Build AI Agents by Voice, Not Workflows

00:00:06 Mark Smith
Welcome to the MVP show. My intention is that you listen to the stories of these MVP guests and are inspired to become an MVP and bring value to the world through your skills. If you have not checked it out already, I do a YouTube series called How to Become an MVP. The link is in the show notes. With that, Let's get on with the show. Welcome back to the MVP show. Today I'm joined by Ragnar from Hesse in Germany. I think I pronounced that right. All the links for the discussion today will be in the show notes. Ragnar, welcome to the show.

00:00:47 Ragnar Heil
Mark, thank you so much for having me.

00:00:49 Mark Smith
Good to have you on. I always are surprised at the people that I come across in the MVP program that have been in a long time. You've been like 9, 10 years now.

00:00:59 Ragnar Heil
Yeah, exactly.

00:01:00 Ragnar Heil
Some years, now the renewal phase is heating up. So it's going to be my 10th year, absolutely.

00:01:06 Mark Smith
Yeah, end of the month, right, as things kick over.

00:01:10 Ragnar Heil
Apps and services and also Copilot.

00:01:12 Mark Smith
Yeah, yeah. Okay, so tell me a bit about food, family, and fun before we get started.

00:01:18 Ragnar Heil
Yes. So I would say my favorite food since I was literally born is maybe pizza. But now, since a few years, I'm a  lover of sushi. So everything. So I'm addicted to sushi. So I have to eat it like once a week. But I also prefer a nice steak. Yeah. So and I love traveling. So traveling is one of my biggest hobbies. And then, of course, when you're traveling to beautiful countries on this planet, eating, visiting friends and restaurants. So these are my favorite topics. I love to travel with my family. So I've got one wife, three kids, and my happy place is just to be on vacation with my family and discover this beautiful planet.

00:02:00 Mark Smith
I love it. I love it. I'd say I'd be in the same camp as you.nTell me, we're in month three already of 2026.What's top of mind for you right now? What are you doing in the tech space? What's taking your time?

00:02:18 Ragnar Heil
So, yeah, so privately I spend far much time working with OpenClaw. This is my new favorite agent. I have a hate-love relationship with OpenClaw because on the one hand, it's a very fascinating autonomous agent, but it takes so incredible much time to keep it alive because every second day, he or she, whatever, is doing suicide and I have to just repair and fix it. So that was my last, like, Yeah, let's say six weeks working with OpenClaw. And professionally, yeah, I'm mostly spending my time currently in customer workshops around Copilot, Copilot Studio agents, plus governance, compliance, and security. So that's my sweet spot. And mostly I'm running between, let's say, like two to five days workshops with my customers around agentic AI.

00:03:15 Mark Smith
Tell me a bit more detail about OpenClaw. How did you come across it, and how has it changed the way you think about agents?

00:03:29 Ragnar Heil
In my presentations, I talk a lot since, let's say, two years about the future of autonomous agents, but I never really saw them with my copilot studio agents. They're not really...autonomous, most of them are retrieval agents, some are going to be triggered and can run some power automate flows integration.But the fascinating thing about OpenClaw is that I really use my Telegram client. I mostly talk into my Telegram client, give him some instructions. Let's take a simple example. I talk into my Telegram client, please create a market research or a competitor research every single day about topic ABC, and please send it to me as a nice summary each day at eight o'clock. And it immediately does a nice research, immediately creates a scheduled job, like a crone job. And yeah, so when I say, listen, please add some MCP server integration to get the data, I just paste the MCP server name into it, an API token, and then it just works. It just works, and it takes me much more time when doing the same procedure in Copa Studio, because there I need to build my Power Automate flow. I need to create a connector for MCP. Many, many manual steps, and here I just literally talk into my microphone of my Telegram client, and then it runs. And that's the core and sweet spot of working with OpenCloud.

00:05:05 Mark Smith
Interesting, interesting. How much... Man, so this just opens up a conversation I've had for a while, which is, you know, I've been 22, 23, 24 years maybe now in... and my background was Dynamics CRM, which then 10 years ago became the Power Platform of which spawned Power Automate. And I just find that Power Automate seems to have so much overhead now compared to what you could do if you drop down to PowerShell or CLI and and just script exactly what you want. The automation, rather than having all this overhead of plumbing, I feel, that's designed for every scenario on earth, but makes it very difficult to do my scenario or your scenario.

00:06:05 Ragnar Heil
Mark, that's the point. That is exactly the point, because I know how to work with Power Automate. I can build flows, yes, butIt is so much easier to literally talk into a microphone in my Telegram client and then it just builds it for me. Because sometimes I'm just lazy. I have an idea maybe in the night. Now it's 10 o'clock in the night and just have an idea. I just talk into the microphone and then it builds it overnight without any effort for me.And this is just the crazy future where I just can put my brain drain ideas into a little smart client and then just builds it.

