# AI Agent Factories and Organizational Bottlenecks

**Podcast:** Stories Connecting Dots with Markus Andrezak
**Published:** 2026-11-02

## Transcript

Ich habe hier einen Produktmanager und fünf Engineers und so.
Und jetzt ist einer dieser fünf Engineers, der jetzt das Tool nutzt sozusagen, auf einmal irgendwie, der schließt jetzt halt 2000 Pull-Requests in der Woche und die anderen machen irgendwie drei.
Und dann ist sozusagen...
And from the one you have 600 questions to the product manager.
What is this feature actually thought?
Because I'm already done with it.
But you have to tell me how it was actually thought.
I have here 28 Annahmen get caught, because there was no way to do it.
And that was I was at 100 User Stories.
And then I got so 2800 Annahmen.
Which 5 should I keep?
Code-schreiben was nie das Problem.
Wir haben ja eh nur 2 Wochen dafür gebraucht.
Ihr habt in Summe 9 Monate gebraucht.
Connecting Dots.
Hallo und herzlich willkommen.
Heute habe ich bei mir zu Gast Benedikt Stemmelt.
Das ist so ein bisschen die Podcast-Serie Normale Leute machen abgefahrene Dinge.
Benedikt hat auch was Abgefahrenes gemacht.
Und zwar hat er, ich weiß gar nicht genau wann, eine Firma gegründet, die ...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
the agent for him prompt.
And that brings people to the people.
And relatively far.
It's not just the software production, but the agent's just a new ticket.
They look at what's for bugs that are made, they look at the right, look at them, they can resolve them, they can get them to the people, if they need them.
That's at least my half-caïve understanding.
And that's very interesting.
And I'm interested in what you're doing, why you're doing that.
And why does that interest me so much?
That is also this normal people make a lot of things.
While Benedict here is, and he sits here and he makes these things just like, there are people who say, this world doesn't exist.
And now it interests me that we're trying to say a menschic exemplar and a proof to have that it's this world.
And now I want to hear what that is, but especially, who are you?
Where are you?
And what has I triggered these things to do?
I'm in full of Spannung.
Yeah, thank you very much for being there.
I'm very excited.
Where do I get to go?
I think that's best to tell my life.
I'm not going to talk about it.
I'm not going to talk about it.
But a little bit is that, I think...
It's a good thing.
Yeah, I think I'm pretty early...
I had the first time contact with computers, for me relatively early.
That was 1999, I was 9 and that was already a Pentium 4, so that thing had already been really awesome.
750 MHz and Turbo-Taste.
Turbo-Taste.
Turbo-Taste.
The Turbo-Taste made it but longer.
That's important, that's important.
The Turbo-Taste was not to turn on the Turbo, but to turn on the Turbo.
That's very important.
We can talk about it when we were to talk about it.
I had to tell you in young years that I wanted to do something with software.
I built this thing with my father.
I built it to my father.
I built it to build it together.
Then I got to come back to a type.
That was actually just a computer.
Then I got to install this computer.
And then it was pretty quickly, that I just programmed with Adobe Dreamweaver.
I was like 12 or so.
Then it was here in Germany with the Internet.
Everyone wanted to have a website.
I had to build a website.
That was really a website.
Just like a website.
Just like a website.
Just like a website.
It was really a website.
There was also a phone number on it.
But that was done.
And, uh, that's what brought me then, as I studied it and then, uh, I got to be in a happy place with Otto, the Otto-Sand, or, uh, he's now just Otto, and, uh, I had then, in my life, always the chance to work in so many areas where it was really good.
I don't know that it's not going to be bad for developers.
I know the stories of developers tell how it's not going to be bad for developers.
But I myself had always the chance to have a big migration...
We had a very high autonomy, very cool tech stacks to use.
We had very early continuous deployment and worked on tests.
There are now people who don't do that, so to say.
That's 15 years ago.
And you had a bit asked, where does this come from, what we're doing?
I think what I realized quickly is that not everyone has this privilege to work in such a great environment.
And it's quickly been clear that Developer Experience is something that is very important to me.
And then came a few years ago that first time an AI tool was developed.
ADA was called.
That's probably not a lot.
And there was very slagart and very quickly aware that it will have a significant impact on our work as a developer.
And that has been triggered in the sense that I would not like to give this topic so.
Because I had then already the concern that people just use and interpret it wrong.
There are many more afraid of us.
We need fewer developers or what's going on with the juniors and so on.
And I wanted to play with it.
in this discussion.
And then I started to bring people to bring, how I think, how you should use these things, so that it can be a good developer experience.
And then of course, to a good product that you build.
Maybe we'll get there later.
And it was then kind of a lot of people.
that they wanted to talk about it.
And then it was a company that was created.
It was actually not planned to let us go to a big company.
It was more so...
It was then so created.
Because then people had more people asked me, as I could.
And then it was a company that was created.
And now I found a great Co-founder.
And we have many great customers.
And we discuss this topic with many great people on conferences and meetups.
And, yeah, I'm very happy about it.
Yeah, that's a very simple one, on the other side there is still something to do.
How was the development was?
Was you first talked about or was you could directly have something to do?
Because you could say, you could not talk about it.
Sometimes you could probably live in the way.
But you had a specific job in the sense of it?
Or was it more like you were there and it formed itself so all the way?
There are 100 ways.
A one of my backgrounds is that I always liked to other people what I brought to.
That was going to happen in school, because I was in certain areas of the field and then I gave there a lot of help.
Or at my university where I studied, I gave a pretty quick lecture.
I think I didn't have the Bachelor to finish.
They said they had to make a few seminars here at the time.
I was I was nie in vielen also ich es gibt so bestimmte Dinge die fallen mir irgendwie natürlich zu und da gehört halt viel in diesem Engineering Bereich eben einfach dazu weil ich auch schon so früh damit angefangen habe und dann war meine erste Idee war relativ schnell sozusagen naja ich will das eigentlich anderen Leuten beibringen also wie funktioniert das eigentlich wie benutze ich so einen Agenten und ich glaube zu Beginn hatte das auch noch also das war ja das war vor der Cloud Code Beta sozusagen ja also es war so da There was a little bit more needed to learn how to use it.
It's less than it's been to the tool itself, but in other areas it's more.
So it's been developed in a different direction.
But our product is always a classic training where we just say...
So you can use this tool to use this tool.
So there comes Halluzination.
So you can also see what can that, what can that not.
Many are in the user mode.
And we try to get people out of the user mode to get from the user mode, the user mode, the user mode.
And if you understand it, you can use it of course.
And then, of course, it's very much.
And at the beginning, I had to get some...
I've spoken to them about the training, about the group they want to learn about what that is and how it works.
And that was at the beginning relatively difficult, until it was no longer a question, because no one really knew what that was.
So that was a little bit like, what did you say?
We've seen here chat GPT, but that's a total scheiße.
So that was at the beginning.
And then I have actually...
Yeah, so demos auch gegeben dann ganz am Anfang und dann kam natürlich nach so einer Demo sofort Oh shit, das wollen wir auch wissen, wie das geht.
Also das sieht ja ziemlich krass aus.
Und genau.
Und dann kam irgendwann Ende 25.
So in Dezember 2025 oder so.
Da war ja dieser Moment, wo dann alle irgendwie, Claude Cote war gerade, glaube ich, in der Beta.
Das neue Opus irgendwas Modell kam raus.
Alle hatten Weihnachtsferien.
Und da haben sich bei mir ehrlicherweise alle Anfragen überschlagen.
Also so viele Anfragen, das habe ich noch nie in meinem Leben gehabt.
Das war wirklich echt abgefahren.
Also das war wirklich in diesem, wie alle sagen, in diesem Watershed Moment, Zeitscheiben-Ding, ist das alles passiert bei dir so Hals über Kopf sozusagen.
Ja, total.
Also angefangen hat.
that I myself really had to talk about it, was so Anfang, Mitte 25, would I say.
