Docker Agent: Simplifying AI Agent Creation

Speaker 0: Okay, everybody. So welcome. Abdulai from AI Teamers, guys. We are a community of data scientists, AI engineers, and software engineers with a focus on AI, and we organize a lot of events in order to share knowledge. So today, we will have a few speakers who came especially for DevOps. Speaker 0: They will share with us what they've been building and AI and what's the enabling you to build with their technology. So we will have Sebastien, from. He will be the last 1 to speak. And just before him, we will have Philippe, Georgie, David, and Guillaume. We will share us what they are building at Docker and, their brand new technologies. Speaker 0: But just before letting them introduce themselves, I'd like Keith Gordon, our host tonight, to tell us, like, in a minute who they are, what they do. Speaker 1: Yeah. Thank you, Abulai. So, yeah, welcome at the Q broadband, privacy of the quarter. So we are a secret security platform. We are basically a developer, security engineer, even working with infrastructure to find secrets. Speaker 1: So secrets is everywhere. When we speak about secrets, we Runner. about API key, gen generality, token, anything that can let you access any, yeah, any resources, any infrastructure. Especially what's on my T shirt. Attackers on breakings and login. Speaker 1: So we help you identify all these secrets, either via VCS, Slack, anything like that. Speaker 2: You can use the product for free. Speaker 1: We have a very generous offer, so please go ahead and use it. And if you want more sauces, anything, just, let us know and, yeah, enjoy the night. It was a pleasure to welcome our communities, here, as developer. Speaker 2: Yeah. And, and, if you are looking for restroom, they're on the right side. Yeah. And enjoy the night. Thank you. Speaker 0: Thank you. But, Jeremy, Speaker 3: that Speaker 0: was too easy. Are you recruiting? Because I'm sure we Speaker 1: got a lot of talent here. Yeah. Good call. Yeah. Yeah. Speaker 1: We we do, recruiting right now. We have a lot of opportunities. Please go on the the big content page. We have the carriers, and we are looking for AI engineers, software engineers. Yeah. Speaker 1: Please have a look. We just raised the our series c. Speaker 2: So, yeah, we are scaling up. Rich. I appreciate it. Speaker 0: Great. So this is the first important QR code. Please scan it if you haven't been scanned when you came in because our AI agents try to follow your attendance. And based upon that, when you don't have when you don't have enough room, we prioritize people who have been to our past events. If you're looking for jobs, Guardian or not Deep Guardian related, you can scan this 1. Speaker 0: And if you are recruiting as well, you can scan this QR code and let us know who you're looking for. And at the end of the event, to let our speakers know how well they did or how bad they did. Please send this 1 so that you can give a feedback, but you'll receive an email as well with the link to that survey because they are here for DevOps. So this is like their kickoff, and we're trying to help them, like, getting better and better and getting ready for tomorrow. So tonight, after Geek Guardian, our Architect sponsor is Docker. Speaker 0: And for those who don't know who Docker is and who are in this room, who are not supposed to be here, please, I'd like to introduce who you guys are. Speaker 4: Hi, everybody. I'm Joflyp. I'm a software engineer working at Docker. So Docker is sponsoring this, this meetup, because we are not only doing containers, but we are also trying to address the AI market and trying to secure your AI agents. In the room, you will find a lot of engineers. Speaker 4: Feel free to ping us, ask us questions. You have engineers working on reliable stuff. You have AI engineers. Don't hesitate. Most of us are working on the tools we are going to Speaker 5: present you tonight so we Speaker 4: can give you details, etcetera. Speaker 5: And we are hiring. Speaker 4: Exactly. We are firing. Yes. Speaker 3: That's the Speaker 1: question I was about to ask. Speaker 0: What I would suggest to you guys for those who haven't, seen it, you have a networking tab just right here. If you click on it, you will see, the profiles of all of the attendees, first and foremost. And based upon the information you gave on your profile, if you click on suggested, if you're looking for a job, if you're looking for a mentor, if you are, trying to hire people on this tab, you will see in the event who you should talk to. It's going to be personalized. So, Philippe, technical solutions architect. Speaker 0: I guess we are supposed to speak at some point. You can help with technical architecture. Cool. And you try to promote Docker wherever possible. So, basically, you're not investing, but take advantage of our platform in order