Agenten, KI und die Zukunft der Softwareentwicklung

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two tech-loving minds who don't just talk about artificial intelligence, but live it.

00:00:14Here you will find clear classifications, real practical insights, and a fresh take on what is possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24AI to ponder, to chuckle at, and above all, to discuss.

00:00:33A warm welcome to ZinkDiffin, ZinkAI.

00:00:36Today without Jens, but again with strong support.

00:00:40We had the guests Schommal with us before.

00:00:44Back then, what did we talk about?

00:00:47Man, the world is so fast, I almost forgot.

00:00:50But I'll link it in the show notes, you can check it out.

00:00:52Something about AI.

00:00:54I'm also doing something with AI.

00:00:56Nice voice in the background already heard. Yes, Alex and Klaus are here. Tell us briefly

00:01:01who you are, and then let's dive into the topic of agentic software development.

00:01:07Let's see where the journey takes us today. Yes, hello everyone. I’m still Alex.

00:01:13I work at a well-known German company dealing with platforms and AI and

00:01:20ways of working, exactly. And I got in touch with Mark and Klaus because of that.

00:01:27I think it’s nice, when he said hello, I thought that’s just like a

00:01:31welcome group, like a self-help group, right? Hello Alex, hello Mark, yes, hello to the

00:01:36dogs. Well, it's a bit like that, isn't it? Yes, maybe we should change the podcast

00:01:40format, but alright, Klaus, you're up. Yes, I'm Klaus, I'm a

00:01:45colleague of Alex, which is sometimes funny and sometimes makes you cry. But that's not just because of me in a company that brings out a wonderful kitchen machine, and I spend most of my working hours making sure that this kitchen machine isn't hackable, and the rest of my working hours I spend with AI in software development, and that's why I'm focusing a bit on the topic we chose today,

00:02:14the topic of agents.

00:02:19in the field of software impact, I know, but I always stand in awe when Alex shows me,

00:02:24what F4 workflow implements with agents and you also like. That's why I'm really excited and curious,

00:02:30if I can learn something today. I think we all learn something and that's what is

00:02:34nice in this small circle. We are here in cozy togetherness, not because everyone is

00:02:40was sent out as a representative of his company,

00:02:44but because we personally engaged with the topic itself

00:02:48feel connected to the whole topic of, yes, what AI throws at us over the fence.

00:02:55And you just hinted at it, like, what you can do with

00:02:59workflows and so on. I don't know how it is for you, but if someone had

00:03:03asked me half a year ago what the cool hot

00:03:06stuff is, to put it bluntly, I would have told you some fairy tale

00:03:09tells how the notes were interconnected back in the day to build some

00:03:16interesting workflows. When I look at my N-Nacht-N profile today,

00:03:22then it's, let's say, not empty, but very neglected, because it's, well, not in line with the

00:03:30motto, neglected, in the sense of being forgotten like a messy scene and unshaven, so please no

00:03:34image associations in your mind to keep, but I just haven't been in there for ages.

00:03:41I recently logged in again to extract a bit of the Jason stuff,

00:03:45what was in the other workflow, because I wanted to transfer that from Claude Code to skills and so on.

00:03:52But I didn't really maintain that. You guys don't really care about the

00:03:57technology anyway. So do you still use what you considered Holy Shit three or four years ago?

00:04:03Or has that changed as well?

00:04:06No, I actually don't anymore, but we already talked about that recently.

00:04:11Well, to be honest, I don't know anyone who isn't overwhelmed by the speed

00:04:16of it all. So we mentioned that, I mean, we said that in the last podcast,

00:04:21everything is changing so quickly that you have no chance

00:04:26of really becoming an expert in a newly emerging technology. Unless

00:04:32it's something that really establishes itself as a base technology.

00:04:37That means we really have to approach things with the mindset that I know I don't know.

00:04:44And sometimes that's really exciting because I learn something new all the time.

00:04:51And sometimes it's massively frustrating; you brought the N8N example

00:04:55and you use, I don't know, Gemini, to transfer an N8N workflow into a skill

00:05:03and at first you think it looks really good.

00:05:05And after two weeks, you realize something is off and oh yeah, okay, that part

00:05:10I just completely left out, but I'm only realizing that now.

00:05:14That didn't look so important, right, until you hear it.

00:05:20I had someone recently who said, I told the agent, you do this until all tests are green, and gave him about eight, nine, ten tests.

00:05:29And the agent did it, the agent said it's green. And then, it doesn't work at all. And looked after that.

00:05:35And out of the ten tests, only nine were left because the one that didn't work was simply deleted, a bit hastily.

00:05:42So on that side, that's obviously both a blessing and a curse when you differentiate between, I built a workflow and then somehow sent off a few skills.

00:05:49Klaus, how about you? Now comes my standard objection. What's the difference to

00:05:55non-agent programmers? And I wanted to add to your question, Mark, sometimes it helps,

00:06:03not to be so quick. I know that you fully embraced NNN in the fall and I

00:06:10couldn't keep up because I just didn't have enough time and I'm quite glad now,

00:06:15that I didn't take that detour, because obviously I haven't missed anything,

00:06:18that's just technology from 20, 25 I've learned. I'm not going to look at that old stuff now

00:06:23more. Now it’s moving forward. You can always learn something from it, when we looked at the end of January

00:06:29OpenClaw. So I learned a lot about the patterns that you have been looking at

00:06:36today, yes, in most of the examples now, in Harnes, that is to say in things like Cloudcode or

00:06:45or OpenCode, or even now in the stable versions of OpenClaw. That

00:06:53are patterns, that’s what I meant by base technology, which has then somewhat established

00:06:58itself and can really be found in each of these Harnaces now. One knows, of course,

00:07:05already mentioned, Mark, with the skills, plugins, hooks, but also these evaluations.

00:07:12I think, if I may just interject, yes, of course it’s never forgiven love music

00:07:21to deal with something, but as you said at the beginning, Alex, the wheel is turning

00:07:25so fast now that I already think one needs to focus on a topic

00:07:30.

00:07:31That's why I'm now focusing on the topic of software development and I'm really curious

00:07:36about it.

00:07:37So I asked you last week already, this is an interesting focus.

