Welcome to another episode of The Edge. Like all the hype around large language models have died down a little bit. We talked about this before. But today, Jasper and I, we dig into agents.
What are agents? How do agent workflows work? And why are they the current hype? And why there is a future?
We explain to you how they’re used. We explain to you what we see, the trends we see in the market. It’s a very fun discussion. I love it.
I hope you do too. Hey, Lutz. Welcome back. Jasper, it has been a while.
I think we’ve been very, very busy for various reasons. But yeah, I wanted to kick off not asking about coffee because that’s obvious. Not asking about where you are because that’s obvious. But asking about like latest news that we heard.
So everybody suddenly is jumping on a new topic that is called agent. And it’s not secret agents. But it seems like we read about OpenAI releasing new agents and Google is planning something. Maybe Anthropic can now control and access my computer and do stuff.
So that’s the new thing, apparently. It feels a little bit like, you know, when kids play soccer, everybody is running behind one ball instead of sitting strategically. Yes, there was the wave where everybody said, we are going to have reasoning. And there will be even models that replace us humans.
Nah, far from it. What I like was there was this post on whatever this platform is called now, maybe one letter, maybe longer, but from Jan Lekun, who said, I told you so. So in parallel, apparently people are worried, or just saying it had to happen, that large language models are reaching their peak. So they’re not getting any better soon.
And then you read about Sam Altman saying, yeah, but the new thing is agents. So kind of look, is it marketing or is there something behind it? I guess that’s the question here. There is something behind it.
And you actually can see how heated that discussion actually is. Mark Benioff, the CEO from Salesforce, I mean, he went on stage making fun, of the Microsoft agents and saying, these are like Clippy 2.0 and they’re completely not helpful. And then he obviously introduced his own agents. You mean the Einstein bot that he’s trying to sell for years?
Okay. Yeah, well. And now point is, he’s a pro, right? He goes out and he does it in order to place his own agentic universe.
So agents are a thing. And why are agents a thing? Yeah. And I guess the interesting part, Yeah.
of the Microsoft agents is that they’re not just agents. The hard part here is maybe we should dissect what an agent really is. Because when we first came in contact with it, I guess our reaction, at least mine was, is that really new? Because it’s a combination of different things, right?
It’s not a new thing like a large language model. Let’s first define what agents are. I mean, an agent is not Matrix, a guy with sunglasses going around. Not James Bond.
Okay. It is like Matrix. It’s a piece of code. No.
But like the agent, essentially, we always said large language models will change the way our user interface acts. And that’s what we see now. Take very defined pieces where we allow an interaction with the code, with the computer in a word or like text-based setup or an image-based setup, making it easier to do one step and one step at a time. And it’s also…
A little bit close to what we called chaining of models. There was Langchain and others. Because I realized, okay, I can’t use one model to rule them all. There were a lot of discussions there, but I just combined different models.
And then that agent helps me which model to use when. This is actually very nicely said. So chaining of models sounds completely amazing, but it’s actually very simple. Chaining of models is what an average manager at any company does, when they have a young new employee.
They give not the task of, okay, improve our revenue. They break it down in step one, step two, step three. That’s called chaining in our data science. So it’s not like what we do at our VC.
We just say to our younger people joining, yeah, you find great companies. Go and find them, right? It’s also not how we work with our portfolio companies, right? When we get a portfolio companies, you’re not saying, oh, come back.
When you did the IPO. No, you actually break it down. And for models, we have to do the same thing. For my online course for Equinel, I actually had this funny revelation that I offered every student in model and co-pilot, an agent, from the start.
And people could not effectively use it. So I actually had to break it down. Think about the problem, break down the problem, ask step-by-step chaining. Okay.
Everybody was looking at RPA and said, oh, that’s simple scripting. So robotic process automation. So I basically take outsourced tasks that let’s say low skilled labor would do. And then you realize, oh, by the way, they’re just saying A or B.
