← Back to search

#125 Restart with 0€: I would use AI agents to make money

Doppelter Espresso · 2026-05-13 · 29 min
relevance 66 4997 words Episode page ↗ Audio ↗
Show full episode description
Die Mehrheit unterschätzt, was KI-Agenten wie OpenClaw und Hermes leisten können:In diesem Video erkläre ich Schritt-für-Schritt wie sie funktionieren und wo sie in der Praxis tatsächlich Mitarbeiter ersetzen. WerbungHermes bei Hostinger holen: https://www.hostinger.com/torben 💸 Code "TORBEN10" für 10% Rabatt ✅_____________________________________________Folge mir gerne auch hier:📲 Hauptkanal: https://youtube.com/@TorbenPlatzer 📸 Instagram: https://www.instagram.com/torbenplatzer 🤝 LinkedIn: https://www.linkedin.com/in/torbenplatzer/
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
How a solo founder could restart from zero using AI agents instead of hiring employees or paying developers.
Benefits
  • AI agents replace your first 1-3 employees
  • Vibe coding builds apps for one prompt instead of huge dev cost
  • Agents run autonomously on a virtual server, not your PC
  • Hermes adds strong memory and a learning curve over OpenClaw
  • Integration everywhere means AI without going to a chatbot destination
Use cases
  • ChatGPT cracks the 900 million weekly user mark while ChatGPT.com web searches fell 6.5% since the start of the year
  • Google Gemini app has 750 million monthly users but AI Overviews in search reach over 2 billion
  • Built a 1-to-1 Subway Surfers copy with Claude code from a single prompt
  • A custom messaging app once cost the host 150,000 euros (total ~quarter million) before WhatsApp made it obsolete
  • Grew from 0 to 10,000 YouTube subscribers; Ask Studio reads top comments to suggest content-gap topics
KPIs / results
  • ChatGPT 900 million weekly users; web searches down 6.5%
  • Gemini 750 million monthly users; AI Overviews over 2 billion
  • ChatGPT Go embedded on one billion iPhones
  • Old custom app cost 150,000 euros (~quarter million total)
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Now just you, your podcast and a delicious break. Off to Aldi Nord. For 2.99 euros you can get your lunch break rolling with sushi. Mmm, delicious! And for 1.99 euros you can treat yourself to a little ice time afterwards. At Aldi Nord you will always find what you are looking for. Sounds good? Then try the Snacktime Sushi Box from 205 grams for just 2.99 euros. Or Mucki sandwich ice cream 8 pieces each for just 1.99 euros. That's good for everyone at Aldi's price. Now in your branch. Aldi. Good for everyone. Double espresso with me and we'll talk about AI evolution. About agents, about external chatbots that are dying out. An outcome that I'm a little proud of, because two of my predictions actually came true. And we'll start directly with the first one: Integration beats destination. What does that mean? Current numbers. From May 2026, so completely up to date. ChatGPT cracks. The 900 million weekly user mark. But classic web searches, i.e. really on ChatGPT.com, have already fallen by 6.5% compared to the beginning of the year. Why is that? OpenAI has a deal with Apple. You probably noticed. This means that on one billion iPhones, ChatGPT Go is now embedded as a system layer directly in apps such as Spotify or Uber. And that's where these monthly users come from. It's even more extreme with Google Gemini, my favorite. The dedicated app, also now a web app or if you download it and so on, has 750 million monthly users. But AI Overviews in search over 2 billion. This is broken down and I found it really exciting. Namely Chrome Browser. I mean, you gradually notice that you're entering something somewhere. It somehow started at Google. You search for something and then this AI summary appears at the top. But is it more than meets the eye? So directly in the address bar, in page usage, in Autobrowse, in Native Integration. This means that with Gemini Nano the AI ​​sometimes even runs locally on the computer, which people don't even know. Gemini offers functions such as proofreaders or combining websites. Then in Google Workspace, that is, in Docs, in Gmail. When you write an email and you have a Gmail account, the completions always come, then sometimes the whole sentence is already there and so on. In Sheets, in Slides there is every time as a side panel, hey, do you want to compose the email with Gemini? I would even say on Ebay. I don't even know what's behind it on Ebay, but they already have this: you no longer have to describe the item yourself, the AI ​​writes it for you. Then Android 17. That means Gemini is there too, deep in the core of the system for complete contextual awareness. The AI ​​knows exactly what is happening on the screen. Now when you watch a video, why you watched it, what topic, what article are you reading and so on. That's the key word again: personalized advertising. Control Panel. Gemini takes over the tasks of the operating system. From organizing the home screen to intelligently controlling notifications. Google Trends Analytics. These search terms are increasingly appearing in the trend browser. If you are a creator on YouTube yourself, thank you again for all the very, very positive feedback from this 0 to 10,000 subscriber following, then there is now Ask Studio. And that's basically, read your comments, your analytics and then give you suggestions. And I'll be honest, something like this happened before in the beta and I could mostly forget it. Then he just looked at what else the audience was watching and then gave you a similar topic. But now, Ask Studio, which I