00:06:47 Mark Smith
How did you deploy it? Like what's your infrastructure for it for OpenClaw?

00:06:52 Ragnar Heil
I've tested a few things. The most reliable currently is running OpenClaw on AWS EC2 server, which is a T3 small, I think. that's the Ubuntu server size. It's pretty cheap. I think it's going to cost me like 10 euros a month. But I never would recommend to use OpenClaw on your own productive environment.So please don't install OpenClaw on your PC, MacBook. Just use a free empty Mac Mini, a free Raspberry Pi, you know, laptop. Use a VPS server, Amazon Web Services, very easy to install. Much easier than Azure, by the way. I would love to use Azure, but it's much easier on AWS. But don't run OpenClaw on your proactive laptop because it literally could delete your whole C drive without your command. And that's dangerous.

00:07:50 Mark Smith
Interesting, interesting. Tell me about, what's the name of your OpenClaw.

00:07:57 Ragnar Heil
Ragstar. Open Ragstar. Ragstar.

00:08:01 Mark Smith
I love it.I love it. And you mentioned Ubuntu, so you've deployed on Ubuntu server. What are you doing around models and the models that you're using to light up your OpenClaw?

00:08:17 Ragnar Heil
Okay. I've got an Entrophic Claw Pro plan. So I started with my Entrophic key. Then, you know, the history. OpenClaw and then traffic had a little fight. First, they started to legally fight because of their close proximity naming, because Claw and Claude. So, yeah, Peter Steinberg, Peter Steinberger from OpenClaw had some issues with the legal team, and then he renamed it from CloudBot to finally to OpenClaw. He talked to some old man to it, and then so them improved. And for this data, he acquired with OpenAI OpenClaw. And then after the acquisition of OpenAI, then Anthropic said, Listen, please, dear users, if you are going to use our credits, our tokens, then we are going to ban you. And then I immediately had to move away because I don't want to get banned from Anthropic. And then I just uploaded, I think, 10 euros to Deepseek. And then I'm using now the Deepseek tokens, which are very, very cheap, very budget friendly. And I'm using them. Technically, in the OpenCloud configuration, you use OpenAI custom provider. So there are two OpenAI settings, the original one, and then the OpenAI custom provider. And then you enter the DeepSeek one, DeepSeek chat. I think that's the name. And it's very, very cheap. So I just paid 10 euros, and it still works. That's much cheaper than buying the Opus 4.6 from Anthropic or other premium models. And for me, it's just sufficient.

00:10:02 Mark Smith
Yeah, so you can still use Anthropic through the API. They're just not wanting you to use it through OAuth and direct.

00:10:09 Ragnar Heil
Yeah, you have to pay for the API, like an pay as you go, but they don't want you to use your included tokens in your pro plan or in your max plan. That's what most of the people use. And if you want to be very open and you want to play around with different models, then my recommendation would be to go to OpenRouter, upload a bit of money, because then OpenRouter is also supported by OpenClaw, and you can choose between many different large language models. You just stick with your OpenRouter configuration, and in OpenRouter, you just change your large language models.

00:10:44 Mark Smith
It's interesting, as in, I haven't confirmed this, but I thought that Anthropic might have backed down from its hard stance on using your max credit or whatever that you have.
 I haven't checked it, but I just heard a few things in the last couple of days that they're kind of walking back their position on being so hard-nosed on it. I'd need to confirm that. Interesting. So Telegram's your main interface.

00:11:15 Ragnar Heil
Yes, exactly. Because the API is just richer than WhatsApp. So that's the reason why I use Telegram. WhatsApp is also technically possible, but it's just a few more steps. And Telegram is just very, very easy to configure.

00:11:30 Mark Smith
Yeah, I'm happy not to use WhatsApp just by very nature, be known by Meta. And I don't want any siphoning going on where I agree Telegram is a great solution. Now, what type of You know, what type of work did you do around things like memory and how it handles memory and also Sol? You know, it's DNA type thing of OpenClaw.