There is a rule engine behind it. I can script this. It’s software, UiPath, Automation Anywhere, pretty successful companies. But then you realize, well, it’s for very, very low skilled work because it isn’t a rule engine.
It is zero or one decision or you can’t have, let’s say, uncertainty. You can’t have unstructured data where then has to be some decision, which meant the promise of automation was not very limited. I mean, UiPath is a huge company, Automation Anywhere. There are many other players, but it has some kind of limit.
So is it now because we have AI, artificial intelligence, can we now do really smart tasks? Is that the idea of agents? You are old enough to know Clippy. From Microsoft, the paperclip, which kind of was waving one finger and kind of saying, hey, you wrote deer.
Are you writing a letter? Right. This is kind of, and Mark Benioff made fun of this because it was limited. It figured out you most likely writing a letter.
Let me help you know in the process of writing a letter. Now, Microsoft ain’t stupid, right? So they definitely tested it. And there are people.
Who need help in writing a good letter and could have used this tool. There are many others who don’t. And we will see now. And this is an agent that like, don’t get me wrong.
It’s a rule-based agent. It was terrible for the ones who knew how to write a letter. So we now have autocomplete when we are writing a letter. And that’s way easier because we can guide it more.
So it’s still true what we said. Very early on in our podcast series that we said it’s a UX problem. So here we will have now agents that are easier to steer because it’s not Clippy asking anymore, but will that will run into the same problems as Clippy that sometimes they are too simplified or sometimes they are too complex. And I think that the wrapper or the high level summary of it is yes, it’s a little bit like a rule engine.
It makes sense because we. We want to solve a certain task, maybe a mundane, repetitive human task. Imagine something you do every day and it’s pretty similar, but now it’s not just send the email or don’t send the email, but it’s sent the email to a certain person based on the header, based on the topic. Maybe send the email in a certain style and I can detect the style.
I can extract it from the email. I have my large language model behind it, but I can be a bit more complex. But I think the interesting question is. What is the limit?
There again, because we just learned large language models are limited. And how can you make sure these agents are not going rogue? Sometimes agents go rogue. I’m not sure that the rogueness is a problem because hallucination as a technical problem is actually pretty well solved.
But you don’t want an agent to send random emails to random people in your in your. No, because you have to break it down. Yeah, OK. To break it down.
What is the email you want to send? But whenever we both sit together and we get it. Pitch onto our desk and then we discuss it. It’s actually funny.
I think there are three groups, right? There are the groups that kind of just wrap an interface around an LLM and says it’s going to be amazing. And that ain’t be the case. No, it’s it really shows.
I think we discussed it a year ago already. It really shows it’s not just the wrapper, but if I have better search like better query, thanks to information retrieval, everybody can do it at some point in time because it’s an API. And as you said, it’s solved. Yeah.
Yeah. Exactly. Now, then there are the people who saying, look, I worked in the industry and I know that I need to do step, in order to get there. So these are the ones who have a workflow in front of them.
They know that the workflow can be simplified by using an agent. And then they’re looking for agent. And here agents is not necessarily a large language model. Agent could be a piece of deterministic software sent mail.
Or send acknowledgement if somebody bought something, for example, an e-commerce. Or it could be a prediction, kind of like what is the most likely thing you want to do next? Where do you want to go? As well as have a smart interface to actually steered and summarize or extract information from words.
Right. And that’s a second bucket. There is a third bucket. Because those companies.
Are very often awesome. But they follow a very structured approach. Right. And you and I, we always wonder how will the world look like if you have agents.
Right. Where’s the journey going? I think also I’m trying to make the connection with RPA. Who was the biggest beneficiary of RPA?
It was the consultants. Because they would go into these large companies and say, oh, that’s your process. Oh, that part of the process is repetitive. Hey, we can automate that.
With scripts. We can run the scripts. We can set it up. But that costs you quite a lot.