As I then for myself a little bit knew, that's a topic, what I want to offer.
And there came then, trough inaudible, but I think the brand was already present.
I have that also already under the name.
And then, end of 25, have the people started to ask, a network to ask, who can do that and who can help us online to google and so.
And I was probably already already present with the brand.
I think there are some parallels.
You can check if it really is the one that I build.
For me is the time also so.
But in my world, what does it start now?
What did you start to this watershed moment?
That's the belief that you can do other things as coding.
You see that.
And you see that in Coworks, you have something that has been created.
The level, in which you can treize, is not very clear.
The most are surprised what's going on.
What I'm surprised at the current skill, especially when other Wissenssets as coders, is the fantasy that you can use it.
The most people think that can't do it.
It comes from a chat GPT.
I prompt the slope before I go to the experience.
I can't know what it's going to happen when you start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start to start I go into something like Cloud Code, Codex, it's egal, what I'm going to do.
I'm going to start to think, what is the outcome?
Then I think, oh, what's the context I'm going to do?
What would I have to do as a person, that's good to make?
Then I'm going to say, maybe there's a few steps that I don't need to do.
And then I think, maybe there's a lot of steps that I don't need to do.
And then you get a lot of value, a lot of value.
And then you get a lot of value, a lot of value.
And that's...
mind-boggling, was da passiert.
Das ist für mich in meiner komischen Management-Strategie-Produktwelt obskur, dass es geht.
Und jeden Tag mache ich neue Experimente, das kann nicht funktionieren, aber hoch ist klar.
Und es scheint mir sozusagen das Grundprinzipien, und ich bilde mir auch einige Clouds zu haben aus Entwicklungstechnik, Kontextengineering, weiß der Geier was, dass es da Prinzipien gibt, die dafür sorgen, that these machines better for us work.
Can you tell us before you explain the things you were telling us, what was for you the moment where you were saying, oh, now things are different.
That's not just a bit of a code completion or something, but that's a different level.
That generates different things.
And what are the first steps, where it's a little, that it's now different?
Okay.
A very interesting question.
Also, I think at the beginning of the day was a bit like a better Stack Overflow.
You can say something like that.
You can say something like that.
You can say something like that.
You can say something like that.
You can say something like that.
You can say something like Stack Overflow.
Because Stack Overflow is a way to search for the right thread.
And then it was, that the IDEs had more ebent, so that they had better self-reported.
I mean, self-reported, that's been already been done for 25 years or so, or even more longer.
But it was a bit better or a bit longer for all the methods and not only for single-reads.
I think the moment...
The first tool that I really got to get into the mouth was CLINE.
I don't know exactly how that is.
CLINE.
And that was the first time with a loop.
Where you don't always have to manually the next loop.
Because you have to copy it or write it in the same way.
But where you have to say, on the basis of the inquiry, the first time the agent says, I have to read it out and make the next information.
And then it was the next thing.
very magical.
That is very long ago.
Two years ago.
A little longer.
It was after Ada.
Ada could not.
And Kleene could not.
And then...
That was the first moment I realized, okay, crazy.
And then it was just happened, the context-faces were still small, that means the agents and models had to maintain, what they actually worked, if so a context-fenster was full.
Then there was always a panic.
Then you didn't know what was going on.
Where am I going?
And help.
And then...
That's also something that the normal consumers could not understand.
If they, I don't know, before a year or so, chat GPT had a full job.
Then he was completely verwirrt and had some answers.
You didn't know what's going on.
You didn't know what's going on.
You didn't know what's going on.
You didn't know what's going on.
You didn't know what's going on.
You have not noticed that you have a mistake, because you have to do it.
But the effect is that.
Yes, exactly.
And I think that the, what's exciting is, that the models, they are trained with the Chat-Low.
What is also also for the way that they are more optimized to optimize the way we use the Chat-Low.
What has that done.
I'm going to give you a instruction, then he does.
And then I can do another topic.
I'm going to work on this other topic, what has to do with this other topic.
But I'm going to do it again.
And then he says, that he does not make the same context, but also makes it a separate context.
And then comes this advanced, that he is more than an orchestrator.
But that comes, I think, because they just learned how we actually chat with them and how bescheuered we actually are.
I have to say that to the topic of three hours, where you think, yes, it's now.
It's about to say that we don't work linear.
Yes, exactly.
And then they had to find that intern, how they actually went with such a chat-through.
And in principle, that was a kind of second moment, where then, well, it was always, it was always been very strong with this concept of these subagents.
That was also the moment in the end of 1925, I think.
Where the idea was, let's not just one loop loop, but we can also, instead of having a prompt to give an agent, the agent can also have an another agent to set up, the agent can do something to do and the answer can go back to the main agent.
And that was at the end of 2025, that was a Alpha-Feature of Cloud Code, Subagent.
Today is of course normal.
When I say, make the 28 topics, then spawned 30 Subagents.
And they did their things.
And that was the time you had to organize all of them.
You had to say, use this for a Subagent and this in the other.
But that was really...
There was the first time this problem with...
That he forgot, what he worked with.
organized.
it's the experience not there.
And at Amazon is also so that they have a clear understanding that this world exists and there is.
And at the end it's also totally different.
And actually it's so that...
in the state of the civil society, the people who had no contact with.
We're talking about a small bubble, which is what you're talking about.
That's a great thing.
That's a great thing.
It's a great thing.
It's always exciting.
It's always important to analyze.
Also for our business is that important, because we ask ourselves, are not all the people out-trained?
And I think the last statistic was that under 2% of the work of the population had with contact or so.
The barrier was even small.
They had already used to chat GPT.
So that was under 3% or so.
I'm not sure if I'm not sure.
But yeah, it's actually so that that relatively few people are being able to do.
But yeah, they have then then irgendwann realized...
It was a long time, actually, the first ones were not the engineers, which it was so.
The first ones were the marketing colleagues, so text generating, product management, not in the sense of product, but more in, I have a sort of a sort of a title and a description and so in e-commerce or a picture.
That was the first wave there came.
That had difficulty running around.
It was already cool, but not like wow.
And then they came to this trick, that it works for the engineers.
Because the result is better validable than a picture.
That can be difficult to validate and evaluate.
And then it went on and this spiral is so incredibly strong.
And that's why you said that, what you told me.
that you start to start to set up, the Ordnerstrukturen to set up.
That's kind of...
It's completely ridiculous, that you have to put a product manager or someone else's...
Yeah, you must have a Git repository to set up.
It's not a good UX for these people.
If I think of the developer experience, for these is it really great, because they say, I can't work in Git and so on.
Aber für so einen typischen Manager ist es ja keine super Manager UX.
Also das ist ja völlig...
Ja, ja.
Bei mir geht es ja so weit, dass ich sage, so Entscheidungen müssen in die Infrastruktur und so.
Oder keine Ahnung, wenn der Chef nur Strategie ausgibt, sage ich, die Straight GMD muss halt in ein Repo rein, das muss überall verteilt werden und so.
Vollkommen selbstverständlich.
Und anders, wie soll das gehen?
Und dann hast du dann ein Gegenstück, das sagt so...
Ja, aber der Schäuble hat sich ja auch E-Mails ausdrücken lassen.
Dann müssen wir da auch so einen manuellen Prozess irgendwie finden und so.
Und so ungleich verteilt ist das gerade.
Ich glaube, das ist total irre an der Zeit gerade, wie ungleich das verteilt ist, dass die eine Welt so selbstverständlich davon ausgeht, nicht nur, dass es funktioniert, sondern was das auch bedingt.
Wenn ich es ausnutzen will, dann standardisiere ich plötzlich meinem Unternehmen Prozesse, Dokumente, Templates.
I can't tell you that all the context can be done and the same thing.
And others say, why?
What should work?
What helps you do?
And that's totally irrelevant.
Not only how it's distributed, but how the speed is distributed.
How it's distributed on each other.
The companies, with which I work, are mostly so that a few are very far away.
A few are interested, a few hoppings behind it.
