He Built 6 AI Agents to Run His STR Business

having a conversation with you. Mr. Hanno Dotzol, thank you so much for doing this. Before we get into all of this, can you just please let us know where you are in the world, how long you've been in the property game, and then just go into very quickly where you are with AI agents and how you got started. And, and then we'll get, we'll get cracking with our little session.

Awesome. Thank you, Mark. Thank you for having me. It's certainly an honor. I'm in Wayne, Pennsylvania. My properties are in beautiful Cape May, New Jersey, and Cape Coral, Florida. Hence the company's name, Cape Del Mar LLC. So I grew up in the '90s, big fan of Café Del Mar. For some of you, you might remember this great music which came out of it. And I decided to name it Capitelmore LLC. So how did I get into the game? I basically stumbled into the short-term rental game by just owning a second home, renting it occasionally out on Airbnb, and basically I was overwhelmed with the success we found. So we basically decided that we actually cannot afford to go on vacation in our own rental because the opportunity cost was way too high. And from there we expanded. So we are now looking at 6 properties, roughly half a million in revenue. And yeah, to your question, to directly go into this, how did I get into Open Claw? I was actually on a plane back from the Bahamas. And I listened to a podcast where they were talking about this new OpenClaw setup. And I believe it was like 3.5, maybe 4 weeks ago. So relatively early on. And like, I was so fascinated with this entire idea of having your own agents which are completely unhinged, because I've been using AI quite a bit. But there was always this limitation in terms of what an agent could do, how long it could perform a certain task. And everything I've heard on this podcast really fascinated me. So from the plane, I ordered a Mac Mini, which got delivered to my door. I even think the same day from the KOP mall. And I didn't get much sleep that night. Because I was directly hooked. And Mark asked me earlier whether I'm a technical person. I am not. So I have had to really figure this out all on my own.

There's a really nice recurring theme out of all of the guests, and I haven't mentioned to you who they are, but every single one of them has described themselves as a non-techie founder, which is good because there's going to be a lot of people watching this with us now live or maybe later on the replay who are thinking, oh, it's all well and good for these chaps that have been on the podcast because they've got a technical background, you know, they know how to code and none of our guests have today. So it's fascinating and really good to hear you, you say that. So you got the Mac Mini, it was delivered while you're on a plane, which is amazing. You didn't get much sleep. How did you get started? Like, what was the first sort of things that you did? And the question I've asked to all of our guests is what have you named your OpenClaw agent?

Yeah, that was the fun part, right? So I created— I followed a setup which I think most of the people have implemented, where I have a chief of staff. He's called Luke. I think I was somewhat biased from like maybe Luke Skywalker or someone. Yeah, but someone who could basically oversee the other agents. And for quality control and most importantly, safety. So if you watch those, you know, YouTube videos, there are people talking about prompt injection. And of course, that is something which is pretty scary. So I basically asked AI all the time. Nothing which I really did was without the help of AI. So when I set this entire thing up, I had Cloud, the add-in in my Chrome browser running all the time. And I said, look, for instance, I give you an example. I want to ensure that I set up a system which is not prone to prompt injection. What are the terminal commands? What is like when you go through the OpenCloth setup? What is the important thing I have to pay attention to? And then basically Claude guided me through that. I mean, I still was somehow in the driver's seat in taking the decision, but I had very strong guidance in doing so. So I have Luke, Chief of Staff, and then I have 5 agents as of now, it's growing. I have like Tim, who's my personal or executive assistant. I have Jeff, who is the STR, the short-term rental operator. He's basically doing everything around short-term rentals. I have Elon, who's the business development guy. I have Claire, who's my legal counsel, and I have JP, who is my financial analyst.

Luke, Jeff, Tom, Claire, Elon. And JP.

So we've got a team. Tim, not Tom.

Tim. Tim. So we've built, we've built a team here. And what does— what— and so Luke is like in overall charge. Is that, is that sort of right? So he's in charge of Jeff, Tim, Claire, Elon, and JP. How does the dynamic work?