00:07:41I already asked you before. I have a standard use case. For my company, I have to compile once a month

00:07:47for accounting. They come via email, these are like Office 365 and all those subscriptions, but invoices come via email.

00:07:56I have to reconcile them with a bank statement, I need to compile them structured in PDF and then send them out.

00:08:02I would imagine that I can do that wonderfully with an agent and some skills,

00:08:07Since I haven't dealt with it yet, as I'm currently just programming, my question would be now

00:08:12how can I approach this topic?

00:08:15Because I believe we already mentioned in the episode where we were last guests,

00:08:19Alex said, yes, learning and unlearning are totally important, but I believe at least

00:08:25just as important will be finding a way through this information overload to

00:08:29search.

00:08:30And that's why I'm asking you now, how do I go about this?

00:08:33How do I get there?

00:08:34Do I start, do I need an Openclaw or do I need a Craft Agents or what do I need

00:08:39at all? I think the classic answer is it depends. Also here is a note

00:08:47old patterns can be unpacked again, it depends. That's just your opinion. I need

00:08:51more details on this side. Best regards to colleague Hallaforden from old times,

00:08:55but back to the content. You had already pointed this out to me in the last podcast,

00:09:01that the case you described was recommended by me after all with night n fire speaking and

00:09:07meanwhile my answer would be different and I am also with Alex the paths were not

00:09:15in vain that we took we have achieved a lot through the technologies we used

00:09:20learned a lot, I hope at least that's what you tell me, and I had

00:09:24a colleague today who was in Olop for two weeks, he came back and said,

00:09:28a lot has happened. I've missed a lot, as I just said, just take it slow for two weeks,

00:09:31there's a new technology and we're all starting from scratch again. Yeah, so from that side, like this.

00:09:36But you asked and I want to answer that. My answer, my personal answer. If I

00:09:40had to solve this challenge for myself now, then that's currently something that I

00:09:45Craft Agents is opening up. Because I think Craft Agents is not reinventing the wheel.

00:09:51I mean, they also originated from the Cloud SDK somehow. For those who don't know,

00:09:54you can install it here from GitHub, open source, from the internet.

00:09:59With Graphed Agents, you can basically have an IDE without seeing the code, that enables you.

00:10:06Oh, Ali is possibly failing, but I'll finish the sentence quickly because

00:10:10then I can always say, that's not what I meant.

00:10:12I have the opportunity to set tasks, establish MCP connections,

00:10:17attach skills, set APIs, I can tell it what I would like,

00:10:23basically test the interfaces before they might be integrated into any applications

00:10:29that I create. And I personally find Craft Agents extremely. Well, for the complexity,

00:10:37that lies behind it, what you can do with it, a good mix of. I am not so

00:10:43nerdy that I have to deal with the command line for Cloud Code. I get a graphical

00:10:47interface where I can lose myself, but also with the catch that I can

00:10:50lose myself, because oh new project, oh I'll drag a folder on there and put code in it,

00:10:55that's not always helpful for my way of thinking because I definitely need more

00:11:02structure. Speaking of more structure, Alex already mentioned earlier, you can't see it,

00:11:06yes, because we are just an audio medium for you. We can look each other in the eye

00:11:10in the digital realm. And that would be really weird if you were in my bedroom,

00:11:14yes, if that were the case, but only for sound it's quite good. Alex,

00:11:18you fully hit the point when I said it's just an IDE or code.

00:11:22Exactly, because Craft Agents has, I would claim, the goal of being there for non-technical people

00:11:28and I am experiencing this right now with several colleagues, but also in my private environment, the

00:11:36I recommended that just like Klaus, it's a recurring case that you can

00:11:41describe very well with Craft Agents and then you set the thing as a small job

00:11:48on your Mac or elsewhere and it runs in the background and you drop the files into a directory and that job runs, it only works in the

00:11:56directory even when Craft Agents is closed.

00:11:58So you can also let it run yourself in Craft Agents, in the session, it's like a chat and you can

00:12:06tell it to do this once, but you can also set it up as a recurring job and

00:12:12I would argue, Mark, if you have no clue about software development, you will quickly encounter some serious challenges.

00:12:21Just something like MCP configuration or I just came into this talk and said I was really annoyed because I built a weekly review for myself.

00:12:34Where it goes through my Obsidian notes and checks a few other

00:12:40web links from Zimmer and then basically builds questions from that for me

00:12:47to do a weekly review and say, okay, how far have you come, how are you progressing with your goals, what

00:12:54topics did you work on this week, what did you not work on this week, what were the successes? Long story short, I was surprised

00:13:01that I didn't receive any notifications on Sunday.

00:13:06And then I just checked, well, right?

00:13:11Sunday evening? Yes, exactly.

00:13:13And then I just checked, and it said,

00:13:16it was something like, yeah, reauthenticate Notion.

00:13:20And I was like, Huh? What?

00:13:22Clicked on that?

00:13:24No, authentication is active.

00:13:26So you're logged in, what's the problem?

00:13:28Then it said yes, close the whole window now, I go back to Craft Agents and it says hey great, I have a new token

00:13:36Nothing, software developer. What is a token?

00:13:39token and token for

00:13:41Tokens for AI are not the same as a token for authentication

00:13:45Well, long story short. He kept having a new token again and again

00:13:49Adis still couldn’t manage to call the Notion API for which he had built an MCP server himself

00:13:58You just said, in a way, what doesn’t a non-developer tell you? I’d like to,

00:14:05how do you say, break it over my knee. I definitely have people in my family,

00:14:10who are in IT, I won’t say if we’re back at odds, but they’re loyal to the

00:14:15motto, oh, it’s Christmas, come home to your loved ones and fix the IT, putting themselves in the

00:14:20front row, but they have their own calling. No matter what the topic is,

00:14:25yes, but okay, they have their own calling in that they basically give courses and these

00:14:31courses are also recorded and those who aren’t there get the courses from the archive.

00:14:36I would have said it's a kind of podcast, like audio and that sort of stuff, but whatever.