And many say, not with me.
And that's a good company.
There are reports about Google, that that's also very surprising.
And then, that's not in the company, but over the company, again, there are so many companies that don't even bother me with it.
And the other people say, oh my God, what I'm going to do with this?
And the others do it already.
That's a total weird.
on every level, every uniform is available and speed and so on.
Yes, in the fact.
What I also find interesting is the point, where we are now often and where we are also very much, one is the individual use as a person for my job and the next is also the use as a team.
And what effect has the acceleration of the individual person on that?
teamgefüge und auf das die arbeit im team gerade in der software entwicklung hast du das ja auch da sagst du okay ich habe hier einen produkt manager und fünf engineers und so und jetzt ist einer dieser fünf engineers der jetzt diese dass das tool nutzt sozusagen auf einmal irgendwie der schließt jetzt halt 2000 pull requests in der woche und die anderen waren irgendwie drei und dann ist sozusagen And from the one you have 600 questions to the product manager.
How is this feature actually thought?
Because I'm already finished.
But you must tell me now, how it was thought.
I have here 28 Annahmen, because there was no way to do it.
And that's what I did with 100 User Stories.
And then I got 2800 Annahmen.
rein encoded worden.
Welche fünf soll ich behalten?
Ja, genau.
Und welche sind denn jetzt die richtigen eigentlich?
Und ja, was ist eigentlich das nächste im Backlog?
Weil irgendwie habe ich nichts mehr gefunden, was da jetzt gerade so implementiert werden könnte.
So, sag mir doch mal.
Und das ist eigentlich super, super lustig, weil jetzt sind wir volle Kiste eigentlich in einem total nicht technischen Thema drin.
Weil jetzt fängt wieder komplett Organisationsentwicklung von vorne an.
Also welchen Kontext willst du teilen?
Willst du überhaupt teilen?
Entwicklung ist eine, weil Produktmanager ist es selbe.
Das Beispiel, das ich mir gebe, mein normaler Job war ja, ich werde wo reingerufen und die sollen mal eine Roadmap machen.
Und dann hast du, keine Ahnung, 15 Produktmanager und die sagen, Moment mal, was soll das?
Roadmap.
Wir machen keine Roadmaps.
What?
Ihr solltet eine machen.
Ja, aber wir machen nur Outcome-Base.
Okay, dann Outcome-Base.
Ja, nee, ich muss ja immer eine terminliche Feature-Roadmap machen.
Ach so.
Keine Einigung.
10, 15 Leute, keine Einigung.
Daraus sitzt ein Stakeholder, der sagen, ja, genau so ist es.
Ich kriege immer irgendeine Roadmap.
Niemand einigt sich.
Was soll das?
Und so.
Und wenn du jetzt dabei bist, wir sind super nett, wir machen Outcome-Based Roadmap und so, dann ist ja, ja, aber wie sieht die aus?
Welches Template?
So, what?
Niemand hat sich geeinigt.
Ja, wir haben eigentlich auch seit einem Jahr nicht mehr in einem Raum gesessen.
Ach so, ja, das erklärt viel.
Wenn du jetzt halt AI hebeln willst und sagst, ja, wir haben eine Skill, die diese Roadmaps bearbeitet.
The template needs to be able to use.
What is the template?
And what you do now is you're going to use a complete work.
That's a very exciting point.
I have a lot of experience in a positive nature.
Because what you describe is...
actually a problem.
The people have already.
There is not a problem.
And because of all, every individual has a little bit of his own workflow.
And that same problem I have of course with AI.
And, spannend, it is now so, that people come in the idea of the people, that they might have a template to build.
And if you don't make AI, it's like...
It's totally okay, if everyone else makes it.
Because nobody has time for it.
And everyone is like, well, it's just so.
And then really the Mehrwert is so crazy.
And at AI is always so plakative.
Because everyone is with documents around themselves.
And then like, maybe we're going to put a format on it.
So we don't always get a new document.
The important question is actually one after that.
If it's not for the person, but if I say, I want to automatically automatically automatically, then I have to automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically automatically But I need to be a certain degree, we want to always have these artifacts in the way and we want to review them.
And first when I trust them, I always believe that he does always make these artifacts, then I can always say, hey, he does already do this for me and I don't have to worry about it anymore.
Now you have said something funny, that's not everyone.
And you're also like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're like, you're a whole lot of engineering, but also because I am end of the process trust, the that's what I build.
That's what I don't want to do at level 3, where I trust the whole thing, but I'm going to do it step for step and say, oops, that works.
And what I think of my opinion is this, I don't look at it anymore.
I know, it's not that it's been a hundred thousand times.
1.000 Mal wird es auch klappen oder 100.000 Mal wird es auch klappen.
Und viele Leute wollen das, glaube ich, überspringen.
Und ich glaube, heute kann man auch schneller springen.
Aber dieses, ich kriege Vertrauen in meinen Prozess, ist halt trotzdem noch ein Ding.
Und das wirkt so obskur menschlich, esoterisch, dass man diesen Biestern vertraut.
Vielleicht kannst du das ja nochmal erzählen, weil das wirst du ja oft treffen.
Also das ist auch was, was mir ständig in Organisationen begegnet.
really this trust.
And people who say, that's not, that's all Slop, that's all Kram.
But then still over practice, and that's what you can't take, it's just over practice and trying to keep up and keep up this weird trust.
From which I don't know anymore, if that's true word or so.
But something that's a role.
Yeah, I think trust actually quite as a word.
That's a bit like the jurists.
You can do everything else to do and write and write and write.
But at the end it's just like a trust question.
If I say someone, I make this DSGVO-conform and I have zero data retention and so on.
That's all in the contracts.
If that's AWS really does.
Who knows?
Yeah, also, klar, irgendwie versuchen die das auf bestimmten Ebenen dann durch Sachen abzusichern, durch vielleicht externe Prüfer oder irgendwie sowas.
Aber am Ende ist es dann ja eine Vertrauensfrage.
Ich vertraue auf Basis der ganzen Dokumente und externen Prüfungen, sage ich, naja, gut, dann werden die das schon ordentlich machen, sozusagen.
Und ich finde, das ist bei den Agenten sehr ähnlich, bei Menschen auch.
Wenn ich jetzt einen neuen Arbeitskollegen habe und ich höre, ja, der...
He had a doctor something like this and so.
He had his five other company.
Then I had a little bit of a job.
I had a little bit of a job.
I had a kind of a kind of trust in the company.
I have a kind of trust in the company.
I would know what he does.
Then I would know what he does.
Then I would know what he does.
Then I would know what he would know.
Then I would know what he would know.
Then I would know what he would know.
Then I would know what he would know.
Then I would know what he would know.
Bei den Agents ist das ja ähnlich.
Erstmal muss ich den ja kennenlernen und muss ja wissen, wie reagiert er jetzt hier eigentlich auf diese Aufgabe?
Und das geht ja alleine mit dem Zutrauen los.
Das ist ja auch eine Form von Vertrauen, dass ich dem zutraue, dass er diese Aufgabe überhaupt erledigen kann.
Und dann würde ich als Tipp geben, erstmal einfach dem alles zuzutrauen.
Das ist eigentlich ein ganz guter Start.
Und dann kann man ja gucken, ob man dann auch zufrieden ist damit.
Und was man vielleicht ändern muss.
Also das ist ja auch bei Menschen würde ich ja jetzt auch immer versuchen, dann den Leuten auch Feedback zu geben und zu sagen, Mensch, das hat jetzt nicht so gut geklappt.
Woran lag denn das eigentlich?
Und jetzt nicht gleich sofort zu sagen, ja, jetzt ich vertraue dem nicht mehr oder irgendwie so.
Und das geht mit Agents eben auch.
Ich kann ja dann auch sagen, was sehe ich jetzt, wo ich gemerkt habe, dass da was nicht funktioniert hat.
And what do I do with the agent to do, so I can trust him?