to meet people and, we can let some really do that thing as well. Speaker 0: So that being said, I think we are ready to start. And Speaker 6: if you Speaker 1: would yes. Thank you. Welcome. Speaker 7: Hello? That's loud. Hello, everyone. Can we switch the screen to Speaker 2: Hold on. Are Speaker 0: you guys in or you prefer turning the lights off? Maybe you can turn the light Speaker 7: All we need to do is spend 10 minutes on Speaker 8: each slide. Speaker 0: So you have 10 minutes. Speaker 7: 1 slide. Okay. Cool. Speaker 1: So I just know. Speaker 7: Hello. So yes. So Georgi and I, we're working for Docker, and we've got a couple of slides for you and mostly demos. We're gonna talk about Docker Agentic. Docker agent is Agentic framework, but it's also what we call the the Swiss army knife of agenda framework. Speaker 7: So there's plenty of things you can do with it. We're gonna show you only 1 thing today. And it is the kind of tools that we are working on at Docker to help you really change the way you develop, really change the way you work with AI, and we use it every day ourselves. Speaker 5: Jorge? Yeah. So this is my name. You can call me George or George if you're in French. Yeah. Speaker 5: I'm a digital engineer or yes. By the call now, does anyone maybe call these days? Yeah. That's me. Speaker 7: Yeah. And so both of us, we started to work on AI about almost 2 years ago. Docker and CP, like, the first project was Godel. Godel is the chat chat assistant you can find in Docker Desktop. Docker Desktop is our main product. Speaker 7: You can install it on Mac, Windows, Linux. And there's a challenge. I always started really slow with GALA, and then we tried to grow it. And back then, it was funny because it was a different world. Right? Speaker 7: It was 2 years ago. 20 years ago, we started with llama 3, t p t 4 0. We had no MCP, no obvious, no nothing. It was the beginning of live coding, but mostly those people were just considered as crazy. And now what do we have, Jorge? Speaker 5: Well, I guess, yeah, now I mean, we have open Staff 0.6 or 0.7 since last week, and I would urge you to try also GPT 5.4 because it's something else. MCP was all the rage for, like, 6 months, and then now it's even going out. Speaker 7: It's you shouldn't be using it. Speaker 5: I guess everyone says you should use now just a shell. And we even have, like, Agentic, orchestrators. You're not anymore just, like, using the AI to auto computer codes. You're using AI to for them to work for you. And, yeah, we even have code bases that are 100% coded by Agentic, redo by agents, bigger. Speaker 5: Yeah. We need bigger. Speaker 7: The question is for you all. Who's using AI to code or do they have okay. A lot of people. Speaker 0: The rest can leave the room. Speaker 3: Like, who Speaker 7: is even coding those days? But, like, way fewer and for the first question. And also, who's using agents for something else than coding? Cool. So we're gonna show you something about Docker agents. Speaker 7: You can reuse it for many use cases. And it's really good for coding, but we're not gonna talk about that. So, yeah, Docker agents, a bit more about Docker agents. It's really a Swiss army knife. So Swiss because you can choose from any model, any tool. Speaker 7: You can really connect to any kind of Model. The list is even longer than that. So it gives you a lot of flexibility, a lot of control. You can use remote models, local models with model docker model runner, but also you can use all m I q models. It's rich. Speaker 7: We use Docker Agentic to call Docker agents itself every day. A lot of us in the team, we use Docker Agentic to call. And it's there's a lot lots of features. We are gonna show you all of them. And it's what's quite nice also is that the first use case to get started with Docker Agentic is just being able to write some YAML file. Speaker 7: You can use it in more advanced use cases with the Go SDK, for example, but we're gonna show you what you can do with YAML. And, yes, we use it in production. Now both on our chat assistant is based on on the current. So you wanna show us the first demo? Yeah. Speaker 7: The the first yeah. Speaker 6: This is my my the Speaker 5: the demo that I that I love the most. Speaker 7: Yeah. So you can you can have very useful agents with the agent, but, also, you can have the least useful agent, and 1 of them is the pirate. Yeah. Speaker 5: This is the an agent that's. So this is, Docker Agentic have, like, Speaker 7: the 3. I guess if I make it smaller just for you to see because we have Speaker 5: I mean, there's much more yeah. I have, like, a sidebar and everything, but if I make it bigger, it goes away. So you can say Speaker 6: oh, are you Speaker 5: and, normally, the Wi Fi is okay. Yeah. It's telling me how how are they, matey? So we have an agent that talks, like, about it. Speaker 