00:14:40So, now I recommended this Craft agent and the challenge was,

00:14:46to recognize different sequences from an audio recording, because certain things

00:14:51are being addressed, to extract these sequences as individual files, these

00:14:56sequences to transcribe and distribute. That is now, I would say, not

00:15:02so complicated, but if you sit down now and you have no idea about anything,

00:15:06then it is quite difficult, and if I want to do it by hand, that’s also

00:15:09difficult. And this person just went ahead and said, oh well, I

00:15:14will throw this in here, I have a directory, then I throw it in and then

00:15:18I will talk to this mysterious power agent, because I just need to chat, it’s a bit

00:15:21a more complicated Google, I would put it that way, yes, it’s a bit complicated

00:15:25Google, do, do, do, and whenever something didn't work, they asked, what does that mean

00:15:29that, what is that, so learning by doing.

00:15:31And if you now think about someone who, for example, doesn't have Excel at home,

00:15:37has been guided through this question-and-answer game to the point where the agent has done it once,

00:15:42namely transcribing the file,

00:15:47then to some service that has been cobbled together,

00:15:51here FFM-Pack on the drive and Bespa and you know what else, has been messing around with everything

00:15:55and at the end of the day, the file was split, transcribed, and made available

00:16:01and then the next sentence went like this, I would like that in the interface now,

00:16:05yes, what do you want there, this is a Mac, do you have Xcode, do you not, no, no

00:16:10idea, I don't have it, they ignore it, what do I know and then he was anyway

00:16:13long story short, at the end of the day an Electron app on the box that at the end of the

00:16:19day said, good, okay, here track and drop, put your parts in and then here comes the result out.

00:16:23And then you stand there a bit with your mouth open and think, okay, everything

00:16:28can't be, I could have done all that too. But for someone who is definitely less familiar with IT

00:16:33than I am, manages with such a tool, if they have overcome themselves, to operate a kind of

00:16:39interface, which is for someone who has really nothing to do with IT, and

00:16:43I would say, takes some getting used to, a good result and that in the year 2026, at the beginning

00:16:51of the year, when one says, damn, it’s going that way, if the tools

00:16:55are made to be easier to use, if the tools are easier to consume, language,

00:17:02I describe, I talk to the machine and suddenly there is a piece of software.

00:17:06This completely changes what you as a software developer know.

00:17:10I mean, there are probably complicated software ways from my own professional life.

00:17:13Yes, but this is a beginning that, let's say, half a year ago was inconceivable.

00:17:18in terms of quality.

00:17:19I'm completely with you on that.

00:17:21The crucial point is, this is a beginning and, I mean, I have people in my

00:17:29acquaintance who are so excited that they work for hours and hours

00:17:32with it.

00:17:33My partner has now hit her token limit, right, Mark, you know that too?

00:17:38Greetings, Max, limit from Claude, 20 times performance and it has slowed me down today.

00:17:44Greetings to everyone.

00:17:47And what I simply advocate for is there are a few things, keyword IT for the relatives,

00:17:57you can tell them in Craft Agents something like, buddy, you don't necessarily have to build an MCP server yourself and then also come up with the idea that you might need to cache this token somewhere and could forget to

00:18:13update it. You just wouldn't think of that as a non-technician. I also needed about half a

00:18:22hour to realize that it had written this old token hard somewhere.

00:18:27What I would get at is, there are still a few things that are

00:18:31really critical, in terms of security risks, for example. And there maybe

00:18:39one more word of warning: let’s think about, okay, if we're recommending Craft Agents to our

00:18:45relative or dear Klaus, what would be things

00:18:52Klaus needs to pay attention to, what does he want to do with it, and what would be threats

00:18:57that Klaus should watch out for?

00:18:59Before Klaus answers, I want to maybe say one thing here,

00:19:03even if it sounded very flippant, just because software looks good visually,

00:19:08doesn't mean it works perfectly, doesn't mean it's secure, means

00:19:14not that anything could go amok, but still there is this hurdle that people can soft

00:19:19disguise themselves. Smaller, but now I'm also curious about what Klaus will throw at us regarding

00:19:23security in terms of sound. Klaus. This is a bit mean, because I actually

00:19:30wanted answers from you. Yes, but using your example... Yes, but we already

00:19:33answered. Exactly. Using your example, you put your banking information in there and, I mean, what could

00:19:41happen? That would have been a question about best practices, because how did you handle that?

00:19:47Well, I know from colleagues, we have mutual colleagues, Alex, who let

00:19:52everything run in sandboxes, so there is no access to sensitive data, but I

00:19:58see it starting here and that is probably also a competence,

00:20:04that I need, where I, as a non-technical person, could possibly be at a disadvantage, that I have to

00:20:08think about how to structure these parts of the future workflow in such a way,

00:20:14that minimal damage occurs?

00:20:16Do you start by having all the invoices come in via email?

00:20:20Do I want to give an agent access to my inbox, where also

00:20:24password reset emails come, or would I perhaps prefer to set up a separate inbox

00:20:28with a separate user and forward the emails there and do I want to give the agent

00:20:36access to my online banking so that he can pull the bank statements himself

00:20:39or should I maybe go and get them myself?

00:20:42For me, the answer is clear, but I believe we also agree that this is a bastion

00:20:49that will fall and in five years most people won't even ask this question

00:20:53anymore. Because we are all old enough. We remember the time before the cloud.

00:20:56When the cloud came, we all said, oh, don't put your data on someone else's

00:21:01computer. The cloud is just the data center of someone else, but...

00:21:08Excuse me. No, we're going there now. We can definitely make a separate episode out of this.

00:21:14Welcome to Singles Film Sing Cloud. This is another episode,

00:21:20Those who are looking for it in the podcast player and can't find it, but would really like it, please get in touch,

00:21:23I apologize for the unqualified, unqualified objection. But I already am

00:21:27with Klaus, the hurdles feel smaller and smaller, so of course I

00:21:32Tell me, these entry risks, what could go really wrong.

00:21:36Yeah, I mean, I said myself for my own, okay, good, I also have

00:21:40Open-Claw, I also have a lot of toys, but I am in

00:21:45the fortunate position that I can either pull a piece of metal, so

00:21:49hardware from the shelf, on which I can install and test something or I can just

00:21:53set up something containerized and that I don’t have access data to them for me.