I don’t know. It was maybe 20x, sometimes 10x of the total of what a new iPath or other companies would get out of it. And now to your point. Okay.
It’s cool. There are a lot of horizontal companies that want to help you building these agents now. Similar to what a UiPath did. I just read Index back.
The former CTO of Stripe. Amazing person. A lot of experience, obviously. So they give you the tool set.
But I think it’s a bit. Similar to other companies we’ve seen in our deal flow. If you don’t know what to use it for. Again, we are producing tools here.
Then I have a problem. And this is deep knowledge, right? I have to know which processes can actually use the agents. Don’t break.
Don’t destroy anything. Don’t go rogue. I’m still sticking with my point here. Or would you disagree and say, no, don’t worry.
The agent can discover that by themselves. No. You have to give the agent very clear boundaries. And UiPath is very nice.
Nice example here. But I have to chuckle if I see companies like stuff like SAP, R3 systems. And they’re trying to make for each step along the workflow. They’re putting in a better interface.
They’re putting in a piece of predictive software. I think that’s part of the second bucket, right? You don’t give the agent just a direction. You actually go step by step.
However, we should acknowledge that this workflow is changing, right? The. Way how we will work on problems is about to change. So companies who I normally get very excited about are the ones who acknowledge that at the moment the workflow is here.
But we do not want to be here because we will bring in an agent and that will change our workflow. Cool. Let’s divide it maybe in two parts. One is because people that are listening maybe are new to the topic, but understanding a little bit around.
What are the capabilities? And then I would say, let’s go into some recommendations, what to look at. And I think you also have some practical insights. Also at Sherry, we invested in some companies.
Most of them are still stealth, but we can at least tackle some ideas or yeah, let’s go for that. Cool. So there are a couple of, yeah, you could call it’s not a framework, but a couple of patterns that these angels are doing. And I think it’s interesting to at least hear it on a very high level for everyone.
If they are thinking about automation augmentation with agents, what they should consider. And the first thing I read was reflection, which is interesting because you also saw a one thinking. So there’s kind of this step of reflection. What is that?
What does it mean? What is reflection? What is an agent reflecting? They’re not intelligent, are they?
They’re not intelligent. And all of this talk about agents and an ability to plan is essentially. Chain of thought, meaning I have a goal to get to. And what is the best way to get to a goal?
It sounds like reinforcement learning, is it? No, not quite. It becomes reinforcement learning once you then have. Did you reach the goal?
Yes or no. And you go back. OpenAI, the latest version of OpenAI 04 is essentially focusing on this. The more complex the question, the longer the answer will take.
Because. The first question, which internally is asked, okay, you have a user request. How do you best break this down? And then probably OpenAI has said, break it down in five approaches.
And then it has another chat GPT call saying which one of those five approaches seems best and then execute on this one. And then is the answer realistic? And then answer. So we have a loop of many, many.
Different calls. But you already see that they’re in between. There are certain structured pieces. Think about how to approach it.
Is that approach a good one? Does this make sense? Can you double check? All of those pieces are elements of an agentic workflow.
And I think there also you see already what is different to just prompting a large language model. Because what happens here is and this is public. Cherry invested in Octomind, a company that does front end. Front end user testing.
They call it self healing mechanism. Because what the agent does and RPA cannot do that because it’s a script. It’s a rule engine. It breaks or it works.
It kind of checks and corrects what the tests are doing and adjust accordingly. So you look at the data. You say, oh, that goes in wrong. Maybe I saw something wrong.
Maybe I look somewhere else. I test out different things. So I still come to my result. But I try different ways without needing a clear rule engine.
I can do this kind of self correction. That sounds pretty handy. As you know, I’m building agents for e-commerce, right? And the approach there is we have search and we have consultative approach.
And so the consultative approach is essentially bring the information the customer is looking for pre sales. But do this under the setup to actually push the customer further down the sales route. And this is what I’m doing. I’m not going to give you like the jokes I made about Amazon.