think was rolled out in my account a month ago, is very, very accurate. So he reads the top comments, looks at what people are commenting on, what discussion arises and then says, hey, I just see a content gap here, I also liked looking, are there other people who are talking about it? No, then he suggests that to me and in the description it says, yes, people exchanged ideas under your video, the discussion was about that and that, so you could make this the topic, it also fits in with the search queries, which are currently higher again and so on. So really, really good, yeah, topic suggestions from Ask Studio. I could go on forever, my point is that when you go to Gemini.com or ChatGPT.com, that's what Destination means. So you explicitly go to a chatbot because you say, I want to communicate with it now. Then you throw away a lot of what is still possible, plus, whether you like it or not, there are probably one or two people who say, woah, hey, AI, I can't hear it anymore, you're going to use AI and are probably doing it at the moment, don't know it yet because it's already embedded everywhere. So, a little fun fact here, for some people their YouTube views are going down. That's, maybe I'll do another episode of my own, just to explain it briefly, if you've previously watched a tutorial or something like that, it was a common method to go to YouTube and look, okay, are there step-by-step instructions, I'll just tell you how to assemble the Markus chair. Okay, and then people ended up making a video, eleven minutes long, then you went there as a newer YouTuber, did the whole thing in seven and were then able to replace the previous one as the top video, so to speak. So common strategy. Look at what works, what is shown above, make it shorter, just make the title and thumbnail more entertaining and that means it is almost always better, i.e. in the display. Now you have to think carefully about which topics is this still the case? I'll say 60, 70 percent, but at 30, 40 percent people stop watching the video because they get the summary from the AI. Or even if you say yourself, hey, I'm drawn to, I don't know, this podcast is, I think, when I get feedback like that, a lot of people have it in their ears, go running, go training, as inspiration and probably just like hearing the voice. But now if it's just about, you have a video, it's 40 minutes and you form the information. Say, hey, I just want to know, which five stocks is he buying now? For example. Now not me, but someone. Well, then you give that to Gemini and say, summarize it for me. And then you have the summary. There are already plugins from VidIQ, for example as a browser extension, until Gemini uses them again. Well, and then he just tells you the five stocks. Then why should I watch the entire video? That's why some people's views are going down, which is why you can now do a few things when changing your strategy so that you aren't affected. But let's summarize it like this. However, the majority still use AI without, let me say, consciously exhausting it for use cases that bring something. A lot of times you're just playing around with something, like, oh, that's cool, I can create a few pictures, but what's the point? And I want to go in there now. And that kind of brings me to the second point, which is agents. So having a buddy who brainstorms with you is a chatbot. This already worked with ChatGPT. Okay, if this is now integrated everywhere, then I no longer need to go to the destination. Then I can just wait, for example on Google or something, until I get the summary. Okay, there's the mainstream. The pioneers are now starting to train agents. So no longer brainstorm machines, this assistance within ChatGPT, your own chat that knows you, but an agent. And that is incredibly exciting. So Real Talk at the point, as a one-man company, that's how I started, I was for a very, very long time, is this a money glitch. And that's not an exaggeration when I say you can replace your first 1, 2, 3 employees and you have access to things, keyword vibe coding for example, that you never would have had. So imagine you want to make a web app, which may well make sense in this day and age. Just two years ago I uploaded experiments to YouTube, where I buy them from Fiverr. Where on Fiverr I say to an Indian guy or something, hey, build me an app. I've had one developed myself and it cost 150,000 euros. So it was complete nonsense. Back then it was a communication app. Before WhatsApp had certain features, I had it done with my two best friends. Put in by a developer over 150 from me alone. Total probably a quarter million or so. And then WhatsApp released certain features, such as forwarding or deleting messages. And then it was obsolete. We wanted to create a messenger like this, especially for people in sales, where you can, for example, withdraw a message. So, and then the feature came, the app was gone. Then, sure, you could now get a game or whatever on Fiverr. That's just one prompt now. I've already shown it in videos, like the main channel video, the penultimate one, where we simply built Subway Surfers with Claude code. And the prompt wasn't somehow, use such and such language and this and that, but create me a 1 to 1 copy of Subway Surfer. And that's crazy. So, the agent can now do a little, a little more. And I want to show you that. And therefore a shoutout to our long-term partner Hostinger. They not only sponsor this episode, but also many, many of our experiments in the field of AI, because they are always the first to go out, whether it was OpenClaw