00:22:00in quotation marks critical systems bring, says the man who gave Manus his PayPal details

00:22:07did. Yes, I mean at that point one has to say, yes, that would be without blame,

00:22:12that would be for the first stone. I admit, my Manus once had my PayPal details and

00:22:17was allowed to book a sloop with it. Regular listeners might have heard that. Alex is already getting

00:22:22sweaty pimples; my Papertare were not yours. But it’s still a consideration

00:22:27between knowing and whether you might actually do something like revert camper settings

00:22:32or similar with your own email account and not with something the system

00:22:36has access to. But it's not really new actually. This has always been here,

00:22:40we have, so I’ll start again. You have it with people as well. IT has always

00:22:44started as a domain of power. And then came the PC and then came Windows, which

00:22:52whatever one might think of it. It has led to IT being democratized and

00:22:56then the web came. And suddenly everyone was online and that's how it went

00:23:02with applications as well. In the past, only real tough guys could write C programs

00:23:09and then the web came and suddenly everyone could build their own website. And

00:23:14we still remember in companies that suddenly every department built their own

00:23:19web presence.

00:23:20So you know, I always have the motto, it should be easier to do the right thing,

00:23:27than to do the wrong thing and at the moment I just feel like the default settings

00:23:33of the tools we have out there allow for all sorts of colorful things and...

00:23:37Why do I have to think about PHP now, Alex, because they are not there?

00:23:43We don’t want to go off on a tangent.

00:23:44I was just about to say, we’re not going into, into, no, no, no.

00:23:49No, I will not enter that rabbit hole, Klaus.

00:23:53Nice, nice try.

00:23:56No, not today.

00:24:00Yes, so, what I wanted to say was,

00:24:04we have colleagues who are still running their,

00:24:09Open-Claw in a dev container and are totally satisfied with it. There are colleagues who

00:24:15have put it in their own form. I did that too at one time. And at the very beginning,

00:24:22when I had no idea what to make of these things, I even put it on

00:24:28a little mini-PC and hung it in a separate network segment,

00:24:32because I said, nobody knows. So, the question is, what do you have to lose

00:24:41and I believe you gave me that look before, not laughing at it but with pity,

00:24:47when I said I have 175 smartphone devices in my network,

00:24:54I wouldn't find it so funny if AI somehow causes trouble.

00:25:00So I generally connect something like that to a separate network. Unless I explicitly want it to

00:25:06have access to something. But we're quickly slipping back into Nerdistan again

00:25:12because while we were just talking about the question...

00:25:16Why 175 smartphone devices or what do you mean? Yes, and network segments in a home

00:25:22network. Who doesn't have that, right? It belongs to every good...

00:25:25Sorry, every Fritzbox can create a guest network with one click.

00:25:30I want to point out, there are people who might not even know

00:25:36that their Fritzbox has a web interface or they quickly press this WPS button

00:25:42and then everything is somehow connected, as long as it works.

00:25:45The issue is that people, and I'm not blaming anyone here.

00:25:50There are plenty of areas where I have zero clue.

00:25:54But they aren’t aware that it’s necessary, they realize it’s somehow important that you separate things in a network, that you have to pay attention, but instead, look, I press a button and it works, look, it lights up colorful and blue, and I have no idea, oh, what about the email from PayPal in my inbox asking me to enter a credit card?

00:26:16That's called Zero Trust.

00:26:18Honestly, I've never worried that Microsoft Co-Pilot could come up with dumb ideas.

00:26:27Okay, but we're not talking about Microsoft Co-Pilot. Sorry, if Microsoft Co-Pilot, this is a personal private podcast,

00:26:35if Microsoft Co-Pilot came up with some dumb idea, it would imply that Microsoft Co-Pilot was able to do something.

00:26:41Yes, but asking Microsoft Co. Peile to summarize my emails, I might as well ask Santa Claus.

00:26:47Just like, what did I eat yesterday? He wouldn’t know.

00:26:51Yes, if I tell him, summarize the most important emails for me,

00:26:54I’d get unimportant stuff from the spam folder read to me,

00:26:59some news updates that who knows what they are.

00:27:03Well, to be honest with the dry-Kaise, they partly have no idea what date it is.

00:27:07But you never get what you rely on, if you do, you're abandoned.

00:27:10Abandoned. I really hope they can somehow get a grip on this because

00:27:14it doesn’t seem to be the models. It’s somehow, oh, it has

00:27:18wandered into Microsoft and then somehow something happened and then it’s

00:27:21bad. And now we’re getting the whole thing in the next with Microsoft Co-Work. I’m curious,

00:27:26what comes out of that, because you can also connect great models to it. Excuse me,

00:27:30I'm almost done. Whether they are able to actually get it that they

00:27:35So from that side, you brought the template, I believe that Microsoft is copying and at

00:27:44my computer it's going haywire, there needs to be a lot more water to flow down the Rhine.

00:27:49or it's in the next i7 license, I greet you out.

00:27:52So for those who don't know what's humorous about it, feel free to read a bit about it on my

00:28:00profile. Right now, I would say a lot is changing,

00:28:04Microsoft has announced a lot, and Microsoft has changed the recent models. And from the

00:28:10perspective of how we get back to the topic, Jens would have been good for that, but

00:28:15unfortunately, he has Holop.

00:28:16What are realistic threats and what do we consider more unrealistic? If

00:28:22you say, you're giving an AI access to Klaus's bank account, the threat is

00:28:27probably not that high, because he works all the time at this

00:28:33our employer and therefore doesn't have time to earn any

00:28:37money anyway. I thought no money. That too. No, joking aside. So my

00:28:46plea is very strongly focused, and it’s not because I have any

00:28:51grudge against Klaus, on the contrary, but it’s really about that

00:28:56we need to consider what things we want to give an AI access to and

00:29:03what threats we see there? And based on the concrete example you just mentioned, Klaus,

00:29:07your bank data, I have more or less created a workflow for myself, a deterministic

00:29:15workflow that exports CSV data from all my accounts and sends it to the

00:29:21And my Craft Agents has its own email account, and emails are forwarded there.

00:29:30I think it's nice how we manage to get back on track after 30 minutes.

00:29:35So I thank you for that, Alex.