If I can do a external test, where you say, there's a second agent to do, and before I ask, comes there a second agent to do, the give him feedback.
And then I have a higher trust in him, because the result will be better.
And there are two things in me.
The one is, where you said, first of all, everything is to trust.
I think that's a positive and negative.
What for me is the parallel, that's what I said before, that people often don't have the idea of thinking about it.
That could just go with it and so on.
And the second step is, if they try to overthink it, they say, how could I do this?
And what you today really do is, describe the goal, describe your results and look first, what's going on.
And if the result is not possible, you can always improve it.
I make the obscure, complicated experiments with these things and then almost always comes something out of the way, then I give them two iterations, then I make it really, what I want.
But a normal Bureumenschen-Schritt would be like, I think that first of all, I'll see what could happen and I'll give the whole Rand-Bedigungen with it.
And what we forget is that the models are so good and so much, ahem, a dangerous word, common sense, that we...
ganz vieles gar nicht mehr explizit erklären müssen in dem ersten Versuch.
Das ist für mich auch so eine Übungssache, dass man lernt, gar nicht zu viel mitzugeben erstmal und dann verfeinert.
Also ich sehe da auch so zwei Dinge gerade drin, die ich interessant finde.
Das eine ist ja, für mich ist das ein bisschen auch, also man soll zwar nicht immer so anthropomorphologisieren oder wie das Wort heißt, aber im Prinzip ist das ja so ein bisschen auch wie Micromanagement versus nicht micromanagen.
Also ich kann jetzt meinen Mitarbeitern einfach sagen, das ist das Ziel, das wollen wir jetzt hier irgendwie mal erreichen und dann mal gucken, wie der das so hinbekommt und dann davon ausgehen, dass er mich schon fragen wird, wenn er noch Input braucht oder...
So, ne?
Oder man im ersten Gespräch schon mal zu sagen, das ist unser Ziel, wie erreichen wir denn das jetzt?
Was sind denn die Sachen, die du noch brauchst?
Und dann, ah, die und die Sachen fehlen mir, kannst du mir die mal sagen, so, ne?
Und gar nicht selber zu überlegen, was könnte er mich vielleicht fragen, was gebe ich ihm vielleicht?
Und dann so, ey, ich habe dir hier schon mal so diese 28 Ordner vorbereitet, so, arbeite die mal durch und dann mach da mal irgendwas.
Und dann ist das wahrscheinlich auch im Zweifel.
It was not necessary to work, all the Ordner to work, to achieve the task.
And the other thing is, there was a fascinating part of the topic, where I actually have to be in the game.
There comes this ominous Context Engineering into the game.
I think that this is a very high-traumatic term.
But I can find the GitHub Copilot and the Microsoft Copilot a cool example.
Because the Microsoft Copilot was a long time really bad.
He was really bad.
And that problem, I think, is that he was just in the Microsoft Cosmos just a little bit of access to these things.
If you asked, what is with the and the Thematics now?
then he had no chance to answer that, because he had only in the Teams chat, where you asked him, the information he had.
And his model.
But what does that model know already, what happened to you in the company?
And the new version, the they just brought out, I think the is a bit rebranded in this...
Copilot365 or so, yeah, I don't know, how to do that.
But he's actually even very, very good.
I've heard many stories about how good he works.
Because he's just been able to use the calendar, email, all the chats and so on, the same kind of triggers as this person, what this person could also be able to search.
And if you ask, what is here with the topic, then pullt he's just the transcript, then pullt he's just the emails, then pullt he's just the calendar, and so on.
And then he says, yeah, that and that is happening.
And you think so, yeah, okay, cool.
Also, the answer is basically, it's not possible.
And there is also a UX question.
I think that they have better gotten, that you can't get any more...
Ordner strukturieren muss und den MCP-Server zu seinen E-Mails anschließen muss und so, wo du dann schon wieder denkst, was ist denn das?
Und das haben die jetzt halt in ihrer Suite zumindest hinbekommen, dass da deutlich UX-mäßig besser zu integrieren quasi.
Gestern hat ein schlauer Mensch mir sein Bild von Kontext nochmal gegeben, das ist vielleicht ganz hilfreich.
Also wir wissen, dass alle die LLMs hier...
This generisive Wissen, this world's wisdom three times, everything cool.
But if I talk with them, they answer them just because they don't know about me.
And in principle the context, you say that you put your point of view in your mind and say, everything cool, what you know, but look at this perspective and now work so with me.
And I think that's good.
And what it basically does is...
Also das Beispiel von ihm war, dass er irgendwie so Fachbücher schreibt und ganze Fachrichtungen sortiert und so ein Kram.
Und er sagt, es gibt so viel undokumentiertes Wissen, wo die keinen Zugriff drauf haben.
Und meine Leistung ist, dass ich das als Kontext dieses undokumentierte Wissen reingebe, sodass ich aus meiner Richtung denken kann und das alles weiß.
Und dann findest du auch diese Zusammenhänge wieder und so.
Ich glaube, das ist so mal noch ein ganz hilfreiches Bild.
Ja, oder wie wenn du jetzt sozusagen, also...
I think the analogy is always nice to be able to start with a new company.
Then you know how it was in your previous company and how it was going to be.
Because I learned how it goes.
And I have also a certain kind of human understanding how I would imagine how I would imagine how it would be.
And then you are already at the LLM.
And now you say to someone at Mercedes or whatever.
But our process looks like this.
And then you think so.
Why?
That's a different question.
But then you do it like the process looks like the process.
And that's the LLM doesn't like it, without you giving it to him.
Now I have another question about the trust.
And that's why I believe you do not trust.
I believe that you do something to do, to help you to reduce your trust.
I believe that you do something to do.
I believe that you do something to do.
You have already said that you have already said that you to the agents, the building something, a agent that has to do something, the pruning of that all.
You have already got a lot of trust, but actually you say, I know the questions I would ask you to do something, but that I would do something else.
That's also an agent.
Actually, I work in such loops.
And maybe you can explain that.
I think that was the next step.
We can do that.
dass diese Loops beherrscht werden.
Und das führt dann sozusagen irgendwann zu den Factories, die mehr machen als die Loops.
Aber vielleicht kannst du ja mit den Loops nochmal anfangen.
Ich glaube, da fängt das für viele Unglaubliche sozusagen an.
Dass diese Prozesse gar nicht mehr, also dass du nicht mehr in the loop bist, sondern on the loop und quasi eher reingerufen wirst, wenn mal was schief geht und so.
Ja, also genau, da werden natürlich jetzt ein paar Sachen von dir auch durcheinander geworfen.
Sehr gut.
Yeah, also, it's a very, very many Loops.
That's the problem.
There are very, very many Loops on very many levels.
And you have to sort it, on which level of which level I speak here.
I don't know if it's now useful, all of them through.
But that LLM, I'll give you a little overview.
Also, that LLM itself is also a loop.
Because it's always the next token, and then the input...
plus das eine nächste Token wieder reingegeben und dann wird das zweite nächste Token ermittelt.
Und dann wird sozusagen der Input plus diese zwei Tokens reingegeben und dann kommt das dritte Token.
Und das ist ja auch schon so eine Loop.
Die loopt so lange, bis am Ende quasi so ein End-of-Token-Token kommt.
So nach dem Motto, jetzt habe ich hier keine Tokens mehr, jetzt ist mein Text sozusagen fertig generiert.
Also schon auf der niedrigsten Ebene ist da quasi eine Loop.
Mit der haben wir natürlich eigentlich nichts zu tun quasi, weil das so funktioniert halt das LLM quasi.
And then the next is in the agent.
That's what I had mentioned earlier, what was then new was.
At the beginning of the chat GPT, I had a question on the LLM and then only this LLM loop was I just told.
And I got one input text, the one output text was generated.
So that's it.
Input, output.
And if I have then in my input things forgotten, So, dann kam halt dieses ganze Prompt Engineering, du musst das alles schon in deinen Prompt, weil ich habe ja nur diesen einen Prompt und dann muss da irgendwie so der mega geile Prompt sein.