If you go to Amazon and you are asking very open like one question, Amazon gives you like on their generative AI tool. It’s not live in Europe. It’s called Roof Force. It’s here in the US.
Amazon gives you a very broad answer. But hello, I came to Amazon. I probably want to buy something. So formulate a sales approach.
Figure out what most likely I want to buy and educate me along that path. That would be an agentic workflow, which is not just a wrapper. Very good. The next one is that the agent realizes there’s something missing.
So I’m doing this testing. I realize, okay, I can’t really self-correct. I have certain goals I want to reach. It doesn’t work.
So I have to expand my knowledge. And I like this part because also we had a lot of discussions about maybe OpenAI will release their own browser. Maybe agents need better access to websites. So actually we need a new browser technology.
So it’s kind of this knowledge expansion by independent search of the agent. Where can I find the information? Where can I be up to date? Well, I mean, they’re released search, right?
And OpenAI makes most of their money with subscriptions, not with API calls, right? So they’re a consumer company. And you recall that I said, like, if at the time there were a research project, I wouldn’t have invested because they are missing the consumer angle. Now they have the consumer angle.
They’re very… Very promising. And they are trying to be there. Now, the interesting part, what we just discussed, like, think about the problem, call in a function, which not necessarily needs to be a large language model.
We had this already with 3.5. At 3.5, at 3, if you would have asked OpenAI 3, then about something, okay, as I was six years old, my sister was three years old. I’m now 60 years old. My sister is.
How to completion, 30 years old, which is obviously wrong. Now, 3.5, OpenAI changed and said, take the question from the user. Is it a math question? If so, figure out what are the elements and hand it over to a calculator.
Give me back the results and let me reformulate this. Meaning, I take the universe of all the questions and cut it down into a structured entity and work on it. And I think that’s even leading to the third question. So, I think that’s the third component.
So, we had kind of knowledge expansion. It’s not in my vector database. I can’t, I don’t find it. My self-correction doesn’t work.
My reflection says something is wrong. But I can now coordinate. So, to a limited extent, I assume, to your point, I can coordinate tasks and say, oh, wait, that is the other model or the other agent that can solve this kind of task. I identify the task and then I coordinate.
Or is that different now? No, I find tasks and coordinate those tasks. And now here comes the, I think, super important message for all of the founders out there. Number one, if the task becomes as generic as figure out whether a calculator is needed and use a calculator, you will not have a business model.
Because essentially, the generic models, this is what, like, this is the long tail. And we have seen this race for the long tail already with Google, right? There was a time where Google was good in search, but not. Good in specific search.
And you got all those search engines saying, well, Google is good, but we are extremely good at this kind of search. And over time, Google replaced them all because it’s a long tail. So, if your next step, that’s your long tail, is something which is replaceable, then this is a problem. The second one, which we should talk about, is how will the user behavior change?
For example, 4G. For my e-commerce example, I actually assume that most people do not want to have a lengthy chat before they buy something. We all hate the interaction, like, with Google Home. Why should we want to shop suddenly with somebody who talks us to death?
Therefore, we’re building it into search and making very tiny little nudges to move people along. Right? So, I’m thinking about where the user. is today and extrapolating from there to where the user is tomorrow.
Yeah. And I guess also giving another example, we hear a lot or we see also a lot of companies addressing the searching documents, reading through documents. So, a lot of my task is actually having a good folder structure, like having some indexes, etc., etc. And now I can just ask about certain things.
It shows me the documents. It shows me where it is. And then it’s more me coordinating in a fast way what’s the content. And rather than, you know, having a structure and a structured thinking there, it’s more following the question and searching for an answer.
I think the last part of what these agents are doing, and this is maybe getting a little bit, yeah, very artificial intelligence, at least what I read. So, we have parallel processing. Obviously, a lot of these agents can work at the same time, which human beings cannot. Or even in RPA, you have sometimes to do these stepwise processes.