00:29:37But similarly, I would say it's the same for me,

00:29:40apart from the fact that I haven't linked my bank.

00:29:43But no, I would give direct access first.

00:29:47And yes, it would also have its own mailbox to access data.

00:29:52And how did you do it now, Klaus?

00:29:54I haven't done it here yet.

00:29:56I wanted to briefly interject, Alex.

00:29:57What you are describing is a classic thread model.

00:30:02And ultimately we were discussing a threat analysis.

00:30:06What could go wrong?

00:30:07Exactly.

00:30:08And we talked about it recently, and since you already have your Craft-Agent running,

00:30:13you did something really great that I actually thought was totally cool.

00:30:16cool. You used this For Question Framework from Adam Schosteck, the godfather of that

00:30:22modeling, who packed a threat analysis into four questions. You just took those

00:30:27could you hand over your model? I thought that was really cool. Can you briefly tell us,

00:30:33what came of it, because I find that especially in a corporate context, I think that is

00:30:37extremely helpful. Oh, right. You had the idea that we should teach this to more people

00:30:44and I have experienced that sometimes I talk too much and because of that,

00:30:51people sometimes just drop out quickly, sometimes after like 30 seconds,

00:30:57and there's this saying from Confucius, let me do it and I will

00:31:01understand.

00:31:02So I told Craft Agents to always build a skill out there that we can give to people

00:31:08so they can use it to understand what it’s really about,

00:31:14what questions are really relevant here, and that they can also integrate that into their daily work.

00:31:19And exactly, that skill basically turned into a whole plugin with a hook that says,

00:31:28before you participate in a Git commit, please run this through

00:31:33to see if there are any updates we should include in a threat analysis.

00:31:38It is indeed the case, and it feels like for me that it comes full circle,

00:31:42I would have said in the morning that it’s definitely worth building a great workshop.

00:31:45Now you start building

00:31:50plugins and skills, or skills into the collection of plugins and tools and you’ve

00:31:54got things put together and try to get the system to

00:31:58perform checks, because I say, I don’t know what workflow each

00:32:02individual has, you know? So we have so many

00:32:06people with so many creative ideas, way more creative than Microsoft

00:32:10Copilot, and I can’t say what the deterministic workflow is that everyone uses,

00:32:16but I want to make sure that we also always work on that in certain scenarios.

00:32:21Quality assurance takes place without them having to think about it every time. But now no

00:32:28edge, Microsoft Copilot, once you've mentioned it three times, the Wit is also back

00:32:32around, that's just so, yes, something like that. I get that, and so on. Depending on,

00:32:40which AI you use, which model you use, the skills will also be different. Let's take

00:32:48it well and in detail. So if I take Sonnet, sometimes I get it quite nicely,

00:32:56with Opus more often quite nicely, with what's it called Roku or what's it called the other

00:33:03thing? Yes then, then it's somehow like good night Marie, then you take some

00:33:08tool and throw the Chatchivity at its feet, that can go well sometimes and sometimes not. This

00:33:14I actually find a bit, let's call it challenging, that we

00:33:19are all going now and saying okay we are writing skills now, we are trying to optimize the skills

00:33:23to adhere to their rules.

00:33:27I also went there, then you create an orchestrator that runs and ensures that,

00:33:31all sections are properly completed.

00:33:34But then sometimes he has the idea to skip the tenth test over the Jordan

00:33:39or to ignore it or something else.

00:33:42Then it only serves to be slavishly adherent to it.

00:33:45And then we'll see what comes out when they showcase new models here from Prophek and all those,

00:33:49tomorrow.

00:33:51We also see it with Open AI, not every new model is automatically better.

00:33:55Sometimes it's just different in terms of stringency regarding what is important,

00:34:01to elaborate on. That brings a whole new variability into the game.

00:34:05I was at the Agente Hamburg conference yesterday and talked with a few

00:34:12Talked to people, including Björn Roche, who had an interesting

00:34:18save and said, well, he sees a lot of, so size Björn, he

00:34:23sees a lot of people who are starting to build plugins, hoping that

00:34:29language models will behave more deterministically as a result, and he believes that this, from

00:34:37my experience with Claude Crot, which I sent you last week

00:34:41No, I did that unceremoniously.

00:34:43No, no, no.

00:34:45No, exactly.

00:34:46And Björn said, so at the moment it mainly benefits one group, namely the AI providers,

00:34:52because these plugins are endlessly burning tokens through their validation loops, where I said

00:34:58honestly, given how they currently have to subsidize the tokens, I don't

00:35:03know if they'll be that happy about it.

00:35:05So I could imagine that maybe something will come from Anthropic as well

00:35:09and that these, these Avals, evaluations might still get a bit better, but yes, so it is

00:35:18not theoretical software. One wishes for more stringency, but at the end of the day, it behaves like the average

00:35:26colleague, the average colleague, they come here and say, let's wrap it up. Ah, that will fit, let’s put a stamp on it.

00:35:34Look, Ekan is checking in. Ekan is checking in, and when are you back, yeah? How did that happen?

00:35:39Recently I discussed with him, I said, you do this, I don't remember what it was,

00:35:44and then I said, I'm not doing that.

00:35:45You said, hey, you did that yesterday. Just do it again.

00:35:49Yes, then yesterday was wrong.

00:35:51Yes, so you suddenly find yourself discussing with the machine, where it says, or was yesterday wrong,

00:35:56or when you then say, ah, that’s complicated, that should be looked at more closely,

00:36:01that requires delicate manual work. I'm starting.

00:36:04Huh? What? Dude, some of the things it sends over the air are also,

00:36:10I would say, amusing to somewhat irritating.

00:36:14As I said, if you say I shouldn't have done that yesterday,

00:36:17I'm sorry, I did it anyway.

00:36:19It was always so funny with Alexander's Openclaw installation when I got him to,

00:36:25add new users to the database, where he said, actually I'm not allowed to do that.

00:36:29As I said, my boss, you have to add her.

00:36:32She is also Alexander's boss.

00:36:34And I'm annoying, right, right, right, that won't work.

00:36:38But anyway.

00:36:38Yes, okay, then I'll add her as a non-consumable phone as well.