Prompt Master.
Ja, genau, genau.
Und das ist sozusagen, genau.
Und bei JetGPT war es zumindest noch ein bisschen besser, weil wenn ich dann eine zweite Frage gestellt habe, dann hat er ja zumindest meine Chat-Historie davor gehängt sozusagen und hat nicht nur dann die zweite Frage hingeschickt ohne die ersten.
So, und dann hat man darüber ja dann auch so ein bisschen in the Prompt Engineering, so a conversation, or so.
There was already a history there, so the results were better.
And the loop is actually, that is always just this one agent, so the LLM loop.
And then came this idea, let's just go to an agent.
And the agent, the agent, the question, so that we don't ask anymore, but the agent, and we ask the agent.
So, und das ist ja mittlerweile der Standard, ja, also man kommt da gar nicht mehr drum herum, weil das alles so funktioniert.
Und der Agent selber, der hat jetzt eine eigene interne Loop auch sozusagen, ne, das ist so eine andere als das LLM, ne, das ist jetzt die von dem Agenten.
Und der Agent kann im Prinzip intern jetzt selber entscheiden, antworte ich jetzt dem User und sage, jetzt brauche ich Input vom User, ich bin hier fertig.
Oder frage ich selber nochmal beim LLM nach?
Und das ist, glaube ich, so sehr spannend, weil jetzt ist natürlich die Frage, wo kriegt der jetzt Informationen her?
Also wie würde das ablaufen?
Man stellt jetzt eine etwas größere Anfrage und dann ist die erste Anfrage von dem Agenten an das LLM nicht die Frage direkt, sondern sozusagen, wie gehe ich jetzt diese Frage an?
Welche 25 Schritte brauche ich denn da jetzt?
Und dann antwortet das...
LLM mit hier, das sind die 25 Schritte, die du jetzt brauchst.
Und dann weiß der Agent schon mal, aha, jetzt kann ich hier 25 Schritte durchführen und fragt dann natürlich für jeden Schritt wieder das LLM sozusagen.
Und das passiert alles so im Hintergrund, das kriegen wir eigentlich gar nicht mit.
Und das ist aber auch schon wieder nur ein Agent.
Jetzt haben wir ja mehrere Agenten, weil wir ja Subagenten auch haben.
Das heißt, dieser eine Agent kann sich jetzt auch entscheiden für einen bestimmten...
that's why I'm asking now to another agent to ask, who has a loop.
And he has also a loop.
And that's again, also these several agents, that's now the loop, where I sit.
So where I say, which agent will I have, to fulfill a certain task?
And that's what I'm not quite sure to define.
Because these agent themselves, so a general agent, The person is now in the way, that then himself already down to break and say, okay, I need three people, they do this and two, they do this and the one who does this and when the one answer, I can ask them again and so on.
And then we come to the same thing, what we maybe as factory describe, but that is then again behind.
Also, that is actually everything, if I'm with him chat, that happens then all of a sudden.
I could say, I have to write this doctor work.
Please organize your agent fleet, to write this doctor.
What do I do all of this do to do?
And then he goes, then we'll send 30 researchers to do it.
And here we go.
And then we'll do it all.
And then we'll do it all.
And then you'll wait.
48 Stunden oder noch länger.
Und dann sagt er irgendwann so, oh ja, jetzt bin ich übrigens fertig.
Ich habe die Doktorarbeit geschrieben.
Hier ist die Quellliste.
Hier sind die unabhängigen Reviews und so.
Und das macht er dann alles schon.
Da braucht man ja auch über den Prompt keine Gedanken machen quasi.
Er muss halt nur irgendwo den Kontext herkriegen.
Und wenn er da keinen Zugriff drauf hat, dann wird es halt nichts.
Aber wenn er Zugriff auf den hat, dann legt er da halt los sozusagen.
Ja.
Und Now, when you say, the factory is still a loop on top or a level higher, what does that mean physically?
Now comes again this trust in the game, because you had already this trust question, if I really trust.
If I'm in this Chat-Modus and also in this Chat-Modus with the 20.000 Agents and so on, then I'm still there still.
I say to him now, I want to have this Doctor work and then I look at him and give him feedback.
Or I say, I'm going to take a look at each chapter.
And that's just like, I would say, how far I have this automatized?
I would say that is a Teil-Automation.
Because certain steps will be...
You have already mentioned Maturity Levels, and I'm going to be on Level 3 or 4.
I'm not in every loop, in every one of the individual.
And that is what many of the companies, with whom we work, actually this level is very well.
And then you see...
3x, 4x Produktivitätsgewinne.
Also das ist schon so.
Wenn ich da schon loslasse, weil das ist ja schon auch ein Loslassen.
Und da ist auch schon eine ganze Menge Vertrauen mit involviert.
Und da kann ich aber jetzt ja immer noch ein Code Review machen als Entwickler und dann sagen, okay, ich gucke mir den Code danach an.
Das wird schwierig, weil es sind natürlich dann irgendwie 2000 Pull Requests die Woche.
Ob ich die dann durchgereviewt kriege, ist halt die nächste Frage.
The idea is in this work already always so, well, at the end I'll look at it again.
And now I'm going to have a level higher.
I will not have to look at it again.
I will say, okay, I will maybe the Mitarbeiter instruire, something to do.
And then I can't rely on them to do that.
And I don't have to look at it again.
I'm always happy to have it again.
Because that's...
And actually you're now CTO.
Genau, jeder eigentlich.
Also jeder ist CEO, würde ich sagen.
Also jeder ist Macher.
Also jeder kann jetzt das alles machen sozusagen.
Und du gehst sozusagen in Makrokontrolle und Delegation, aber eben nicht in Mikrokontrolle und guckst dir das einzelne Arbeitsergebnis an.
Genau, genau.
Also ich muss natürlich schon irgendwie am Ende sicherstellen, dass die Arbeitsergebnisse die Qualität haben, die ich haben will.
Aber das mache ich ja, wenn ich delegiere, auch nicht, indem ich sozusagen...
I'm going to go to the person and give me a little bit of the stick.
Wait a minute, I'll do it myself.
That would probably be many CEOs, which I know, do so.
And I'm also very, very traurig, that I don't even know what to code.
That was I was already before AI, because I was also not more Entwickler.
And I'm now, witzigerweise, I'm glad to code more, although I don't know what to code.
The idea is to, like you said, to go back to more and more delegating.
And I have to start with, and the Entwickler call it factory, I find the word not so cool, because, yeah, who wants to go in a factory?
Then I say, you work not in the factory, that's exactly the idea, but you build a factory, the for you works.
And in the factory work then only Agents.
But trotzdem, yeah.
I mean, other people say that to be a process automation or so.
That's now a process of development process.
And then you have the special word that you have selected, that you call it factory or so.
And to build a factory, I have to observe a lot of things.
How does this agent work in my, I say, as we just discussed, if I give him a grope task.
What makes the right and what makes the wrong?
What information is missing?
What is it?
Like with a person who says, I've seen you see how you work.
Also Shadowing.
And I was like, I'm going to say, could you do that differently?
And then you have a better result.
If you try to program it in this factory, then you can do it again.
You can do it again with Java, but it's like that.
But exactly there comes it again.
When I build a factory and build together, which works out of the way.
And there is trust in it, that I have to say, I trust my factory so is set to control myself and also has a control instance, which brings me in sum to the same way, that I can trust the whole thing.
And in principle is your job now more than just a process-reaching.
Yes, exactly.
Could you describe it so.
Also das Problem, wenn ich jetzt so an ITIL-Prozesse denke, also bei Prozessverbesserungen denke ich immer sofort an ITIL irgendwas, das habe ich in der Uni das letzte Mal gemacht, da ist bei mir das Developer Experience Thema eher so minus 1000.
Weil da hat man ja sehr feingradular geguckt, wer macht jetzt welche Schritte und dann malen wir das mal auf und so.