7: Yeah. Yeah. So it's very useful. Yeah. The nice thing about it the good thing about it is that you can It's great. Speaker 7: It's only you need to be able to write this to have this Agentic. And you've got a full tool, a full agent. You can talk to it through an echo SDK. It's really powerful. It has no tools, just a good prompt, funny prompt. Speaker 7: Yeah. This I mean, the 2 we qualify for, like, your only the model and destruction. Speaker 5: Like, the weaponized is just, like, for the tool to show your message. But, like, in 2 lines of code, you have an agent that does something useful, which is, like, make you, smile. Speaker 7: Yeah. We're gonna try to do a bit better than that. So, yes, specialized agents. So you work anywhere in a company, big, small. You might have APIs, knowledge bases, things that you want to use through your agent. Speaker 7: You want to give access to those tools to your agents because you need to solve complex enterprise workflows and all that Staff. And we're gonna build such Agentic, today in front of you. The only thing that we're not gonna do is we're not going to use a enterprise API. We're gonna use a Pokemon API. But when I say Pokemon, think boring enterprise. Speaker 7: Pokemon, boring enterprise. But yeah. Maybe you've got lots of APIs, existing web APIs that you might want to give access to your agents. So we're gonna do exactly that. We're gonna take an open API that can be used to query your knowledge base about pokemons, and we're gonna give that to our Agentic. Speaker 5: But, again, imagine enterprise and you have, like, your APIs defined in an open API. YAML. Speaker 7: So, yeah, to give access to this API to our agents, we're gonna it's gonna take only 2 lines of YAML. This is a this is the more complete example. So we're gonna give instructions to the agents. We're gonna give it a model. We're gonna we're gonna choose a fast 1. Speaker 7: Quite expensive, but quite fast. And we're gonna give it what we call a toolset. So the toolset, in fact, under the hood, what it's gonna do is gonna grab the open API spec. It's gonna say, oh, I've got now 97 endpoints I can use. Each of those endpoints will be 1 tool for me, the agent. Speaker 7: So we're giving basically 97 tools to our agents. Speaker 5: In 2 lines. Yeah. Can you yeah. Let's run it. So let's open it. Speaker 5: So, yeah, we have the welcome message. So, yeah, we don't see, but, like, technical tools, we can actually list all the tools that we found in Speaker 7: the in the open API definition that we have. Speaker 5: So yeah. So I have no idea what are Pokemon. So yeah. Yeah. Speaker 7: You can list the Pokemons. You can list the reg regions. You can get the split the spec of the Pokemon. You can get the egg group. I Speaker 5: don't know. Speaker 7: You can get better. So, anyway, you can ask Solutions to your Agentic now. I'm gonna ask you to French. It's it's basically asking the statistics of the the starter pack of, Speaker 5: Pokemon. There's 3 Pokemons. Yeah. Just like yeah. Speaker 8: And so Speaker 5: it has the tools, so you can call them. So you can see it did 3 2 calls here. It called for, 3 I don't know about those. I don't know what it is, but, like, it calls 3 3 remote tools, and then it got the answer and gave, like, the customer. It's it's hard to Speaker 7: yeah. Ruby is the best for some reason. Speaker 5: Yeah. Yeah. I think it was right. You know? Speaker 7: So yeah. So so so this is to show you that just having an existing tool with its open API spec, you can play it into an agent and the agent can start doing some interesting things for you. And you don't have to think about the API. You just ask a question to it. But we came up with a a more advanced, version of it. Speaker 5: Yeah. More advanced, which is this 1 is kinda nice. Doesn't work. Doesn't work? Yep. Speaker 5: Oops. Okay. Speaker 7: You know what you what you have to do? Okay. So it's basically the same, but we gave it 2 extra comments, a slash quiz and slash battle. Slash slash quiz, the agent will interrogate the API and will come up with a quiz for us, the user. It's gonna ask us questions and it's gonna give us a score. Speaker 7: And if we have time, we can organize a date of a battle. The battle will be, like, the the agent will organize a battle between 2 sub agents. Each of them will pick a Pokemon and will fight against each other. Yeah. Speaker 5: We we we try the agents after that. Like, this is, like, a multi agent system where you have the root Orchestrate Agentic and then the red and blue trainers. But when you when you talk now, here Speaker 7: we're Yeah. Before yeah. What's what's interesting also is that each of those agents, we give them a different model. You can use really you can have a a fast trainer or, an experienced trainer or whatever, and you can choose a Model, but you can give different