Shipping agents, agentic workflow and what to build with Malte Ubl

[Music] And I think we really need to crack on. So our first I mean as an MC one of the things you want to be able to say is our next speaker needs no introduction. And it's true. Our next speaker needs no introduction. Malta Ubie. All right. Good morning. Uh thanks for for having me here. I'm the first speaker. I get to define the vibe of the conference. Um so should we tune it up or down? All right.

app up app. All right, let's go. Let's [ __ ] go. All right. Um, so I don't know. I don't know why I made this title actually building agents and not just talking about it because I'm very much just yapping. We're not building an agent right now. But the goal of my talk today is that you come out of this talk and you have and I hate this word methodology in your head for how to select the right uh project to pick for the first agent you

build. Um that is actually successful project. All right cool. Um to ground ourselves because the first talk of the day um and I think this will be helpful for other speakers as well. Who here has already shipped an agent in their in their career? All right. I I I call it 40%. Um my my thesis is that in like a year like everyone will raise their hand. Um

because if you actually figure out how to do it and and and which projects to build um these are very successful and very helpful. Um, but agents are still a new technology and whenever you have a new technology and this might have been I don't know mobile apps in 2010 or I don't know the web in 1994 you you define things right it's not that's a it's a measure of immaturity in a way that you still have to like talk about what something is um so let's talk about what agents are they are according to

anthropic I think this is probably the the most popular definition LMS autonomously using tools in a loop and like this definition is like is actually horrible because it doesn't tell you at all what it does, right? It's just like it's actually an implementation detail exposed as a definition. Um, but it's still useful if you actually want to build one. Like that that is really useful like this is what you're doing. Um, got this definition here from Google AI systems designed to perceive their environment, make decisions and take autonomous action to achieve userdefined

goals. All right, so this is actually useful, right? actually an actual definition that um so it's some something about like doing something autonomously that has an actual goal set to it. All right, I will contribute a definition myself. I'm going to go out of the way because you can't even see it. Um agents are a new kind of software is my my thesis that we always wanted to build but couldn't for economic reasons because it was just too difficult. Um to illustrate this even worse, look at the

worst ever uh diagram you see in your life. So this this blue circle represents all software that's nice to have that we would love to build, right? And then like we couldn't actually build all of it. There was some stuff we shouldn't never have built in the first place, but like um it was it was difficult to fill out the these circles and and agents essentially are just making the filling in the circle of software that would be awesome to have.

And and the the reason why they do that is because they can you can now like kind of you can just rely on the model to make decisions in a way where that that was just really really difficult to build in the olden days, right? because you would have all these if statements, all these special cases and now you can rely on the the emerging behavior of the model to fill these things out and like more concretely you know this new kind of software is really automation of workflows that were previously non-automatable right again