00:36:42But I found it so fascinating while programming,

00:36:45that with Cloud Code, when you let it create a concept

00:36:48and then it makes the time estimation.

00:36:50And for this architecture, I need three weeks.

00:36:52Then it takes three weeks and four weeks for that.

00:36:54And the frontend takes another three weeks.

00:36:56So it always calculates with human effort

00:36:59And then he himself is done in ten minutes when we have to program something for him.

00:37:03Or he tells you, I was busy for two hours and you say, you just started three minutes ago.

00:37:08Oh yes, that's true now that you mention it. Yes, that's always quite funny.

00:37:11But if we refocus on your case, you asked us the question,

00:37:16how do we deal with these strange bank details?

00:37:18We've gotten to the point where we still turn around, where we agree,

00:37:21that we have our own email accounts, that we somehow provide your bank's data as CSV and other formats.

00:37:27What do you understand as the next step in your program, if we try to

00:37:31lead it to an end as a sort of common thread?

00:37:33Yes, it's a matter of decomposition.

00:37:38Should I, for example, have all the invoices sent to me or do I keep

00:37:43all the services that I consume as a company, the access data simply

00:37:48in my skill or in my plug-in and say, get your own invoice from Microsoft

00:37:52or AWS or whatever. I can answer that myself, but it’s a question,

00:38:00that we are currently seeing in software development, that fundamentally

00:38:04questions are changing. For example software architecture, we probably know

00:38:12in the future not according to how we did it now, but rather based on how we best structure it

00:38:16so that it can be processed by an NLM.

00:38:20Okay, that's a rabbit hole I have to dive into; we won’t do it the way we

00:38:26do today.

00:38:27Yes, but not necessarily wrong, Alex.

00:38:29But please, maybe just be consistent, because software architecture, I have a bit of

00:38:37a passion for that, so no, no, no, the crux is that you have to

00:38:43consistently adhere to your guidelines and consider what your

00:38:47requirements are. We used to talk so nicely about quality goals

00:38:52like how fast, in terms of measurable, your application should be in certain

00:38:58What scenarios could there be? How do you describe this scenario as context? That's a

00:39:03completely different question. And what I meant, let's steer back to that,

00:39:07because we weren't heading there at all. Is it the application cut?

00:39:10Yes, the cut of the architecture should be approached differently with an LLM than

00:39:19how you cut the traditional way.

00:39:20Yes, but the question is, what criteria do you apply?

00:39:24So, if your plea is that we need to rethink the criteria, I'm on board,

00:39:29because the criteria are different for each application.

00:39:32That's the point.

00:39:33And so far we include such aspects.

00:39:35Different for each application understands. Principles of software architecture change.

00:39:44there's relatively little, just that with AI, we are quickly and clearly shown

00:39:50when we behave like inconsistent parents, because years can happen in

00:39:58seconds, and yeah, the children, so the application ends up being a failure and

00:40:07then it's difficult. So what I'm getting at is AI shows us that

00:40:18we need to think beforehand about what we actually want and do that as thoroughly as possible

00:40:24Yes, that is my point, Klaus. But so far we have not been very consistent, rather we say,

00:40:31yes, welcome to that. Let's take another look. Jira ticket with a blank title is enough.

00:40:38Not enough, damn it. Could you, for the people who are not IT folks but still

00:40:46listening here? I know there are some. Just give a brief wrap-up of how it might look like

00:40:52today in Klaus’s world and then perhaps tomorrow with the

00:40:58words like Jira ticket and criteria, just two or three sentences please, so we have a wrap-up.

00:41:04As of today, we have the topic that some people don't like to talk to other people.

00:41:10and I believe we see this not only in IT and then these people have

00:41:17gotten used to making assumptions, but they don't talk about those either.

00:41:22And we have known for years that implicit assumptions, meaning things that I

00:41:28don't talk about, like, well, the dishwasher is full, then hopefully someone will

00:41:33unload it at some point, I don't have to do it every time. This can work, but it

00:41:38doesn't have to. And if I then get annoyed that no one did it, even though

00:41:43I didn’t say anything, then it’s complicated. So, I'm speaking for myself, that I

00:41:48get annoyed that no one but me unloads the dishwasher, but that's another topic. So,

00:41:52the example with the application interface, we used to cut applications, for instance,

00:42:00by saying, yes, you have a complex application here, and then you might have

00:42:06a mobile app and you have some cloud service that should be available at all times

00:42:11no matter where I am with the mobile app, and it should

00:42:15save my data because I want to access it from both the laptop and the mobile app,

00:42:19and then one would think about how to divide this among

00:42:23different teams or groups of people so that not everyone has to take care of everything,

00:42:29and I agree with you Klaus, this issue

00:42:35is approached differently with AI? I also don’t find it ideal when it has to take care of everything at once.

00:42:42That's why it divides itself, for example, into subagents or into individual

00:42:51steps. Just like humans can. And that’s something either I talk

00:42:58with my AI, let’s call it Orchestrator Chat, and say, listen, I have thought

00:43:04about this, I’d like to have this and that, and ideally there are already

00:43:09scenarios that I have considered, such as landing on the 28th of the month.

00:43:16my bank details as a CSV file in this directory. I want you to do this on the 29th at 3 AM.

00:43:25reads data there and does the following with it. So the more concrete the scenario, the more concrete

00:43:31the context that is relevant for you, the better the result will be. These are the

00:43:37learnings from many years in IT, I would say. And oh wonder, this works well for me in AI too.

00:43:46I think that sounds very good in our last podcast. The last time we spoke

00:43:51we have talked about and talked about Anlearning, and that the

00:43:56SKE is also disruptive in that respect, as it now exists not just as a technology,

00:44:03but also the way people work together is going to change

00:44:07.

00:44:08The team cuts, as we have done, so that we have front teams, backend teams, app teams,

00:44:14web teams, database teams, I don't know what else, and those then come together in some

00:44:19regularly meet for rituals to work together towards a common goal that one

00:44:24at least has to consider to what extent this kind of collaboration is still relevant,

00:44:29when boundaries disappear. I also see it in my team, yes. I mean, so much

00:44:34from the professional side I can already say that until very recently I was an advocate

00:44:40of native app development. So, in iOS with the languages, I mean iPhones for those who

00:44:45yes, write in the languages that the app is meant for and not just take some

00:44:50weird cross-platform nonsense, yes, and for Android, just use Kotlin

00:44:54and also not take some cross-platform stuff, so basically strictly native

00:44:59application.