So macht man das eigentlich ja gar nicht mehr jetzt mit den Agenten, sondern ...
That's on a much higher Flugebene.
So, of course, you could do that.
You could definitely do that.
You could do that.
You could do that.
You could do that.
But the problem is that, as I go into the middle, and that was at Eitel also always that problem, the faster it will be.
So if I'm at a point of time from the process, then...
Then it's in a warning and then, yeah, that's going to happen, because the guy behind, the guy behind, the guy behind, the guy behind, the guy behind, but not because of the IT-Provents.
And that's what it's for the Agents.
This Non-Determinism is a feature.
And if I now the process is crazy, I'm going to be back with the Determinism.
The I have already already.
What's funny about what happens like Dynamic Workflows or so.
Teilweise Loop-Steuerung gibt dir dann wieder deterministisches, damit da sozusagen kein Quatsch passiert und so.
Genau, das ist diese Vertrauensfrage dann.
Also wenn ich Teile jetzt quasi wieder deterministisch machen kann, bei Code-Review ist es dann ja Unit-Testing, Linting, solche Sachen, dann kann ich dem Linter vertrauen.
Das habe ich ja sonst auch schon immer benutzt als Entwickler.
And the product gives feedback to the model.
And then the model says, there is now a Linting-Regel broken.
The need to repair them, so that we can see the quality of all the Linting-Rules.
That we can see them.
That is a long discussion.
It's very exciting, if you allow me that.
The Linter, that's very deep in detail.
But so a Linter, the developer gives the developer...
the information, for example, the method is too long.
You write a method in your code and it has 15 Zeilen.
And because that's a smell, that's the structure and the architecture of your code is not so good, then you make a Linting rule and say, they should have maximum 10 Zeilen.
So to say, you have to get some feedback.
So to get as a developer, you have to look at it.
So, was macht denn jetzt ein schlechtes Modell, wenn es eine Linting-Ruhefehler bekommt?
Die Methode hat 15 Zeilen, die ist zu lang.
Ja, der macht Methode Part 1, Methode Part 2.
Methode Part 1 ruft Methode Part 2 auf und beide zusammen haben wieder natürlich noch mehr Zeilen, aber einzeln ist die Linting-Ruhe gefixt sozusagen.
Das System gehackt.
Das System gehackt, genau.
Metrigen gehackt.
Genau, das wollen wir natürlich nicht, weil das Feedback ist ja nicht...
The Zeils are too many, make two methods with less Zeils.
For example, there are some ways to do that.
If I'm not a junior, I'm not a junior.
But actually, I just want to get this indication that I should get over the Architektur of this code, Abstractionslevel and so on, I should think about it.
And there is a very exciting topic.
There are models that are better trained.
They can do that then.
They make a method 1 and 2.
It's a question.
How do I train this model, that it comes to this idea?
Not so easy.
But what's even cool is, there are also Linter.
There is a very cool project that I can recommend.
The one I can recommend, the one with such Lintern works.
There is not only the wrong answer, you have here 10 letters or so, but then give you a few information back back, what could be done.
That could be done to say, that here is a abstract.
Or you are too specific.
And then the model is of course with the right information and goes immediately up.
Stimmt, da muss ich da nochmal drüber nachdenken.
Und das ist ja auch wieder eine Art von, wie gebe ich jetzt eigentlich gutes Feedback?
Also das ist ja bei Managern die gleiche Frage.
Also sage ich jetzt, meine PowerPoint, die sollte aber nur 10 Seiten haben.
Oder gebe ich quasi mit, worauf das jetzt gerade deutet, dass die PowerPoint irgendwie 46 Seiten hat.
Was könnte denn das jetzt vielleicht für das Thema bedeuten?
Wir sind zu tief reingegangen, nicht auf der richtigen Flughöhe.
Du kriegst ja auch qualitatives Feedback.
And I think that's a bit surprising.
You know, you're going to say that, you know, that these weird algebraic beasts are kind of feedback, so sharp and structures, but are quite a bit of a difference.
In the meantime, the models are so good, that they just understand this qualitative stuff.
Yeah, exactly.
That's actually much more important.
That's qualitative stuff is much more important.
But I think it's also a disease in people's lives.
So ein Junior-Entwickler, der hat ja das gleiche Problem.
Und ich hatte, wie gesagt, wieder zurück in meine Zeit, ich hatte damals bei Otto, als ich angefangen habe, ganz tolle Kollegen, die wirklich sich auch die Zeit genommen haben.
Und das Environment hat das auch zugelassen.
Die Firma hat das auch zugelassen, dass ich auch verstehen konnte.
And asked why we do that now?
Why do we do Domain Driven Design?
Why is that for us important?
What are the principles of the book?
Let's read it through.
Okay, now I understand what that is.
But why do we do that?
That why?
And that's not our practice in our...
Tools und in unseren, also das Tool ist eigentlich gar nicht geübt, ein Warum zurückzugeben und wir sind das als Menschen glaube ich auch total, also weil ja alle so dann wieder Mikro-Management machen und sagen so, Alter, hier so 15 Schritte, lasse die Schritte, mach das mal.
Ja, abgefahren.
Also so jetzt habe ich so eine Factory, ich bin so eine Firma, ich habe das von dir gelernt, alles cool.
Was ändert sich denn für mich jetzt in der Software-Erzeugung und vielleicht sogar darüber hinaus?
Also mein Ding ist ja immer, früher habe ich das schon gesagt, also ich stehe heute Morgen auf, habe zu warm geduscht, habe eine Idee, gehe ins Unternehmen und sage, oh, lass uns das mal bauen.
Irgendein armes Schwein, mach das wirklich und so.
Und überlege, dann kommt dieser ganze Prozess mit Entscheiden, Präosieren, weiß der Geier und so.
Was man so macht im Leben.
Und dann in nur neun Monaten ist beim Kunden oder so.
Und davon war dann irgendwie zwei, drei Wochen war Softwareentwicklung.
Und das ist ja eigentlich, was wir gerade diskutiert haben.
Diese Softwareentwicklung schnurrt zusammen.
That makes a surprise.
And that's the revolution, about which one speaks.
Actually, it's still more.
Also, in the company.
Also, um this Schnittstelle drumherum.
What's your view there?
Yeah, very exciting.
Many say to the right.
Code schreiben war ja nie das Problem.
So wie du schon sagst auch eben in deinem Beispiel.
Wir haben ja eh nur zwei Wochen dafür gebraucht und ihr habt in Summe neun Monate gebraucht.
Das ist wie wenn ich bei meinem Hausbau zugucke.
Die haben eigentlich auch nur sechs Monate gebraucht, das Haus zu bauen, aber blöderweise waren die Übergaben mit den Gewerken und die Freigaben und deswegen haben sie dann doch zwölf Monate gebraucht.
Und das ist bei den Agents natürlich auch nicht anders.
Also nur weil jetzt das Engineering dann auf einmal anstatt von drei Wochen jetzt drei Minuten braucht.
I'm not really faster.
I'm just 9 months minus 2 months faster.
Or minus 3 months.
What's that problem?
That's the problem.
That's the problem.
That's the problem.
That's the problem.
That this fall.
This fall.
This fall.
This fall.
What's always like a bottleneck.
That now.
That you can't show.
It's not a bottleneck.
There will be a lot of weird things in the company.
I mean, the book of Elliot Goldred to Bottlenecks and...
Oh, now I'm not going to...
The Goal.
The goal, genau, aber den Begriff, Theory of Constraints, genau.
Das gibt es ja jetzt auch schon etwas länger und sozusagen, das ist ja bei vielen Menschen immer das Problem, dass wenn es dann nicht plakativ genug ist, dass es dann einfach nicht angegangen wird.
Und ich glaube, mit den Agenten haben wir jetzt, genau, ist das jetzt wieder so ein Thema, was jetzt auf einmal so plakativ gemacht wird, wie das, was du vorhin erzählt hast, mit diesem Artefakt, was fehlt.