00:12:00 Ragnar Heil
So currently I'm just using... Okay, so I already told you a few minutes ago that my OpenClaw is dying every second day, so I have to reinstall it so many times. that my memory, which I just added, I haven't installed lately. So currently, it's running with just default memory settings. And that's actually the biggest pain point. I think I know I do everything right, but the gateway stops to work. And so I need every single day login into the AWS console, run open claw, dock draw, fix, and just fix it. It's just very time consuming and that's the most...So please, if you are really, your listeners, if you're just afraid of working on consoles, if you're afraid on running Linux scripts on Ubuntu, don't go to OpenClaw because it is really nerdy. You have to jump there and run a lot of shelf scripts, of course, which I get created by my large language models. So I don't write this script, but it's very, very manual work and very difficult. to maintain these agents. And regarding the use cases, to be very honest, sometimes I'm playing around with some innovative stuff like poly market trading. I've uploaded a bit of money in poly market to see how betting works. I don't use a lot of data, so I use like actually zero data, which is confident. I never use a customer data. I never company data or PII data. I just have a lot of non-PII data in my Obsidian and then it access Obsidian and I just do some studies around the content there. But it's not, I didn't use it very heavily because in other systems like Copilot, you know, there is my knowledge. Copilot has my access to my exchange and OneDrive and SharePoint, Teams chats and Teams channels and so on. So there is the wisdom. OpenClaw is just limited. I just use it because it's technically very fascinating, but there is not so much wisdom inside. Actually, I play around with some skills, but it's more for me the interest to understand what OpenClaw is, how it works, but I don't give too much data into it. What about you? You have a background in CRM. Would you give OpenCloud the access to an Dynamics MCP server, or are you also very afraid?

00:14:36 Mark Smith
So no, I've not given it access to my M365 anything, so therefore it can't access anything on that side. But What I worked on in the first couple of weeks was getting OpenClaw very stable. And so I have set up a lot of scripts that, for example, I have a nightly run script that updates Ubuntu and make sure it's always on the latest build and also all security patches are applied. I have another script that runs like Cronjobs that update the OpenClaw build based on as soon as the new build's available, it applies it and then it runs a routine to validate. So by the way, anytime I bring in anything that's external, it goes through a rigorous security scan agent that I've set up. So I do a deep analysis and I ground it on, it must look at what's happened in the last 72 hours from a security out in the public domain and it validates that. Mine's totally locked down. I've deployed mine in Azure. So I went through the process and I have Ubuntu set up. I originally built it on a Windows machine in Azure, and Windows is just horrible for it. It's just, you know, and so now I actually have it deployed on the server. but I've also got a deployment that is a node. So a second, a open core environment that you can set into what's called a node configuration, and that's on a Windows laptop. And so that gives me the hands to do stuff because obviously my Ubuntu service is headless, so I don't have a browser or anything like that on it. I've scripted security scans that happen regularly. Every half hour I have a heartbeat run and it is constantly, if it identifies a problem, I've set it up to proactively fix the problem. So I've now gone maybe a week and a half, maybe two weeks where no crashing, totally stable.

00:17:00 Ragnar Heil
Oh, I need this script. I need it because now... To be super honest with you, I'm losing my fascination with OpenClaw because I literally have to fix it every single day. It takes me, it doesn't take me more than 50 minutes, but it's still this 50 minute extra time around maintenance. I don't want it. So it just needs to run. I would just want literally.

00:17:21 Mark Smith
Yes. And I think in six months time, it will absolutely be that. And like I've seen a lot of people go, oh, wait for six months. But for me, I'm intrigued because it's taken me further into understanding agents than I have ever been. Like, you know, understanding why we have a user.md file, why we have a sol.md, what does the heartbeat do? How does memory work? So I've built my own 3-tier memory system that has short-term, mid-term, and long-term. And I got it to build it all for itself, but I often start all my prompting with, a research piece. So I go out and say, Okay, look at all the blogs and what everyone's saying on it, and how are people solving memory problems? And then I really build that, and then, excuse me, what I noticed, it started to write stuff into the different MD files that shouldn't have been in there, duplicates. And so then I put in a rule that says, You can't write anything to this file unless it sits within the definition of what the file is, otherwise you have to ask me. And so then I'll question and go, Well, why are you putting it in there and not in there? I've used GitHub more than I ever have in my life because I love how with MD files, everything's just human readable, right? I can go through, that makes sense to me. So I love that. One thing I've done is on that node, I have Claude Desktop installed, and in Claude Desktop, you can connect to MCP servers. So I then got it to build its own MC server and MCC bridge, because what was there, it looked at the code base and said, Hmm, it's not robust enough for what we need to do. And so it built that all out.I said, I'm learning all this, I've never done it before, and I'm loving it because I'm learning how to solve all these problems.

00:19:30 Ragnar Heil
That is maybe, maybe, maybe, you know, we are in a very, very early stage here. We are going to have a completely different conversation in one year. But now for me, the thing is we are learning so much within just a few days and a few weeks. So I think I never learned so much around Linux scripts than in my whole life, because I'd never planned to learn so much about scripting. Linux, running Ubuntu servers, running them and getting and keeping them alive. But I need to learn it to get this agent running.