instructions to each of those, trainers. But let's start with the quiz. Speaker 7: And the quiz, you would have to participate, I'm afraid. Speaker 5: Yes. Because I tried it earlier, and I was at 0 5. So so the first question so this is yeah. We have we have a piece of tool, a special tool that we have in Docker Agentic that's, Speaker 6: It's Speaker 5: yeah. It's a it's a it's a special tool. I I know. I yeah. I Speaker 2: What? You are right. No. Speaker 5: No. Yeah. I know. You you were right, but I didn't put it. I put something else. Speaker 5: Anyway, yeah, this is a way for the airline to give you questions and so we can show a nice UI. So, basically, yeah, this is, 1 line of 1 line of configuration, and you can even, like, have this your agent have to talk to you like this. What is this? Speaker 7: Uh-huh. Otter. I don't know. The interesting thing here is that the tool we gave to the agent is a tool to ask questions to the user. And this is typically what you would do, for example, if you use code to code, you have a plan mode. Speaker 7: And the plan mode will ask you questions to verify Speaker 5: that what you asked as a set Speaker 7: of features, for example, it has enough information to implement them. So here we get exactly the same tool which is a way to ask you Solutions but for a game. But it could be for any other things and the other hand will decide by himself, by itself when to ask you questions and what to do with the answers. Speaker 5: Does anyone Speaker 3: know? Speaker 5: The second 1? Speaker 7: You're a junior trainer. Sorry. Speaker 5: I guess we can try battle just, like, if you want. Yeah. Speaker 7: That's all. It's gonna take this. So here, basically, we created a more complicated YAML with 2 sub agents, and each sub agent will be coordinated by the top level agent. And it's gonna ask us questions so that, we can define constraints on the on the battle and the Pokemons. Basically, the agents will fight against each other. Speaker 7: There's gonna be a lot of text. And most of us, we don't understand what's going on. Speaker 5: You know, you're not really follow, but what's what's interesting is you you see, like, the root agents calls each 1. And I was like, hey. What are you doing? Then it will answer, like, handling this, and then it will go on, and then someone will go. Yeah. Speaker 7: And under the root, it's using, again, the same API. It's using the API to know all the attacks for each Pokemon. It's it's using the API to know what are the impacts of each attack. All of that, it's backed by real data that's not data that you find on the web. It's something you find on your own APIs. Speaker 7: And if we wanna show you maybe the yeah. Right? I don't know. Maybe we show we want we can show you the the yellow for that. Yeah. Speaker 7: It's really not do you have it? I don't know if you have it. No. So when it comes Solutions, the toolset, we gave the the user prompt. This 1 is a 1 engineer. Speaker 7: It just allows it to ask you Solutions. And then we define the we define the slash commands. Each of them is a separate prompt. And when the user types those slash command Minette tool, it runs, the prompt. And then we've got sub Agentic, and each sub agent has its own description, instructions, its own model. Speaker 7: We can be we can give it special tools. I don't know. Maybe maybe a sub agent needs to calculate, to compute things so we can give it a share or we can give it another kind of tool. And maybe another agent needs to search the web. We can give it tools to search the web. Speaker 7: All of that, imagine an an enterprise, you can really create your, agents that are more, very complex. There's lots of, features that you're gonna get for a very small amount of effort from Docker Docker agents. You're going to maybe you Speaker 2: can skip the next 2. Yes. Speaker 7: Yeah. We didn't show you that. All of that is in Docker Agentic, and I think that's that's Speaker 3: it. Speaker 0: Do we have questions? We have, like, 5 minutes for questions. Speaker 1: Yes. Thank you for the talk. I have a question about, because you showed, like, 80, Speaker 0: 97 Speaker 1: tools. So I'm question is about context overloading. So, how do you handle that, since, it Speaker 9: could have, like, multiple, Solutions and it could overload? And my second question is about skills. Speaker 1: Is there a because you Speaker 9: didn't show that, so is there any integration with skills? Speaker 5: Yeah. So the for the first question, for any toolset that you add in your YAML file, you can also filter to the tools. So you can give it only the tools that you want. So for example, last year, there was, like, a GitHub MCP server that had 101 tools, which was I mean, I don't know why they did it, but you can just filter out, tools that you want. And the second question was Skills. Speaker 5: But Yeah. Speaker 