They can dynamically adapt. But there was this direct… Directed as cyclic. I think directed as cyclic.
No, I’m German. Sorry. Graphs like DAX. So, we don’t have RACs.
We have now DAX. And that sounds like a bit like science fiction, that these agents start really orchestrating, paralyzing, complex nonlinear workflows. Or is it marketing again? In a limited space, it should work.
Because in a limited space, you kind of can break it down. So, like talking about the long tail, you break different parts of the long tail. And Google is doing this. If you search, it’s not one search index.
The question is, how deep do you need to go into the long tail? And it’s actually funny. Like, you know Glean, an enterprise search company. They are using large language models.
They are awesome. They grow like crazy. Crazy. But they get funding for $4.5 billion.
Yeah, it’s a pretty high multiple. Yeah. I mean, $600 million in revenue to $4.5 billion. It tells you how hyped it is.
They need to understand the long tail and search. Now, the question is, will OpenAI, u.com, Perplexity, all of those, or Microsoft for that matter, will they be able to understand the long tail? And that’s a super fascinating discussion. I haven’t seen anything amazing coming out of Microsoft lately.
But I’m pretty sure they should work on that. Cool. All right. So.
So we now understood. It feels like a lot of ants running around having certain tasks. They can talk to each other. There is a super ant coordinating all of these.
They’re not super smart, but they are kind of smart and specialized. But they can all work in parallel. They can all coordinate. And they have this function or this kind of goal they want to solve.
Actually, may I? Like, they are not super smart. It’s a huge statement. For all the people who believe.
They’re like, oh, we need now ChatGPT 7, 8, 15.7, right? And there will be this one general model that is replacing us all. And we made this show so much fun of people who talk about AGI, right? Whenever somebody puts a document in front of me where the word AGI, we had analysts in the jury actually writing documents.
It’s like, AGI, I don’t think looks like this term, right? Because we made fun of this. Now, those general models won’t exist. They won’t come, at least not for now.
We will replace them with smaller, faster, simpler models. Yeah, well, Sam Altman might disagree. Maybe we can invite him to the podcast one day. He definitely will come.
He definitely will come. It will be a good discussion. Okay, so, but now, dear founders, dear builders and companies. Hold on, Sam Altman.
Sam Altman. He will not disagree. I’m pretty sure. And Sam.
If you hear this and you disagree, drop me a note. I owe you a beer. But he definitely won’t disagree. Because what he is doing at the moment is he is orchestrating the long tail of consumer search.
And he is doing this with smaller models. If he would do this with a big model, it would take way too much time. Yeah, plus we, I mean, also we discussed it. There are so many things that the large language models are already solving.
And now we get into it. Yeah. So much work we can do automating that, augmenting human beings. I mean, just the Glean example shows enterprise search.
There’s so much potential. Just people not searching anymore, but just doing the work and thinking. And I think that’s something to focus on and build great companies. So what could be those great companies?
If we think about the first RPA wave, obviously that was outsourced BPO. So very clearly defined labor, very low skilled. Now. We come at a higher level so we can make more decisions.
We can search as we learn. We can summarize, we can analyze. So there’s a lot of stuff that you can argue are still lower skilled, but maybe an analyst at some companies. And I’m not saying if it’s consultancy, investment banking, or VC, but yeah, it’s some of the work for sure.
But is it that we now go for all the industries where they’re not digitized, they’ve never seen a computer? Yeah. Or would you say, no, it has to be a certain level of sophistication because still you need a certain quality of data. We discussed it in previous podcast.
Or are we really opening up, let’s say the system of records and old industries and can finally work with the data? We had the SAP example before. Yeah, I think the data is essence, right? So we are not opening up the data, but like companies who that have data, the SAPs of this world.
Yeah. If the user flow is changing, right? And that’s the reason why we’re pushing so far. Now, the other thing which will be a winner, and I don’t think I said it this way, but like you triggered it early on, it’s obviously consultants, right?