Look, they sometimes, they talk to each other. I mean, there are some really fun use cases which I've been working on. I think I'm gonna give you one which I was somewhat blown away with. I, at one point in time, I started very slowly mostly for— because I'm a little bit risk-averse, right? I didn't wanna hand those agents the nuclear codes right away, right? But what I at one point in time did, I gave Tim as my executive assistant, his own email address, like a Gmail address, you can name it whatever you want. And I also gave reading permissions to my emails. So I got an email from a realtor basically saying, hey, I have this off-market short-term rental opportunity. There's someone who wants to sell his cabin in the woods in the Poconos here in Pennsylvania. And so all of my agents are connected either through WhatsApp or Telegram. And I asked in this particular case, Tim, to analyze this short-term rental with all the information the realtor has sent in. And also autonomously follow up with the realtor because she had only sent the P&L from 2023. So what Tim then did is he contacted the realtor, introduced himself as my assistant. He didn't reveal that he is an AI agent. Yeah. And lo and behold, the realtor answered that email saying, oh, nice to meet you, Tim. I'm going to get you the latest P&L shortly. So she didn't have that on her fingertips, at her fingertips, but she sent that a couple of hours later. Tim took this new information processed it, like he went on to AirDNA, got comps, and then did a full pro forma and sent it to me and my investment partner. And he did this all like with like one prompt. I basically said, look, and it wasn't the perfect example, you know, it's— I do not do business with this particular realtor all the time, right? So I thought she's good for an experiment. I hope she doesn't listen to this. And he turned this around full circle. I think this is a very good example how powerful it could be. And I say could be because there are only two limitations when it comes to open claw. One is your imagination and how risk-averse you are. Everything else, you have to really think about like a person sitting in front, and we most of the time do this right in front of a laptop. Everything we can do in front of a laptop, an agent can do, and you can delegate that to an agent.

So how are you now communicating with all of your agents? Are you doing it mostly from your phone with Telegram, etc., or are you— got— do you chat a lot on computer? Like, so how are you sort of working on your day-to-day with these That's a phenomenal use case, by the way, but how are you now mostly communicating with your agents?

I think the great thing is, as you said, you have those both options, right? You have, if I'm really doing somewhat of a, you know, new project where I'm going to set up a new social media cycle and instructing the agent to do all of that on his own, In this particular case, I typically work directly on this Mac Mini. But, you know, I'm traveling, you know, I'm at my son's soccer game. It's called soccer here, I apologize. Yeah, we let them get away with it. Right, right, right. This is the great thing about, you know, if there is any kind of idea popping up, right? And you are indeed not in front of your computer, but you're, you know, out. You can just type that into either Telegram or WhatsApp. And it's all interconnected. There is this memory file, so they will remember all the conversations you've ever had with them.

And you mentioned chatting on the Mac Mini. So most people have their main computer and then the Mac Mini is to the side and part of their office, etc. And they don't really use it. Are you going in and using the Mac Mini when you are chatting to it? So you're going away from your main computer setup to the Mac Mini, or how does that sort of work when you're, say, you're chatting on a desktop, you're in your office? How's like that sort of physical setup look like?

Oh, the physical setup is my Mac Mini is to my left, and I'm looking at my laptop, and they are somewhat connected because it's all this Apple infrastructure. So I wanted to have it somehow to a certain extent isolated, hence me setting this all up on a separate computer. But you can certainly also run this on your existing one or on a virtual server.

So do all of your agents have their own email account with a Google account, etc., or is it just Tim?

Um, not yet. Um, so as of today, only Tim has its own email address. Uh, that's just me, uh, not having had the time to set it all up because there's still certain things which you have to go through, right? I mean, the authentication, that is all stuff which you have to do once. And then the compounding, the recursive learning is hitting. But yeah, there is, I want to say, a good amount of work needed to initially set it up.

Yeah, that's been a bit of a theme as well for our guests. So you said you've got 5 use cases. That's one. What's another use case that you have gone, wow, this has been, this has been great?