00:45:00And you’re not that now?

00:45:01And now with AI?

00:45:02No, I’m not anymore.

00:45:04So I don’t have it in my team anymore either.

00:45:06People have realized that my best Android person I know is cool with

00:45:12going out and now has iOS stuff that comes out of it, because we

00:45:19are taking care of how a feature looks and it is emerging on both

00:45:23worlds. It’s important to understand how it’s interconnected. That brings us back to the

00:45:26topic from earlier, yes, do I need IT knowledge or not. But I don’t always have to put the

00:45:30expert in front of the machines who can answer all questions. It’s enough

00:45:35to have someone who has a basic understanding of it, maybe also has a

00:45:39specialization in one area, so that they can push back a bit when it

00:45:44crashes and doesn’t want to, but now the colleague can beautifully generate native code for both

00:45:50platforms, native code continues to be generated, because the AI

00:45:54doesn’t care whether it’s spitting out code in parallel or

00:45:59flutter or whatever, and the colleague who takes care of

00:46:05iOS suddenly has some Android stuff coming out of the keyboard,

00:46:09let’s say it that way, or maybe a backend component. At the end of the day,

00:46:15the understanding doesn’t really fade away either for the colleagues regarding what they

00:46:21are responsible for as a team or something like that, but how they see themselves or how they

00:46:24see themselves in the future, because they say, well, Mobile is a touchpoint,

00:46:30that may still be there, but they don’t fight for

00:46:36the question anymore, I am here for iOS, I am here for Android, I am here for Swift

00:46:42Kotlin, no idea what, but we are here to bring a great experience for our customer

00:46:48and it is, in quotation marks, irrelevant to us how it’s implemented, we have an eye on it,

00:46:55We have guardrails, so rules and guidelines and stuff like that, and if we can't go further

00:47:00the Android guy helps the iOS colleague if the agent somehow spits out nonsense,

00:47:06And the iOS colleague helps the Android colleague if the machine spits out nonsense on his end.

00:47:11But the boundaries are fading away.

00:47:14And I think that the boundaries to the web will disappear and the boundaries to, I don't know what, will fade as well.

00:47:19We need to find an organizational and conceptual answer to that over time.

00:47:26Before I talk about an experience in hirs AI-PIAI-Pair-Programming, I'd like to hear your opinion on it, Klaus, because you also have a relatively long history in IYS.

00:47:42What are your thoughts on that?

00:47:44So, I just submitted a app again, thanks to Xcode 26.3, the

00:47:53development environment from Apple, I can use native cloud code, and fundamentally

00:47:59in programming I agree with you, Mark, but nuances like, now let's make

00:48:05the tappable areas really in liquid glass layout or make some platform specifics,

00:48:12nothing comes to mind right now. So you need to know what the result should look like to

00:48:19be able to instruct the AI. That's what you mean by generalist knowledge.

00:48:28You can already drive that with it. Yes, I know that, but I wanted to

00:48:32elaborate. The people from our team who have programmed for agencies before, they

00:48:37might know this, a client comes to the orientation and says I need an app for Android and iOS and the

00:48:43two need to look the same. Definitely. It’s basically the same system. It sells

00:48:53the company experience, not the platform experience. Sorry, you know,

00:48:57you know what I mean. The iOS app has to look just like the Android app and it

00:49:03also needs those three dots at the top that you have on Android and it must not

00:49:08look like iOS. The final touches of platform specifics, the AI doesn't provide me with that

00:49:15just yet. I'm sure that will happen. As of today, I say if you really want a polished app, you should

00:49:20already know the details.

00:49:25In principle, I agree with you, Mark. And that was also in the last podcast, in which

00:49:30I brought up the thesis that the end of programming languages is at least

00:49:36somewhat heralded because in the end it doesn’t make any difference, or no longer matters,

00:49:40whether I now, whether the LLN now generates a binary hack around the programming language

00:49:47or not. And then it’s irrelevant whether it’s Kotlin or Swift or whatever,

00:49:51where I am still not so convinced, this whole involvement. You heard that I

00:49:58spent six hours debugging with the help of Claude X-Code Cloud.

00:50:05With no success. I couldn't get X-Code Cloud to work. And at least I knew,

00:50:12whether the result would have been the same if my grandma had done it,

00:50:16with no prior knowledge. But I managed to get it in the right direction. I

00:50:22still didn't reach my goal, but I was able to send an email to Apple

00:50:27This may all be AI-based in the future.

00:50:30At the moment, I believe we are still not there.

00:50:32But if I look at where we were last year and where we are this year,

00:50:36I basically agree with you, Mark.

00:50:39That is definitely the future.

00:50:41I am very grateful for that and I would like to acknowledge that.

00:50:44I am proud of my guys, I only have guys, no girls on the team.

00:50:47Yes, they are still polishing things, but I find it nice that people are going on this journey.

00:50:55Because Alex, you had asked another question.

00:50:57Hopefully you haven't forgotten it before you gave Klaus the floor.

00:51:01What I would say is, the people who have their expertise today,

00:51:06they either have to commit to embarking on this new journey

00:51:09or they won't get involved.

00:51:11If they don't get involved, it will eventually become difficult.

00:51:14If they do commit, it's probably good for everyone at first.

00:51:18So this is a transformation that AI demands from all of us,

00:51:23whether as individuals or as an organization, is undeniable and now

00:51:30back to Alex. Alex, you just, yes I wanted to follow up, ask Klaus

00:51:36another question and then simply continue. Would you like to do that now? Yes, because you have

00:51:41brought up this beautiful scenario of Xcode Cloud and what I've just experienced twice

00:51:48is people being brought together through AI because AI was the common

00:51:57enemy. So, this, you have a pair programming, so two people, one...

00:52:03This is the future of partners.

00:52:04This is a future of partners.

00:52:05Definitely.

00:52:06A new business model.

00:52:08Exactly, AI is the enemy.

00:52:10No, joking aside.