Because the user flow will be changing. So good time for yet another wave, McKinsey and BCG, if you’re listening to this. Accentures, congrats. Okay, so let’s do the final here.
If I want to build something with an AI agent and you have done it yourself, we also did a little bit, we published something with our harvester, but that was obviously not as sophisticated. If I start, so I say, hey, I really like this topic. It’s reflecting, it’s automating, it’s unstructured data. This is pretty cool.
And I have a problem I want to solve. What are kind of questions you ask yourself at the beginning just to get it started? Maybe we can begin with that. It comes down to the value prop, right?
Won’t be a surprise for you to hear if I say this. Is there a workflow that needs solving? If you look at healthcare, like how do you do pre-authorization? It’s a typical workflow.
And there are companies who are trying to do this now, right? You have pre-auth, you know that you need to distill information and you could now get the information in a structured format, send it over to the insurance. The insurance will use an LLM to see, does it map to my contract and formulate the argument. And then you can go back and send it back again.
That would be, there is a workflow. That workflow requires human input and a structured step. If you have something like this, you’re awesome. And you have a place.
Now, will that workflow change? Yes, sure. So depending on how are you positioned for that. And I think you are addressing a workflow that is pretty known to everyone.
So I’m on a classical e-commerce website. And the way I would usually look for something I could buy is through filters. So I click buttons, I click colors, I see pictures and I realize, oh, okay, that wasn’t what I was looking for. So I changed the filter.
I realized they don’t have the filter I want. That’s at least how I experience sometimes Amazon and other things. So I’m kind of hacking the rule engine. That’s how it feels.
Now, this is actually funny because we have those two elements. User flow and data or like model as such. E-commerce engines very often rely on the text data of the products. If you are Walmart or Footlooker or one of some of them, you get your products from your producers and they kind of staff all the keywords and they do black hat SEO essentially for your search engine.
They want to be shown at every search. Now your search engine is text based. And that’s kind of shitty. But you can do that.
So all of us, we get used, like we know that search is not ideal. We search for a product, okay, but we will not search for something complex. And that’s a problem. So we are addressing, like my own company addresses it this way, that we clean up the data.
But as well, we make it then easy searchable. So like we’re working for jewelry store, for example. If you look for a ring, you find a ring and it gives you filters. But those filters are, do you want a bracelet?
No, that’s not the thing. I wrote ring, right? Now the best next filter would be golden ring, silver ring, right? But if you look for a style Gothic, then it will not work.
And for that, we use now an LLM to saying, okay, we have a computer idea and an AI idea of what a Gothic ring should look like. Can I search in all of the images for rings that have a Gothic style? And that is what I bring in. That is what I bring out then.
So we improving search essentially by doing two things. We, A, clean up the wrong data, data wins, as well as we using the stepwise approach, clean up data, look for something. If that’s not there, use an embedded model to figure it out in the images to bring it up. It’s an agentic workflow.
And if I even expand that, so I’m still on my, you know, orchestration, parallelization, all these crazy ants running around. Would that be, I do a general search. I say, I want to have that and please buy it for me. So then the agent would understand is Jasper.
That’s what he did historically. So that could be one agent. Then the other agent would go through the web. This information retrieval we discussed, getting smarter, latest information, find the stuff.
Then maybe comparing prices is another agent doing calculations. And then one agent that’s probably more the RPA level, goes to the shop and orders with my credit card automatically. It feels coming back to the interface discussion. For me, it feels like, hello, I would love to buy the certain thing.
If the price doesn’t exceed X, just do it. There’s a lot happening in the background. Is that the future or is that possible? That’s going to be the future.
That’s by the way, the ball game for perplexity. They’re recently launched something in this market. We are far away from it. And that is because we humans don’t get to trust that workflow.