or Hermes, and say, okay, here is a one-click solution for the mainstream. This means that when you go to the site, you not only have all the tools you need for the foundation, keyword self-employment, for example hosting the site, AI construction kit, yes, for the site, for your web app, for the e-com shop and so on, but also virtual servers. And that has two direct advantages. I'm about to go into full detail about how an agent works and so on. Also for anyone who says, I may have heard of it, but explain it to me in simple terms. So first of all with an agent you have the option, you can download it, it's open source, so we'll talk about OpenClaw or Hermes. And then you have to install it in your console, for example. It's a strange feeling when you don't know anything about it. So most people have probably never had a console like this open and typed CMD or something like that. I say, can you still get into it? For example, you go to the OpenClaw site, copy that, put it in, then install it, then sometimes it says, I still need this and the package. But you can scurry through it. The other thing is, besides this, well, I don't know, I have no IT skills and it feels weird. The second problem is that the thing is on your computer. Means it only works when the computer is on. You may be a little restricted when he does something because he's just using resources. And, well, of course, you can tell him that you can use this and this and this, but you are not allowed to look into these files here, for example. In my experiments with OpenClaw, it happened that he did things where he didn't have permission. So, of course it is on your computer and can theoretically see all your data. And it's too tricky even for me, I'll tell you honestly. There's a little Slack in there, which was available straight away with OpenClaw. Some hackers get on my computer, can pull out some, I don't know, seed phrase or whatever and empty my crypto wallet. Now, for example. And a virtual server means you go to Hostinger and they host for you, for your own, so that's not the case, you don't share it with others, but you have your own virtual server. So. And for that, I'll just show it here or just show it, you go to the page, provide Hermes Agent. KVM2 plan is preselected. For most things it is completely sufficient. Now you can use discount code Torben10, support us and save another 10% on annual plans. Then continue to register and now it takes 2-3 minutes and then this virtual server will be set up for you. You then get a user and password, for example select Quick Setup, I would recommend it to everyone, it's really simple and now you only need two things. You can then link an LLM, for example Opus 4.7, with an API or what I would recommend, the OpenAI subscription. So we almost got through with our subscription for the last main channel video and then we don't have to buy any extra tokens anymore. And optionally, where you want to talk to him. So in this mask, so to speak, that you see now, or do you say, hey, I want Discord, WhatsApp, Telegram. I always use Telegram because I don't have any other contacts there and then I can control it from anywhere and if you give it a task, it runs on your virtual server. Not on your computer at home, but on the virtual server. If he then needs to act, so to speak, with data or whatever, you can easily download it and then have it on your PC. So to put it simply. Now the topic of the AI ​​agent itself. Compared to a bot, this is an autonomous software system that pursues goals. So it's not, hey, I have a question, he gives an answer, but I give a goal. To achieve this goal, he plans tasks, makes decisions himself and looks for the tools he needs. That's the difference. An AI agent is proactive. If you tell him, I want this and that, of course he asks you a few things, I'll show it in practice right away, but he decides and he looks, what is the most elegant and best way to get there, so to speak. At the same time, especially at Hermes, which is the USP of the agent, I also add, he has a very good memory and an integrated learning curve. So the opportunity to learn. This means that if he sees that something isn't working as expected, he will make adjustments and not make the mistake again. That was kind of the problem with OpenClaw. But let's go through it step by step now. Point 1, goal setting and perception. So I pretend and he perceives it. To do this, it calls up data from the environment. For example databases, the web or APIs. So APIs, if you have linked them, for example, an LLM or something like that. I say to him, day trade for me. So just the goals that I put in myself. Then, because I sometimes travel a bit on the Solana blockchain, there is, for example, when someone creates a Solana token, the first one that was created, which is called that, is often the one that has the most value. So I say to you, what was the first token called so-and-so? I'm not taking a name now, not that someone will then create a meme expert and somehow make money out of it or something like that. But that's harder than you think. So it searches the entire Solana blockchain for the creation date and some of them were created two or three years ago. Or, as we've already done, which NBA team will win today? And where can I get the highest quota? So first of all just the goals now, that's step one. Help me buy Aventus Creed with the highest discount code. Shipping to Germany. Check for yourself whether there is a script from me in a file that I share with you and only call me if there is an error in it. We have already