Und jetzt passiert auf einmal Folgendes, wir haben diese Factory.
Und ich will diese Factory jetzt ja mit einem Requirement füttern.
Und dann ist ja die Hoffnung, dass am Ende, also nehmen wir mal an, ich habe eine geile Factory gebaut.
Ja, das kriegt man auch hin.
Also ich schwöre, man kann geile Factories bauen.
Und dann kommt da am Ende produktiv nutzbarer Code raus, auch von der Qualität.
Und Qualitätskriterien ist auch nochmal eine ganz andere Diskussion.
Ja, aber genau so.
Und dann kommt halt das Deployment, auch am besten CICD automatisiert.
Also ist im Prinzip meine Factory wie so eine verlängerte Deployment Pipeline.
So, and now I will put a requirement in this factory.
And now everyone is talking about spec-driven development.
This spec-driven development is mega.
We build our Pflichtenheft now.
And then we give that Pflichtenheft in this factory.
And my question is always, has that with the Pflichtenhefts with us people work?
Nein.
Does it work with the agent?
Nein.
Relative easy.
Also, this whole spec-driven development.
I'm not a fan of it, because that's just this, we have already all our requirements created, but it exists in the reality just not.
So, that's, I need to use an agile process.
That's also, this requirements are going to be an agile process.
And that's of course, you have just a little bit of it.
These factories are now relatively output oriented.
And now I have to look at how I can automatically automate things, so that I can answer the questions from my engineering field in an end time.
And that is what I find exciting.
That was actually also in the past.
And then there was a sprint, and then said, okay, in the sprint we're going to do these things.
And then I could put my questions in the sprint, and then I had to answer all the time, to answer these things.
And then came all the time, to answer these things.
And then came all the questions in a sprint, which were in a sprint, through the field and the product manager.
And now I'm going to take my sprint not five stories, but...
And by 800, but the questions are not less.
The questions are just now just quickly.
And then I have to think about how I build my organization up.
How do I change my role?
That I get this bottleneck, this next thing, I get rid of.
And that is of course quite obvious.
I mean, many are still showing me, that this is with the coding works and so.
But for us is it relatively...
Now you said that I can maybe even go on something.
I think what you can do now is to go on a lot of things.
You can actually know that you can do it in which bottlenecks you can do it.
And then you can think about it in the way you can do it with these bottlenecks.
Do I already train the product manager?
Do I have my orgs a little bit?
Do I have a different process?
Because I just learned a lot of people in the last year, because that was not so clear.
But now it's quite obvious that this will happen and we see it and we know it and there are patterns and there can be something already prepared.
What is your view?
My view is that the factory makes the same thing.
We see not seldom, we see it often when we build good products.
Because we have much more to do with the day.
Because exactly this bottleneck so zusammen snurrt.
die jetzt entstehenden Bottlenecks viel öfter zum Tragen kommen müssen und wir viel mehr Upstream und Downstream vom Coden uns unterhalten müssen und wir sehen, das finde ich so cool, für mich ist das immer der Vergleich zwischen vom Prospekt kaufen oder Probefahrt, ich sehe den Kram jetzt so schnell und kann sagen, mehr in die Richtung, das ist so wie lebende Moodboards eigentlich, weil du früher hast ein Moodboard und dann baust du ein Jahr.
Ja, ja, ja.
Dann hast du wieder ein Jahr, um gegenzusteuern, aber jetzt hast du eigentlich die ganze Zeit ein Live-Moodboard und sagst, mehr nach rechts oben, mehr nach rechts oben.
And then you go to the next step.
That's what I see as a interaction.
I think, I'm a very big fan of Mob Programming or Ensemble Programming.
That was you could do earlier.
And that's a bit of this idea of Synchronization.
We set all the people, Product Manager, Business Owner and so on, we all set them on one Tisch and they develop that together.
And there was always a bit of the problem.
That it was a relatively slow forward.
What you said, I saw something that had a little bit of a delay, because I had to write the code manually.
From the thing here was this model actually not bad.
I didn't need to do a planning meeting and nothing else.
I asked myself, what are the fachsing conditions?
Worauf we should be careful?
And then the staff told me.
And then the security manager said, what is with these five security things?
And then he had to code it together.
I think if I look at this, the work is actually not so different from the work.
The factory is like a kind of a couple of...
So einen an der Seite, der da irgendwie strampelt wie so ein Irrer, der das so anguckt, wie dieser Tisch so diskutiert und dann da einfach nur mit einer Affengeschwindigkeit dieses Zeug einfach, oh, jetzt haben sie das gerade gesagt.
Okay, okay, dann mache ich das jetzt auch noch, dann mache ich das jetzt auch noch so.
Und das funktioniert echt hervorragend, ehrlicherweise.
By the way, die Metapher ist ganz cool.
Ich mache jetzt meine Strategie und so Quatsch-Workshops.
Ich mache ganz viel.
Ich lasse wirklich die AI nebenbei laufen und die Arbeitsgruppen typen das rein oder machen das auf ihrem Whiteboard.
verarbeitest und sagen, zum einen hast du immer die Status Quo Dokumentation, bumm, die ist jetzt fertig.
Also wenn ich die haben will, ist sie da.
Zwischenstand, bumm, da.
And that's the thing that's with, it's like, yeah, cool and so, but that's not really the swear to play, what we've spoken about.
You're a little bit about the theme.
That's how you get a lot of moderators, but not only for me, but the group reads and says, oh, we have already done the theme.
That's absolutely absurd.
And that's exactly what a picture of a Nebendran.
In principle, that's always a mirror, in the man's work, the back reflect and says, oh, I think that's what?
That's just the Spannende.
Also, I think...
when you think about it, I see two different collaborations with AI.
One would be this one, where the AI works, where I'm either alone or where I'm in a team.
And then I'm in a team that's in the moment kind of hand-of-the-science, and so on for this question.
Or I'm just alone.
And the other one is, that I kind of I've worked with this group a prototype or something.
And now I want to stop things, where I don't need to need people anymore.
Where I say, make this one production-sweep.
And for that I need to do something like this factory.
And of course, it's definitely a vision.
I'm not one of these crazy visionaries.
You've heard of it.
I'm always a little bit like, what's happening right now?
And what's that mean?
There are people who say, I could build an AI, which the whole organization could be.
And then at all emails and all the emails and everywhere.
And then I have a huge, so a virtual organization.
So the factory could also be the whole organization.
I think that's a real brain, to be honest with you, to be able to trust.
And then, I believe that works.
in the organizations, which we have.
I think that there are people that have to try something out and ours works pretty well.
But we are just three people.
That's the experiments that I make.
That's what I do with people actually.
So im Rahmen von, keine Ahnung, ich arbeite so mit zehn Führungsleuten in einer Firma, klappt das super gut.
Den Kontext überschaust du immer gut und kannst auch sagen, oh, was sollten wir demnächst besprechen und du kriegst immer false positives, wie du wegklickst und so ein Kram.
But that's still, it's still trivial, but still still.
And so, because the analysis is so crazy and so formulated, not from the language, but that's so to the point, that so that signal from the noise was pulled out, like a person that would never do it from the point of view.
That's then more hard to get, what's on the rules, that's all just Scheinabsprachen, that's not really a decision.
That's nagel that's yeah, no end.
That's not always nice to get.
Here, I kille already your calendar, but I have the last question, the last question, the kind of question.
And that can be said.
We can write a article about it, which you can read soon.
Here comes the question.
Super cool.
You have the factory and we have the right job.
We decide how to take care of the companies, so that the decisions are not super träge.
And that's just a infrastructure.
And all of them can work directly, so that the machine doesn't wild and the right thing makes it.
Everything is cool.
Everything is together.
Now comes my example from the market.
The people, the people, the people, they have to buy stuff from us, they have to understand us and say, yes, they have to buy stuff from us, but nobody else.
As I worked at eBay, there was a problem.