And probably for a good reason. We’ve seen, by the way, we’ve seen many travel agent AI companies in the last, I don’t know, 10 years. And you just imagine that thing putting you on the wrong flight. Yeah.
I mean, like, and the pitches is like, wouldn’t it be awesome if an AI organizes your honeymoon? No, no. I really like the discussion. If it’s great, yes, but maybe not.
About my honeymoon. There is an approach how to find your own. This is an experience in itself. But let me break this structure rules down in another example.
Let’s go back to the very simple search. You go to a jewelry shop and you look for a ring. Okay. Then you find ring.
That’s a facet. You don’t need AI. You look Gothic ring. Now you need a large language model to encode what you need.
You look for 585 ring, meaning, okay, that’s a type of gold. Oh, ha. Now you do not want to look via an LLM because it’s very specific. So it’s Gothic, but it’s encoded in a number and not in the word Gothic.
Right? So you need to build out a model which understands where you are. And that model includes later on what I have in my head. So you need to build out a model that understands where you are.
After you’ve done all this, you can choose where you want to be, where you want to be, a recommendation to create the next best search and to deliver on the promise of using AI to help you sell. So maybe also to digest it here, it’s not just, let’s getting attacked by a fly, it’s not just that I automate workflows, but I even increase the precision of what AI can do by chaining, but it’s not chaining, it’s even putting together different models, even a bit of rule engine. And I think we also discussed for a couple of companies, there is even a hybrid approach of saying, hey, there’s a rule engine, and there is also AI involved, large language models, whatnot. You would actually choose, and to your point, based on the output you’re expecting, the best approach in a future world, or maybe already today, the agent can choose the best approach.
But sometimes you probably also need some human involvement here. Yeah, so I would be very careful of saying the agent should choose the best approach, because that’s essentially saying, we just need to wait for a bigger model. I think there is a huge market for startups where the startup founder knows the stepwise approach. And the startup, like you call the agents, and the startup founder says, I know the steps.
All what I need is I need a guy like Lutz. Who programs those ends and put them in a chain. And now, boom, I have my solution. And I think that’s way more realistic than saying, oh, I want somebody to plan my honeymoon for me.
By the way, funny, I talked to a large retail chain, and they said, yes, we’re building something by ourselves. We want people to come on and saying, I want my Thanksgiving dinner planned. And one of my guests has a nut allergy. I mean, dude, no, that’s not going to happen.
Well, you could do it, but let’s see what happens, right? Yeah, no, but also this is like, you know, as I started my career, we all thought about that our fridges will talk to us and fill it up. I mean, it has not happened. Why?
Because we are not. But I think the point you made is very, very important. If you go for high precision, yes, with an agentic approach, you get closer to that. Yeah.
where I initially, I came from healthcare. So my initial idea was, can I create a nurse for the pocket, right? Somebody who motivates me to make sports or to eat better or to sleep more, all the good things you need to do in order to be healthier. And I realized very, very quickly, it’s easier to lame or the stakes are too high, like in one side of the other.
So it becomes extremely difficult to follow this. So I scaled down the complexity and now I’m selling stuff, like I’m selling something as beautiful as my coffee machine or jewelry or something like that, because I still need to have an aim. I still have an agentic workflow, but the stakes are not as high. Precision recall.
All right, Lutz, thank you so much. So today we spoke about what agents are, not just secret agents doing work in the background, but actually solving new problems that could have not been solved by, let’s say, technology before, or at least very complicated by rule engines like RPA. So hopefully we see much more automation coming out of that. We also spoke a little bit about the capabilities of these agents.
I compared them to ants. I hope the example was okay, but I just imagine all these agents running around, but not being too smart, but at least knowing what their individual task is, how to work together. I think that’s also very important. That’s what agents are.
And Lutz gave some nice examples of how to approach the problem, what to think about, and what you can actually do there in practice, and what does not work. So I really like the conversation. Thank you so much and happy Thanksgiving. Thank you.
I think this is a good start. Yes. Awesome.