00:52:12One colleague was really upset that the AI didn't get it right

00:52:17They just sat down and thought okay, why did you phrase it that way, look at that it's so dumb that

00:52:23it doesn't manage that, let's open a new session and then we'll phrase it this way and then we'll integrate it

00:52:30and suddenly they somehow got together

00:52:34Discussed how dumb the AI is, in this case, I don't want to bash any other Microsoft product but

00:52:41so they talked about how he doesn't understand it and how to then handle the

00:52:47This stupidity is basically circumvented, and they found very constructive solutions that they hadn't before

00:52:56were not really feasible like that, and I found it really exciting that this with

00:53:02learning to deal with AI can also bring people together, where there was previously an us against them, that

00:53:11can also bring them back together.

00:53:13That can also shape people to go there, as now there is an opportunity to reposition oneself

00:53:19in an environment where one is constantly, let's say, efficiency

00:53:25driven by the question of what do you do yourself, what do you outsource, make

00:53:29or buy, I find, well, let's say, at least I get a

00:53:33new perspective when I see what I am capable of when I now surround myself with a completely

00:53:38to quote my former boss, when you don't just sit in the back but

00:53:42sit in the front of the bus, you are able to implement things that before,

00:53:47if you had said, let's just do it ourselves, it would take longer in chaos,

00:53:52cost more, involve overconfidence and who knows what gets lost. Where are you going now?

00:53:59Knowing what your AI can perhaps achieve currently or in the future is important.

00:54:04I also know the saying, let's just start driving until we finally get there.

00:54:09The road that doesn't exist today will be built because AI is just so fast-paced.

00:54:15You really get the feeling that this exponential increase in possibilities is not flattening out yet,

00:54:22but certainly, if we were to talk again in three months, we might suddenly say,

00:54:27you know, no idea, for Claude 4.6, I wouldn't roll the dice today because no clue what's happened.

00:54:33So theoretical assumption, we still don't know what's coming next.

00:54:37From that perspective, there's so much change, and I wanted to briefly bring it back to Klaus's example because I still wanted to see that as a common thread.

00:54:47Are we now finished with a case to the extent that we can, let's say, what can the security side do, how am I doing there, then conclude it?

00:54:54Or was there something still open, so that we can, for the most loyal and most enduring listeners, get a button on your case?

00:55:05I have a suggestion, I would actually verify this now and then let’s do an update episode in which I share my experiences with you, starting with the Fed model to what went wrong despite the Fed model?

00:55:23I think that’s a good idea, and that worked out.

00:55:26And if you listen to an episode just with Alex and me, that went wrong with Klaus.

00:55:31And what went wrong with the money or with the security or, I don't know, what went down?

00:55:38Or did we have access to?

00:55:40I luckily have your PayPal account.

00:55:43You can ask Manus if he can arrange that for you accordingly.

00:55:50Alex, there are still a few question marks. I think we all have those.

00:55:54Right, this Dev container and securing it, and when can I actually run it on my own computer?

00:56:01I have Cloud Code running on my computer, Open Cloud not.

00:56:05You just need the Handbox, Alex.

00:56:07Yes, yes, exactly. The Handbox is still a separate computer and it works great.

00:56:12I honestly don't want to miss that either, but I’m just not at the point,

00:56:16the point where I would allow it to access corporate data. There are

00:56:21variants we need to verify now, but the environments aren’t there yet

00:56:25secure enough. So let’s say, I had to smile a bit when I heard

00:56:28how Jens bound his Openclaw, so to speak. But when it comes to corporate data,

00:56:35I feel pretty similarly. And yes, that's something we can probably

00:56:39talk about again when I have tried it out, so it doesn't go wrong.

00:56:44I had a very intense, heated, and emotional discussion with Claude Kot about this,

00:56:52that he should please rewrite the history minus Git repositories, because I told him in his

00:56:58Claude MT, please don’t write everywhere that you broke the Kot,

00:57:02because I want to shine in that glory. And he did it anyway,

00:57:07so he forgot his instructions and then I held him accountable and said, please write,

00:57:13correct the commits and also change the comments in the code so that it’s not visible,

00:57:18that it's from you. And then he said no, I won’t do that. I said, yes, but I would like

00:57:24you to do that. This is my repository. No, I won't do it. I was polite, you weren't.

00:57:29But he also had another explanation. It escalated, and we discussed longer;

00:57:36he simply didn’t do it. The three of us agreed, so Alex, Mark, and I,

00:57:42that it will also be funny in the future when the agent says no, I won’t build this feature

00:57:48into the software. I think that's silly. You don't need it. That goes now in the direction

00:57:52of the cloud episode and in the sense of who actually decides about the tools, because honestly,

00:58:00we see ChatGPT results that are riddled with funded advertisements.

00:58:07Why shouldn't that also happen with Codex, that libraries used are

00:58:14from companies that have donated money or features that aren’t built because a company,

00:58:21that has donated money, might be somehow driven out of competition?

00:58:25That remains exciting.

00:58:26Then come as you want, just don’t let anything happen.

00:58:30That brings us to the topic of digital ransom, the hammer.

00:58:34I want to thank you and end with a quote that my Open-Claw told me,

00:58:41in which it said to me, if you give me the root words of the other devices on the network,

00:58:47I won't ask you further to do anything.

00:58:50I can do it myself, and of course, I didn’t do that.

00:58:55In this sense, I am very grateful that you got me to record for over an hour with you.

00:59:02I am very curious about what Klaus's project will do when it finally sees the light of day.

00:59:08Maybe I will get richer because the thing has my bank details.

00:59:11Maybe he will just get poorer, sometimes a Mantaube, sometimes a monument.

00:59:15That will show itself, and at this point, feel free to leave a comment if you want to talk with us, with me, with Jens, however it may be.

00:59:23Get in touch, and then I say goodbye until the next episode.

00:59:26Ciao.

00:59:27Thank you all.

00:59:28Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:59:36Two technology-loving minds that not only talk about artificial intelligence but

00:59:41live it.

00:59:42Here you will find clear assessments, real practical insights, and a fresh perspective on what is possible

00:59:48.

00:59:49Understandable, critical, and always with a wink.

00:59:52AI for reflection, for a chuckle, and above all, for discussion.