1: But there is a limit for how much tool Speaker 9: you can have. Right? Yes. Speaker 5: It's I mean, OpenAI is 128, but never go above 15 because the the other end will just it wants to to use the tools, and then then it will use them in, like, interesting ways, and you'll you'll end up with something that doesn't work for you. But, yeah, we have tool to create. Speaker 7: I think it depends on the use case. But, yes, if you want to have a specialized agent like this, really, what you need is to give the smallest possible amount of tools. And the open API spec will give you access to all the tools, but then you can really pick the 1 you want. And it will make, using your your agents more, Speaker 5: more efficient. Yeah. What's nice, I think, about operation is, like, by default, it doesn't do anything. So what you have in your YAML file, that's what we send to the other end. So you can you have you can really tweak what you what you what you're sending to the other. Speaker 5: And skills yeah. If you're adding if if you have skills through your YAML grade, then we'll be able to use skills. That's important. It's a it's a nice way to think of it with, Agentic, how they and even different elements, like, how they work in different setups because it's like it's decorative. It's just the YAML file, and you can really try out different things. Speaker 7: Yeah. Okay. Quick question. Like, so all the agent are separate container, and if they are given, like, same API, do they conflict with each other? If, like, if there were, for example, we we write the API. Speaker 7: Right? Would they be conflict as they use same file or should it exist in different container? So here, nothing used any container here. Like, the the agent here is not using a container. It's not running inside a container. Speaker 7: So all the agents, they all run-in the same process. You can have you can connect to other agents through HOA, for example. So Docker agent can expose itself as HOA or can consume HOA, agents that are running elsewhere. You can also run that inside a Docker sandbox if you want, or you can run it inside a container also. And if you need to have the same set of tools, but, usually, there's no there's no conflict between serve agents if they have the same tools. Speaker 7: It's not a problem. But Solutions, you know, 1 agent, you can give it tools that have the same name. Like, let's say you've got 2 open APIs with the same 2 names, we try to not rewrite the 2 names so you can end up with conflict. So there there are there are ways to solve that, but it's, really, it's not really an issue. Speaker 9: Could you please share more information about how behind the scenes you are handling, the memory between agents and inter Agentic communication maybe and, following up to that. This is very fast and good for local testing, especially with the online models. But, where do you see currently in terms of building something on top of this, like, for more than 1 person? Speaker 7: Yeah. So the, Speaker 5: so we have 2 so when you're in a in a multi agent system, we have 2 ways, that you can, define agents. So we have the sub agents, which is, the parent agent sends a message, and the sub agent has, like, a no doesn't know anything. The only thing he knows is, like, he gets a new, completely new, context. And we also have handoffs where it's basically in between like, in the middle of your conversation, we can speak to switch to another Model, another agent. Sorry. Speaker 5: So those are the 2 ways we can we have for inter inter agent community. Speaker 7: You can also share share me more. Speaker 5: And there's also yeah. We have a memory tool, and, also, there's, like, there's a 2 2 tools that we have at toolset that can be shared. So you can also have, like, I don't know, 3 agents and have 1 memory, 1 to do, toolset that they share and they can, like, work on, something. We have 1 more thing. Yeah. Speaker 7: For that question, maybe you can ping us after, and we can discuss about scanning it and how we use it at Docker to to scale our cloud. Speaker 0: Definitely. There's care for the networking, so please come to them and ask questions if it wasn't clear enough or if it was too interesting Speaker 2: and you couldn't ask anything that you have in mind. Speaker 0: Thank you very much. Thank you. Yes. And our next speaker is I'm let me introduce yourself. Speaker 7: Security. Speaker 0: Last 7 minutes left. Speaker 2: Goodness. Alright. So, yeah, you know, software engineer are very scary about giving permission and stuff like that. That's why. Okay. Speaker 2: So someone told me that that should be only demo, no slides, so I don't have anything ready to to share with you except except demos. So, I want to talk about Docker inboxes. So this is a bigger what what what I need to do with that. Speaker 7: What? I don't know if I pull