00:20:02 Mark Smith
And here's the thing. I feel as the further we go down on our AI journeys in the years ahead, that the interface or UX as we've known, sorry, UI as we've known it, is going to go away because user interfaces are typically very menu heavy. Like, here's where you need to go, but I'm going to go all the way down. And what I've discovered, and I mean, I used to teach Linux 30 years ago. So a long time ago, right? And now it's kind of like, oh, that's all coming back to me. It's all, and I'm loving the CLI. I'm loving. the friction-free nature of being able to build automation.

00:20:46 Ragnar Heil
Because it reminds you like you worked on Linux when you were like 18. I also remember when I had my first customer projects. Somewhere in 2000 or 1999, when I'm running, I think I installed Oracle on a Unix Solaris or whatever. So this, going back to this, giving me some nice flashbacks.

00:21:06 Mark Smith
Yeah. But what it's teaching, and I don't think I've got this as much in the Microsoft, is the actual deep underlying layers of agents and how you should work with them and how you should engage with them  And like, the... For example, the sol.md file, I, with OpenClaw, sorry, not with OpenClaw, with Anthropic, I went, you know, using the latest Opus model, I had a conversation that maybe ran for two hours as I prompted it to go find out best practice. And then I do want a lot of reverse prompting. So I say, I need this file to be absolutely complete. We're not leaving stuff to the future. We're doing it now. And I want to make sure, one, it doesn't get tampered with. What is the implication in the other files? But what I want you to do now is question me a minimum of 20 questions until we nail down what I would like to be in the SOL file. And so it ends up, because see, there's actually like a size limit of how it uses tokens. to what you want to keep your soul, or any of those MD files in, because you're really, you're doing system prompts, right, with all those before it actually gets onto whatever you just said into Telegram. And that's what's really made my system solid  I feel it's, you know, it's robust And as I say, I wake up at the morning, 7:00 AM, it's done a full health check. It says, Hey, I found these three things, and here's three ways I could solve them. Do you want me to do A, B, or C? And I just go, A, like A1 character.
 
00:22:47 Ragnar Heil
And the interesting thing is you learn so much around architecture, which you otherwise you would not learn because, things like Copal Studio, there is not so much around like on security. So you have to go on the Defender side. Then you have to move to the pure few side and talk about the compliance part and the governance part and so on. So yes, Microsoft is super strong in the ecosystem around security and compliance and governance, they are strong, but here you see everything a little bit minimized because vendor and pure few can be also be like a beast. It's super complex. And here you learn just the basics.

00:23:22 Mark Smith
But what I did is I pointed it to Microsoft's responsible standards, and they've got about 5 to 7 responsible levels of AI. And they've got tools like, there's a tool called Pyrite that allows you, I've actually installed that all into my environment. It has that as a core system prompt. So making sure I'm putting the guardrails in, and I think the exposure has been amazing for getting in and really getting all these concepts locked away, because I think in six to 12 months time, we won't even think about them anymore, but like, I feel like I've got the building blocks to, to move forward with.

00:24:04 Ragnar Heil
And this is the cultural change, like one year ago, we needed, when we wanted to integrate an MCP server, we had to go into the JSON file of, I don't know, of a cloud configuration, work deeply into JSON files, but now we don't do it anymore. We just use our Telegram client, add the name or add the, we just paste the MCP server name. And then we maybe we paste our token, then we need to make sure that the token is going to be deleted later. Otherwise, there is also a risk when pasting tokens. But I've also running a nightly ground job, which does also a security assessment. And it tells me, listen, you gave me yesterday a token. Please, it's now stored in my environment file. Should I... delete it. So it's nice. So on the one hand, we see a lot of security issues and threats in OpenClaw, but on the same way, it's telling me how to improve its own security, which sounds like a paradox, but it works.

00:25:07 Mark Smith
Ragnar, it's been so good talking to you. You're the first person I've had a conversation about OpenClaw, and I was so pleased when you said, Ah, I'm working on OpenClaw. I'm like, Finally, somebody else is doing it.

00:25:18 Ragnar Heil
Yeah, exactly. No, let's learn together because this is so new, so exciting. My recommendation is if somebody is not afraid of diving into scripts and all, and so you should start now. Don't give too much access. Work on your empty old laptop, clean up, but start now because it's so exciting because we are experienced this autonomous, really powerful agents. which we knew in the past only from PowerPoint slides.

00:25:46 Mark Smith
Yeah, yeah. So true. Ragnar, thank you so much for coming on the show and sharing your insights.

00:25:51 Ragnar Heil 
Thank you, Mark.

00:25:57 Mark Smith
Hey, thanks for listening. I'm your host, Business Application MVP, Mark Smith, otherwise known as the nz365guy. If you like the show and want to be a supporter, check out buymeacoffee.com/nz365guy. Thanks again, and see you next time.