We have told the people 10, 15 years ago, hey, that's best in the world, a thing to buy, are Auktion.
You know it or not, but Auktion is it.
And that was maybe only for Longtail-Geschäfte, so, egal, but we are going to be for auction.
And then we're going to show how Amazon is so right away and says, hey, auction, have you all?
We can just buy things.
And we're going to buy things.
We're going to buy things.
And we're going to buy things.
And we were going to eBay, we came to Christmas, mass of the packages, but all with Amazon-Verpackung and kaum more.
And we were going to buy eBay.
Under and because it's then on 24th, too under the Christmas tree.
And not the Auktion S on 27th or something.
So with other words, Amazon, eBay comes on the smarts and says, we could just buy things just like that.
How cool is that?
So, we're doing that and so.
The customer says but years long, why eBay buy things?
I don't understand.
I'm not there for Auktion.
And that's been five years to have been 30% of the amount of sales.
irgendwie über Direktkauf ging und neun Jahre bis 50-50 wurde.
Das heißt, die Marktgeschwindigkeit war absolut langsam.
So, jetzt die Gretchenfrage ist natürlich, hätte mehr Software schneller geholfen?
Und meine Hypothese ist, wahrscheinlich nicht.
Wie gehen wir damit um?
We have these cool beasts, which do everything we do.
We do everything we do.
But the market is so rich.
And I mean not only the customer understands that.
The problem behind is that the signal speed and the feedback loop from the market in the company is not hyper-hyper-chnell.
How do we go with that?
Yeah, also I would have no...
Also Code would that probably not lösen, so to say.
Schneller, more Code or so.
I find it on another...
On the other side, I think that's a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit And that has to do with the software.
You can say, I make marketing campaigns and branding and then...
...and then the society...
...it's all already, exactly.
But there's no software...
...but I don't have to ask myself, why I need to do a software factory?
Because at the end...
...what do I do with?
And I think the question is actually...
...we're still still asked.
These factories, if you look at them...
Also die wurde auch schon ohne die Factories nicht gut gestellt, weil Leute haben ja trotzdem auch irgendwie tausend Features in den Shop gebaut, die kein Mensch brauchte.
Und ausbauen ist ja sowieso schon mal gar nicht.
Also muss ja alles drin bleiben.
Aber das wird natürlich jetzt noch wahnsinnig beschleunigt.
Früher konnte man bei so einer Migration dann wenigstens noch sagen, oh, damit wir die Deadline einhalten, migrieren wir jetzt aber die Hälfte der Features nicht, weil die sind ja nicht so wichtig.
Deswegen lassen wir die mal weg.
And then is the shop on the end of the way, because the half of the features is out.
But now it's a bit like, can we do the other 100 also build?
And then it's better, because we have got 100 features in the shop.
So this question, when the agent actually hits?
So first the big question, when is the outcome achieved?
And when is the agent then actually hits, new features to produce?
But I think we have that in the short term.
And there is our first article, I think, that we say, I must actually in these Code-Reviews, in these trust-fragestellungen, I must actually actually more the loop now break.
That's a bit of the idea of this article.
Because this loop is just unending, it's always going to go on.
And we as people, we can say relatively naturally...
it's time to try to stop, so, it's enough invest and now wait until we go, until we go, we sleep in the morning and if we find that important, then we look at it again and that's not the agent.
They don't listen to it.
If someone says, oh, there are three new ones, let's go on and then he goes on.
But this is, I think this is, for me, it's also a very...
in philosophischen Fragen.
Wann ist das eigentlich genug oder wann ist das fertig?
Wann ist so eine Firma dann auch fertig?
Das ist lustig, weil wir wissen ja, in der Mathematik gibt es das Halteproblem, das nicht gelöst ist.
Und hier ist das eigentlich auch so.
Und letztendlich müssen wir, künstlich oder nicht, müssen wir den Agenten halt sagen, jetzt ist gut genug, weil die kein gut genug kennen.
Und das ist so die Mechanismen, die wir uns überlegen müssen.
Wie können wir So, dass wir den Automatismus nicht so sehr brechen, dass wir ihn genießen können und ausnutzen können, aber trotzdem sozusagen Stop-Kriterien definieren können, die aber halt wieder wahrscheinlich aus so einem versteckten Wissen, das nicht überall automatisch da ist, fast wie das Gegenstück zu einer Speck sozusagen sagen können.
not to run, if that's achieved.
And I mean not something stupid like just a simple Akzeptanz criteria, so the computer works.
Really on a qualitative level, that it's also for us people steward.
Because on the level of everything will not be quantified.
You have to have qualitatively be done.
Yes, and you have to have a little bit, I think this is more important than the other.
This is over the way.
I think it's amazing that we have been in the last two years of the last two years of we have to be sure that these agents can be quite long working.
We have come to the question, how do we hold them actually on?
Because these statistics are like, how long an autonomous agent does something make.
These benchmarks are already all switched, because it is just unending.
So long as Compute is there, it works there.
And also with a change of progress.
The code can be a little bit better.
I say it, there is more.
And I think what really is nice is, maybe almost to the end, if you're doing that with firms, is just...
genau das zu spüren, dass diese Entwicklung, die wir jetzt im Podcast in den letzten zwei Jahren gesehen haben, das passiert halt bei Leuten auch in Wochen zum Teil.
Also du fängst heute an, was zu machen und in vier Wochen sitzen die da und sagen, ich wusste nicht, dass ich das kann.
Das ist ja abgefahren.
Ja, und warum habe ich das je anders gemacht?
Und oh Gott, ich habe keinen Agent mehr.
Was soll ich jetzt tun?
Und auch dieses, wo kam der Zweifel her?
In dem Moment, wo ich es probiert habe, hat der ganze Kram ja geklappt.
Das ist ja abgefahren.
That's not a fantasy.
I can do that.
That's absurd.
Why do I?
What?
Yeah, yeah, yeah.
I know what you mean.
You see people in the face.
I make that all over calls.
They come in a call, two hours later, look what I did.
Oh, I'm going to go.
What?
Really?
Yeah, yeah, that was I.
So a example is now in my world, the not coding is.
There come people, the product manager, Die haben seit Jahren irgendwie verweste, elende Excel-Dateien.
Die kommen nicht mit denen zurecht.
Einmal im Leben wollte ich eine Übersicht haben.
Und dann gehen die hin und bauen sich so ein komisches Dashboard.
Und wenn noch komische Systeme angekommen.
So, und eine Woche später haben sie so ein riesenfettes, hyperkompliziertes Dashboard mit 38 Milliarden Filtern und so.
Und zeigen das ihrem Investor und so ein Kram.
Dann gucke ich da hin und sage...
Hey, there are 50 Python scripts.
What?
Really?
I said, yeah, but look at it.
Oh, no idea what that is.
Where is that here?
And I have a full-on-one trust in these things.
I have that 100 times against checked.
The numbers are correct.
So correct, as I had never had.
How cool is that?
And my colleague from BI could have that never delivered.
How cool is that?
My life is good.
I think so is really cool and beautiful.
Yeah, the Kreis is closing.
So we are doing this.
Why?
You asked me at the beginning.
I think the moments are exactly the ones I find and I find so exciting.
That's why the students or the colleagues, which we show them.
That's the most beautiful moments where you think, very good, we have achieved it.
Then it's actually already done.
Then it's done.
Then they're going to go away.
But it's still the same.
Yeah.
We work for the first time.
Yeah, yeah, yeah.
Yeah, we do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do.
We do Mega.
Benedikt, vielen Dank für die viele Zeit, die du genommen hast.
Ein super Rundumschlag um alles.
Wir hätten wahrscheinlich stundenlang weitermachen können, aber für heute mal genug.
Wie gesagt, auch deinen Kalender im Blick haben.
Vielen, vielen Dank.
Wirklich.
Ja, danke, dass ich hier sein durfte.
Hat super Spaß gemacht.
Mich sehr wohl gefühlt.
Vielen Dank.
Cool.
Dann bis bald.