that. Yeah. Okay. It's good. Speaker 7: Alright. Speaker 2: Come on. Speaker 3: Inception. Speaker 2: No. That's fine. Speaker 6: At some point. Speaker 2: Yeah. Yeah. Come on. The very last Okay. I know. Speaker 2: It's just because I'm I'm sharing what I okay. Inception. So, yeah, the, what I want to share with you is how you can run your Agentic, in a safe way, using your word, you know, because we are all fed up to, just, answer to the, to your my the agent to Speaker 5: say, okay. Speaker 2: You can read that file. Okay. You can access to this website. Okay. You can, I don't know, make this API call and blah blah blah? Speaker 2: So, but as I said, there is no way I give, Yellow Mud on my laptop to any agents. No way. So how can we unlock that? Yeah. I know you are doing that, guys. Speaker 7: This is this is there. Speaker 2: But I I don't want to do that. So we introduce the custom boxes. And, what it will do, it will create for you a micro VM with its own Docker, engine running inside, and that will start a container with your Agentic installed in, in the eGInsight. So for example, I'm I'm starting, a cloud, Agentic on a on a project, and that's it. It's done. Speaker 2: You you have to secure a way to run your agents. Speaker 3: No. No. No. No. No. Speaker 3: No. No. No. No. No. Speaker 3: No. No. No. No. No. Speaker 3: No. No. No. No. No. Speaker 3: No. No. No. No. No. Speaker 3: No. No. No. No. No. Speaker 3: No. Speaker 2: No. No. No. No. No. Speaker 2: No. No. No giving autonomy to the agent or just, giving less permission to just restrict it. It's how can we give as much as autonomy as possible and and and, and be sure that we can control what the agent is doing. So this is the purpose of of, the. Speaker 2: So it's not really easy to do a a demo with, Mike, but, anyway, thanks then. So, I've just asked the agent to, ensure so this is a I I Staff my sandbox against a simple web project with a back end and and a front end, and the back end is in Python. So I will ask the agent to ensure the the all the, back end test now ring. But I I didn't set up anything. So, basically, what your agent will ask you, it will ask you to, okay. Speaker 2: I need to do pip install. I'm allowed to do that and blah blah blah. And then it doesn't ask me anything. It's just doing it. Okay. Speaker 2: No way I run that on my machine. No way. The same for NPM or whatever. But now I'm in the secure and and and you you figure out that at some point, you you did okay. I need to to to run that command. Speaker 2: You did it by itself. Cool. So we'll just stop it. Maybe I will Speaker 6: open a second 1 just at Speaker 2: some point that it can help me again. Solutions, we need to, execute an application. So I want it to delete, during that time. So let me no. Okay. Speaker 2: So what I want to show you it's okay. Sorry. I forgot to to remove the the logs. But, what I want to show you is that, if I do, docker exec s bakes sorry. It was called sandbox before. Speaker 2: So that was docker sandbox. It's s bakes. S bakes l s, I can see all the sandbox that have been started on my machine, and I can, also do some exec command. And that exec command, I will, ask my, dev, Speaker 6: that's, dev board. Yeah. Speaker 2: My dev board machine to, to Speaker 8: do, this big, Speaker 2: things like dev board. What what are the process running inside? So just to ensure that I haven't any any contact with my host. So, yeah, what what we can see that we have an agent that is running, with Sleep Infinity. We have a container again, a Docker daemon Runner.. Speaker 2: And then there is my Docker compose app process that is building the application on the other side at the moment. Speaker 6: So this is cool. Speaker 2: There is a TUI also, and maybe I will need to let's see. Okay. And if I do just s b x, I get access to a nice UI that will show me all the so I can create a new sandbox if I want directly. So, I didn't tell you that there is, by default, closed codex, copilot, the current agent, that are available, but you can create your own image. This is the current image for me. Speaker 2: So you can create your own Docker image with your own agent and your own configuration. And, if I go back to my, dev board, sandbox, I can see all the traffic done, by my Agentic. And you can see some stuff I've looked and and some of their past. And this is because we have, some network rules that are applied. So when you start your sandbox, you can define, 3 mode. Speaker 2: The first 1 is deny all the requests, so you will have to give access to any URL to your agent. You have the hello hold, which is okay. You can you can access the whole Internet, do whatever you want with the risk. And you have a, balance mode that is, giving you some rules already defined, like, the the access to default cloud network infrastructure, or you could also have all the access to