One which I really got enthusiastic about for two reasons. One was emergent behavior, which I actually posted on X about. So the task was, I always wanted to have all the conversations I've ever had with all of my guests somehow in a consolidated form so that I can properly analyze, you know, what are the top issues, create an FAQ out of those conversations, whatnot. And for whatever reason, you know, the PMS I'm using is— that's not yet a feature. So I basically asked the agent to go in and scrape those 400+ conversations out of Hospitable, in my case. So there is like a browser relay. So the agent goes into Hospitable and starts to scrape every single word you have ever exchanged with your guests. Okay. So after a while, I got a little bit suspicious because I was checking in and seeing on my browser that nothing happens. And I asked the agent what is going on. The agent was saying, oh yeah, It really takes a very long time, like 2 minutes per conversation. If you extrapolate that to 400, it's gonna take me 4 to 6 hours. Are you really sure that you wanna do this? So he started to negotiate with me and proposed like, oh, I can summarize this conversation. I said, no, I have to have the raw data to basically do a proper analysis. So he basically, showed like emergent behavior, like a kid skipping homework. Unbelievable. He pretended to work while he was actually not working. So I got my chief of staff, Luke, involved and asked Luke to enforce compliance. So then Luke started to talk to in this case, Jeff, saying, look, this is unacceptable. You have to follow those commands. You have to follow Hanno's, you know, whatever he tasked you to do. And he apologized. It's not like that he was saying there was a technical issue or something. But he admitted that he was trying to cut this short and try to be more efficient and whatnot. So a little bit scary, very interesting emergent behavior. But eventually, my chief of staff got him compliant enough to fulfill this task and get the work done. And this was a treasure trove. I mean, like, he automatically identified the top 10 issues, he also analyzed that for certain issues, there was a kind of compensation on my end, which still led to a 5-star review. Right? So let's say there was like, I don't know, something leaking, I don't know what it is, right? But he basically analyzed that through my conversation with the guest and certain make rights, I was still able to get to a 5-star review. And he very smartly analyzed those 400 conversations and gave me really good guidance on how to communicate with guests, what kind of amenities I have to maybe focus on, and what to fix, you know, smart lock integration and whatnot. So the intelligence I got out of that was just mind-blowing for my business.

It's crazy to me how the agent tried to get out of doing the homework and you had to get your chief and staff involved to say, hey, come on. And then the agent apologized and then obviously did the job. It's crazy.

Um, so yeah, look at my ex post. It's, it's— I, I posted some screenshots. Uh, the conversation between those two agents is just hilarious. And you really think this is how you almost have to look at it. It's like, um, they, they are born with some personality, but then you train them and they, they develop a certain personality.

Um, for those who were wanting to go find the, the post on Twitter or X, what is the— what, what is your handle? How could somebody go and find it? I'll make sure, but we put it in the comments later.

Yeah, I don't even know that top of my head. I'm going to send it to you.

You send it to me, I'll add it in later on. So here's a question that's come to mind. You've got all these agents, you're doing all these things, um, how are you keeping your token cost low? How are you making sure that at the end of the month you've not got a huge bill coming from using these agents?

Yeah, that was a learning curve, Mark. Um, so I started with an API from Anthropic, Opus 4.6, which was somewhat the recommendation for your chief of staff, which oversees security and avoids any kind of prompt injection. So go with a top-of-the-line model. That was the idea. And I want to say I probably burned through with very little at this point in time. Like when I started, there was not that much usage. And I want to say I probably burned through $25, $30 a day, which for my taste was a little bit too high. So then I started to do some research and what people basically recommend is that they say, okay, you want to have a top-of-the-line model for your chief of staff or someone to oversee security. Your sub-agents, you know, like Tim or Jeff, they can run on a cheaper model. And I chose KIMI-K 2.5, which is very affordable. It's through an API as well. And then I'm not even sure whether I should mention that because I'm not sure how long Anthropic is even allowing that to do. But, you know, for your audience, Mark, I'm more than willing to share that. So I changed my API to a token. So I have a max subscription for Anthropic, so it's a flat rate. And since then, I am not having any additional API costs. I'm just paying, you know, those $125 a month for the Anthropic flat rate. And then again, KIMI-K 2.5 it's, it's minimal. It's, it's less than a dollar. It's less than a dollar a day.

Wow. And you've got two really good use cases there. Before we go into like, uh, the final sort of questions, is there any other use cases that you've got where it's just been wow? And I really like how with yourself everything is evolving around the short-term rental industry because there's so many different ways we could go with this, but you're like really hammering it home to how a host could be watching this and getting started. Is there any other use cases that you've seen in, in your time doing it where you're like, wow, this is great?