tested it. Or, I've done that now because it fits the YouTube episode really well, tell me the top five reasons why MrBeast is going viral. So. If you say that to ChatGPT, why is MrBeast going viral, then ChatGPT looks for secondary sources on the internet, for example blog articles, I don't know, imagine if I had a social media blog, then the, ah, social media blog looks for Torben Platzer and gives the reasons. But that means that if I tell MrBeast, then these are the reasons that he is now telling here. The agent does it completely differently. And I had him break it down. Step two is thinking and planning. So reasoning. For example, he is now using the LLM for this purpose. So it can be Opus 4.7 or Codex, i.e. ChatGPT 5.5 and so on. This is the brain. And the brain breaks down the goal into small intermediate steps. I think it's relatively understandable. So, I use this with my OpenAI subscription. Great job, yes, we had to buy more. All I can say is whoever uses Opus 4.7, token costs are immense. Step three. Trading and tool usage. So that's the corner thing. And that's also the moment when he often says, if you haven't connected the tools yet, hey, I need this and that. So I need access, for example to a search engine, to an email program and so on. Yes, okay. Now for the examples. Well, just now said, he demands what he wants. Here, that was still OpenClaw, he said, I first need Notion, where your script is in there, at the prompt. I need Notion and I need Twilio if you want me to call you. Then I connected it and then it worked. Other video, the odds API for this betting story. Also got them for free. Then he says exactly, hey, look, just do this, this, this, connected and then he has the odds. Hyperliquid. Trade-If for me, the last video now, regardless of whether it's Hermes or OpenClaw, we've already done it with both. Trade-If for me, at Hyperliquid. Then he also says it has to be connected to your account. Then he checks it and says, sync, out, works, hook. Or, no, it's not possible yet, I need something. So that's first, I'll get the tools. And now I have it on MrBeast, because I can also show you there, his answer, the top 5 reasons, is a lot more well-founded than if you do it on ChatGPT, or that's where it is, depending on how good the blog article was. As if I said, tell me exactly how you did it. So what did he do? Point 1, I have loaded a relevant skill, namely the YouTube Content Skill. I'll do this for myself briefly here. You'll see it all displayed anyway. This describes how to read transcripts and metadata. Step 2, he went to the MrBeast uploads and queried them from the official YouTube RSS feed. I didn't tell him that. I just said what are the top 5 reasons. So he himself, so to speak, determines that. He says that you get the uploads, title, video ID, link, published, date, description, views, likes and so on. So how they perform. Extract. Title structure, strong money buzz, challenge, mechanics, visual clarity, who wins, what happens and so on. So look, what words are used, yes, what is the trigger point, so to speak. And then he bundled together common viral mechanics. Top 5 reasons, clear premise, high stakes, simple game rules, strong visual hook and built-in retention-open loops. And there's this built-in retention open loop. To do this, he has to know what is happening and when in the video. Keyword cliffhanger. When does Mr. Bison speak cliffhanger? He evaluated this very carefully through the transcripts. And you can get an answer like that if a secondary source, including ChGPT, if a secondary source did that. But this is current, I can ask about it at any time. And I can now say, for example, whenever Mr. Beast uploads, I want you to analyze and tell me straight away, did he do something different? How many cliffhangers does it have? At what minute does he have it? And so forth. So this goes far beyond a bot, because the goal is, I want the success factors, I always want to have the current ones. So. Now the comparison, Open Claw and Hermes. I would like to say that Hermes is my current favorite. But Open Claw also has a few advantages. Namely, a large ecosystem, more integration. That means you have, it's just been around for a long time, it's been in the hype or trend or whatever for a lot longer. There are more community skills that you can download and simply more connections where tools have already been released, so to speak, or perhaps an extra interface to Open Claw has been created so that you can work with it. So it's still, if you take it that way, more mainstream. The big drawback is this learning curve. So theoretically, if Open Claw solves a task and I don't know, let's say it's a complex thing, i.e. solves a task within a goal and it takes him 30 seconds to do it, then the tenth time it will probably take him just as long as the first time, except, and now there are two exceptions, he can spontaneously create automations, crone jobs for it. Or crone jobs, they are always only done by the user, says Wiederkerne, but he can still create such small automations or he can also code a second agent for it and so on. But it happens spontaneously, which means you don't know whether he does it or, as I just said, you tell him explicitly. So, hey, look, this and that happened, you now have to save it so and so so that you can be faster next time. And that's actually exactly the gap where Hermes goes. Hermes has a system in the background that constantly checks in a loop, so to speak. Yes, so save everything and then self-improve. And that's what it does when you talk to him, he's