked, right? It wouldn't have clicked. And what I think is true is, so of the people that have reached out to me so far, maybe half are technical engineering anything, and the other half are like, they're salespeople or customer support or like administrative, like company leadership type folks. And those people— are ready for this. Like, they— like, the reason they reached out to me is like, I've seen these tools, I'm— I want to use these tools, tell me what to do. And I'll give them a list, I'll give them some advice. I have this thing that I kind of like send out to people now as my like AI kickstart. And what I find is those people, the non-technical people, they move faster than the technical people.

Mm-hmm.

I get— I get an email back the next day, the next morning, like, I tried it. Here's what I built. What do you think of this? I'm putting this together. The technical people are stuck for probably a lot of the same reasons I was stuck, which is like, I was trying to use code. I was trying to use old tools. I like, I had the wrong knowledge to use this new stuff. And these, it's more about like kind of being ready, I think mentally for how you would put AI into your business rather than knowing that the toolkit like this, the agent that I built that does the mutual intro thing, there's no code.

Yeah.

There's literally no code in it. If you could build Make.com workflows and do like decent prompt engineering, like how do you write good prompts, you could build the thing that I built. No code whatsoever.

All right. If— who are you— who would you want to respond to this and how and about what?

Yeah, I think there's a lot of like very small agencies or individual contractors in kind of sales ops, customer support, that kind of stuff. Those people should be spending 20% of their time building AI agents and workflows for the stuff that they're doing every day. They just don't know how yet. And for the ones that are ready, that are excited to do it and want to try today, I want to— I wish I had the video where they saw the video and they're like, oh, I could follow along with that and I could build this myself. Like, That's, that's who I want. It's not engineers. It's probably not CEOs. It's like someone who's trying to build right now. And they're just, they're building a bunch of make.com workflows, or they're, you know, they have a process that they give to their virtual assistant, or they have a process they follow themselves. Like, you could make that a lot better, faster, easier, bigger, and more reproducible.

And you want them to just email you?

I want them to email me. Like, yeah, at this point, I don't, I don't have anything to sell you. I have my like AI Kickstarter that has really good kind of responses. But I want people to talk to me. Honestly, I want something to sell these people. I just don't know what it is yet. Is there a class? Is there an accelerator? Is there a guide? I don't know.

And the more people ask, the more we'll figure out what that is. What's your email address?

My email is adam@gateway.xyz.

Adam, I gotta tell you, I was gonna end it there, but the thing that I would love is you just told a story. We, we went through it in story format with clear, like, milestones that people can learn from.

Yeah.

I would just love to not have to sit and type this out, but to find a way to share it as a story. And I keep trying to get ChatGPT and, um, and Claude and others to do this, and maybe I don't have the right prompt, but It doesn't even have to be a perfect story. If it's 80% of the way there, I could, I could edit it to, to the finish line. But that's where I'm really struggling.

This is really interesting. I literally have— I'll send you a Loom. I took a Loom yesterday as I was like running out the door where I was trying to walk Jesse through exactly this, which is you take a call transcript, a couple steps. I think it's 4 prompts deep. And, and all the prompts are generic. You could run them on any transcript. So like the first one is, What are the stories that we talked about here? What are the key takeaways? Just like, give me the stuff, you know? And it spat out some stories. Then you say, okay, take this one story and give me all the information about it. And then you run that prompt and it kind of cuts out like, okay, now the transcript is just this story along with some takeaways. Great. Now take that and turn it into a format that I want. And honestly, it worked really well. And it is, I think it's 4 prompts. And so I'll send you the Loom. I think it will blow you away a little bit on what it can do and how generic the prompts are. Like, that could be an AI agent as well. Like, every single call that you're in, oh, absolutely, run these 4 prompts in this order. You need a little bit of input, like some of the insight to still— these were the 5 stories that were talked about, like, What were the truly interesting ones?

Right.

Well, that's— your brain has to pick that, right? If you pick the wrong one, it's— your output's not going to be very good. But the tool does all the work to say these are the things. Okay, I want number— I want B. All right, take B and fill that out. Give me all the context. I'll send you the Loom. Let me know what you think. I think you'll like it. I think you'll want to use it for Bootstrap Giants. And Yeah, that might be a thing we, we actually, that might be one of the next steps. I would love that.

That really is—

So that AI can transcribe your call transcripts.

Just like the, the stories of our lives are happening, but I, I don't wanna write the first draft of the report on it.

Right.

Okay. All right. I like this a lot. We'll do more of these updates basically if you feel like it. And then if the audience response is, is, is good and we'll end it here. Bye everyone. Keep sending Adam and me some feedback.