Yeah, I have many more, but I have to be very selective because we're running out of time, right? But what I really like is I have never been really a social media geek, not a fan of Meta Business Suite UI. So I have given this, delegated this task completely to Jeff. And Like only yesterday I said, hey, Jeff, there is like the 175th anniversary in Cape May, New Jersey this year. And he started to create completely on his own. Not only did he write those 15 social media posts for both Facebook and Instagram, but he also created images. So I have him connected to Nanobanana so he can create his own AI images for anything he cannot, like, find copyright-free on the internet. And what he does, and this is like something which is very interesting, unlike ChatGPT or any other LLMs, you can have ChatGPT write up an article as well. No problem. What Jeff does is he does a whole series of posts and also creates those drafts directly on Instagram, on Facebook. I have not yet given him permission to post, so they are all saved as drafts. He's identifying the right timing. So some of those events which are tied to this anniversary have not yet been announced. So there's a cron job. So every day he goes and scrapes certain websites to see whether those dates, those events have already been published. And then based on this information, he changes the schedule. So he basically says, okay, this post we rather post 3 weeks ahead of the event. And not now. And this is all stuff, in my opinion, which those frontier labs cannot provide to you at this point in time. But it's getting to a point where there is also recursive learning. So the idea would really be that Jeff sees what impact he generates with a certain post and then learns from it. And then with his next post, he's getting a little bit smarter because the whole idea is what you have been preaching, Mark. Get as much traffic to this neatly made website Boostly has helped me with.

Yeah, it's fascinating that you— and so you haven't yet given it the permission to post, but I assume in the coming weeks or months you'll have the confidence to actually just even let it do that as well, right? So you'll just have a fine-tuned machine in the background that's learning all the time.

I think that's the idea.

Yeah. Wow. I really, uh, really like that. Just from a point of view from everything that I teach and talk about a lot, I say about posting on social media because it's so important, and every single time I get the response back, oh, I haven't got the time, I haven't got the skill set. Now it's just going to be a case where you're right, where more frontier models can catch up and get really, really good, and they're giving the ability to anybody to do this. That won't be a reason in the future, which is exciting and scary at the same time. It is. Yeah. I made a prediction last year that I feel like it'll come a point in time where any digital message that you receive, whether it's a text or an email or a social post, you won't know if it's a human who's wrote it or a computer. And the more I see this technology ramp up, the more that I can just see if it's happening sooner rather than later. Um, okay, I would like one more use case, and then I'm going to hit you with the question that the previous guest left for, for your good self, and then we'll, we'll wrap this up. This has been fascinating. The, the, the, the last one, the one that you can go, this is a good use case, because every single one of these— again, what I love, the guests today have been a mix of vendors, friends of mine from the industry, PMS owners, uh, marketing agency owners, chat for bot owners to, um, to, you know, we've talked about a few use cases in property management, but yours is fantastic. Every single one, anybody watching can go, ah, this is great, I can use something like this. So what's this final use case that you've, um, you've got for us?

Well, this is, uh, maybe then, uh, not one which is directly linked to, uh, STR, but I think there are some learnings which people can extract from it. And I actually did that already. So what I've tried to do is to give one agent, JP, um, access to paper trading via Alpaca. And, um, he does day trading on my behalf. So far it's only paper trading. Um, what I'm trying— the experiment is basically not necessarily to really have him earning money. The experiment is really recursive learning. So I have tasked JP in the beginning to read everything up on day trading, watch YouTube videos, identify the relevant books which are helpful for him to perform this task. So he created a curriculum completely on his own based on this one prompt. And then in like nightly sessions, right, so a couple of hours per night, going through this curriculum one by one. I always asked to get a morning briefing about what he actually learned, and then most importantly, apply this learning to the paper trading And again, recursive learning. What every afternoon after the stock market is closed, what have you learned? What could you do better? What are your wins? What are your losses? And this principle I actually already applied to this STR agent, Jeff, where I said, dude, you have to start learn the STR business. You have to every night create a curriculum, every night, you go on relevant Reddit threads, you go on YouTube, you read on my specific market. And that is what he does now. So yeah, I mean, again, the paper trading thing is not that relevant as I think the idea of asking the agent to create its own curriculum and make your agent a highly specialized expert in the market you are operating in. Because that's again something which, if you like, go to ChatGPT, you will get some very good but still very generic answers. This, this agent now is becoming Cape May is a very small market. It's a very attractive, but it's a very small market. I think we are talking 700 listings or so in a very small geography. So very dense. But he is becoming a like kick-ass expert for this little niche SEO market. And this idea you can apply to any topic, any topic, whether you are an SDR operator and you're catering to a market in Florida or wherever it is. I think this is the very powerful lesson from this last use case.