just more steadfast, more consistent and, I don't want to say smarter, but when he explains something to you and you read it and say, okay, that makes sense and you attack it, I've tried that a few times, like, woah, I don't know if that's such a good idea. With Open Claw he gave in very, very often and always said, that's right, you're right, I'll do it differently. Although perhaps my request, I want to check again briefly, you're sure, wasn't right at all. So he was on the right track and I just wanted to unsettle him, that worked. And of course that gives me a bad feeling as a user because I think, okay, does he actually know what he's doing? And I challenged Hermes and he always defended himself. And then he said, yes, no, that's because, and then the reason came and it was consistent with how he explained it to me before. And this trading challenge, I don't want to spoil too much now, just watch the entire video, you'll just notice how theoretically, at least in terms of argumentation, it gets better day by day. And that's worth a lot. Today, three or four days after the end of the challenge, I said to him, hey, how was day one again? Day one. And he knows exactly that, yes, on day one I made such and such a mistake. You know what? Then you said to me, hmm, some OC sometimes loaded it and said, unfortunately I don't have any of the data anymore and so on. Yes, both are user-friendly if you do it through hosting. Yes, if you have a VPS, you can always use new agents for it, so to speak. But I would say for most people Hermes makes more sense because he is much more pleasant to work with. So, I would also like to include the negative point of Hermes, which is a bit more conservative. And that's, I thought at first, maybe it's because of the LLM, but they have their own, so with Open Claw it was, there's a Soul MD, so to speak, his soul, his character. And Hermes has something like that too. And Hermes is, when you say, trade for me, here's 1000 euros, then he tells you ten times, hey, I don't really want to do that, are you sure? The money can be lost. And Open Claw was like, yeah sure, what should I do? Is everything red or black? So, sometimes I even told him that you can only lose 25 dollars per trade, then he loses 50, I say, hey, what was going on? He says sorry man, I did a mistake. Well, that didn't happen with Hermes. Well, this self-learning ability makes it more pleasant. So, more consistent, more attentive, more steadfast. Now, the final word, and that is very, very important to me. This isn't supposed to be an episode like you often see or hear on YouTube, I think, hey, get an AI agent, but think about it, I've now shown you a few use cases and I hope they've explained them in a reasonably understandable way, think about what you want to do with them. Yes, the agent is just, yes, no longer as good as the prompt, you can't say that because this prompt thing is no longer important. Anyone can, you can phrase it however you want, you may have to write it the second time, but at some point you will check what you want. But what could you use it for? So, they can program, they can set up automation, check the top news every day, summarize it for me into a topic. Alert me when you connect Twitter API when such and such topic is trending. Write me new scripts based on this script. So in the agency we use this more and more. So, I would say in the last few months, probably three times as many clients have come to us and said, hey, let's set up the agent. Well, we had already suggested this to them three or four months ago, there were still a lot of people like that, yes, I don't know exactly yet. And now, everyone wants that. And this should also be said at this point, as I said, this is not an exaggerated statement, but an agent replaces one or two employees. and that's how it is. Yes, depending on what you do, you might have to buy tokens again, but you get tokens for 80 to 100 euros, which now cost something a month, which you might buy again, in addition to the subscription, you don't get an employee for that. That's a fact. Just blindly, yes, I have it now and what did Torben say, come on, I'll calculate who will win the game today, that's of course nonsense. Yes, that's why, you have to know exactly, I use it for my scripts, for research, for automations like find me out or get in touch if and all the things that I've just presented here, so to speak, for presentations, for web apps, we're rebuilding our shop with it, yes, then for example Hicksfield connection, then now cloud code, Hicksfield works great, you can also put an agent in between. Well, then it's not me who goes to Hicksfield again in the last episode and creates the videos and then uses Cloud Code for the web app, but rather I tell the agent, build them, he looks for what are the best videos for conversion and then controls the others again. So, the meta level, so to speak. I'm very curious to hear what you have to say about it, feel free to write in the comments, which of you has an agent at the start and if you don't have one yet, I can really only recommend Hostinger, I'll say it freely, we use it for all of our experiments that have something to do with AI, already at NNN, OpenClaw, we have our hosting there, at Hermes, so we have everything, we used the web, the AI, website builder, I've shown it here before, so it's simple good for your own foundation, especially for people who have a one-man show, for me it's the money glitch. In that sense, see you next time. I'll make another one now. And then I can get another one now. And then I can