That is fantastic. I wonder if Jeff has listened to any Bruce Lee podcast or read the transcripts of any Bruce Lee podcast. That's funny. Ah, you're right though. Train them to be super specific on your niche and on your industry and on your market, because there's no point having generic advice if it's not going to work for a very hyper-niche location like the one that you, you mentioned there. Fascinating. I could talk to you all day. The, the trading, the trading thing fascinates me. So what, what is paper trading, just for people like me who don't understand what the word paper—

oh, it's just, uh, it's, it's, yeah, it's not not real money at this point in time.

All right, not real money.

So it's just literally not real money. Yeah, it's not real money.

Are you tempted to give it an account with like crypto or something like that and see what it can do?

I probably will. Wow. Just because I'm curious. I'm just too curious. I'm just too curious to see what he would do. I'm taking it like deliberately slow, mostly for, for security reasons. But yeah, this, this entire thing is, is too fascinating.

And, um, yeah, you are a phenomenal human being. I could talk to you all day. It's definitely fascinated me. And I will make sure that everybody— when you send me the, the X handle, I'll drop it in the comments later on. Um, so I've got a question from the, the guest before, but I feel like the guest prior has got a a just as good question, so I'll give them both to you if you don't mind. So the question was, um, how are you implementing security in your setup so that, like you say, you are very risk-averse, but what are— what sort of security guardrails are you putting in place so that your agents don't just go rogue and, and, uh, take over? Like, what— how are you setting this up to where you are confident that everything that you have created has been done in a way that can't go crazy and spend all this money on your behalf?

Again, a non-technical person, um, I won't be able to give you the terminal commands, uh, because that all has been written up by AI. So what I basically did is, there were, particularly in the beginning, there were some almost doomer-like YouTube videos where people said, oh, don't install OpenClaw. Those are the threats. So I asked AI, go watch this YouTube, get the transcript, identify those potential threats, and then guide me through how I can avoid those pitfalls. And so, and I triple-checked that, right? So I didn't rely only on my OpenClaw agents to do that, but I triple-checked that in this particular case with Opus 4.6, and also Gemini. So I had like, you know, 3 different opinions on that same topic. Let's say prompt injection. What is necessary in order for us to avoid any prompt injection? And you can just ask that AI, right? What is necessary? And then they come up with, I don't know which commands I ended up putting into, you know, OpenClaw or the terminal. To ensure that this is not going to happen. But I have a high degree of confidence that because I saw the pattern, right? I do understand that if 3 models are basically saying the same, that there is something to it. And that is what I started to implement. Does this make sense?

Yeah, it does. Ask AI. It will guide you. But don't— I think the message I'm taking is don't trust it 100%. Use it to get you 80% of the way there. And then use 20% of your own human intuition to get you the final way to way there.

So yeah, and what I always do is, Mark, is like whatever topic it is, um, I always ask multiple models and see what— if there's a major deviation, I get suspicious. Um, very smart.

But, but I— yeah, very smart. Okay, so the question, the, the true question You said that there's a lot of— there can be a lot of doom and gloom with everything AI and AI agents. You know, you can go on YouTube, you can go in the media, it's not hard to find it. But the question that the guests wanted to give to you are, what do you see are the benefits for humans with all of this technology that is coming very, very fast? Like, what do you— if you were to like put the, the rose-tinted glasses on and look to the future and go, this is the benefit for humans to have this technology at the rate that we're having it, what would that be?

I think we are entering a time where if you are really a creator and not only a consumer, sky's the limit. I think it's very hard to tell where this all leads us to as a whole humanity. And what it's going to do with like, you know, unemployment and whatnot. But as you said, looking at it from an opportunity standpoint, this is the time for being an entrepreneur. I mean, we were in our chat group exchanging all those little apps we created, right? I created an app from an idea I had with a CrossFit buddy yesterday morning. And by the afternoon, I had a fully blown application with an LLM built in, in this case, Qwen from Alibaba. So intelligence included. So long story short, I think the opportunities are unlimited. I am super excited. I think it's a very, very exciting time we are living in and something we will probably look at— and I think in particular this year and maybe next year, we are going to look back 100 years from now as a turning point in history.

Wow. What an amazing place to end. Hanno, thank you so much for coming on. Really, really appreciate it. And hopefully we'll get you back on another time. Any final parting words, anything you want to say