the Agentic API, obviously, all the all the the the Model and Containers registry also and so on. Okay? Speaker 2: So if I go back. K. Yeah. But with the screen sharing, it's complicated. Okay. Speaker 2: No? My ID is doing weird stuff. Okay. So if I do, s b x, the dev board. Exec. Speaker 0: Dev board. And then I want to Speaker 2: call something like, the the DevOps website. When I run it, you will see that I'm blocked. You say, no. You you don't have access to that. If I go back over there and go to the network, you can see that DevOps appear over there. Speaker 2: And if I click on it, I can hollow it and say, okay. No. No. I want to be able to access to the Devox website from my agent. Come on. Speaker 2: So if I did it again, I will get access to the DevOps website. So I can finally grind, define the rules network rules I want to use and allow or not my agent to access to certain network or not. Okay. What about secrets? Because, your agent is running inside sandbox heap container Runner. Speaker 2: inside his own VM blah blah blah, and, you don't want to share your secrets to that agent. So what, the there is some comments that you can run about the secrets. I can't see anything. And, so, if I do a secret, I will see all the existing secret that have been set. So I have a global configuration for OpenAI and GitHub. Speaker 2: You can decide to configure secrets only for specific sandbox. And, how can I create a a new secret that could be, very easy? I see. That's good. There's big secrets and and then just ask to get the my entropic key, available inside my sandbox. Speaker 2: So you have to know that when you do that, especially for the global 1, you have to create a fresh, sandbox because that would be done at the initiate the injection will be done at the injection of the at the creation of the of the just 1. Okay? Great. Now my application is running. So if I do this big exec no. Speaker 2: Run. If I go back to to the board and I ask, to to Staff the application, So this is obviously, I'm I'm working on compose. I'm using compose. So, I asked the agent to run the compose app so it will speed up the application. But how can I test it? Speaker 2: It's inside the sandbox. I don't have access to what is inside. So, we had it some, some command that allow you, the reports Speaker 5: to publish, Speaker 2: some some ports. And then what I'm able I forgot to show you, but the application wasn't for me. And then, look, I have access to my wonderful application. Okay? So the application is running inside the sandbox, but I give it access to it. Speaker 2: So and just to show you, I can do yeah. I know. I know. I know. If I do a port and publish if I do a port and publish oh, come on. Speaker 2: I I need to list the the the wrong 1. If I unpublish the right 1, I don't have access to to my, application anymore. But it's still running inside the sandbox. Okay? What if and that's the last demo. Speaker 2: What if I want to, have multiple as in running on the, the same Speaker 9: the same source code? Speaker 2: What I can do is I can run a second outcome codex. I can how come Speaker 9: I should have a second 1. Speaker 2: Okay. No. Okay. So my second, because because I need to create it. Yeah. Speaker 2: Yeah. Yeah. I should give it name first. Okay. And what I would do is I will use dash dash branch auto. Speaker 2: That would create automatically for me, a git walk tree, and then, the source code would be split. And and and they won't touch each other, so they can work, together on the same on the same stuff. So you can see that there is a work tree that have been created for you automatically, And then, I can I can ask him the exactly the same question, and he will start working or whatever I want? That's it. Maybe just for you, I have slide, obviously. Speaker 2: Take away, The the main point is not too bigger. The main point is to give autonomous autonomy to your agent not to to block him, but also you have to to let them work alone. But for doing that, you just need to have the right boundaries, to allow them to work. And you have to manage also what they can access to. So having a fine grain network policies, be sure that you can manage your the security injection into them, make will make your life easier. Speaker 2: And for the 1 who want to test it, this QR code is giving you access to, workshop. Based on the same application I have done, it's a step by step, workshop where you will install, s b x, and then you will have all all the stuff I show you. More you will be able to go more in details and see how you can manage poker very kindly, verifying the the network rules and and stuff like that. That's it. Speaker 0: Thank you very much for the workshops and, your codes, but we expect you guys to come back and do real live workshops if you don't mind. But you are doing a 3 hours workshop at the Nutibox, and all of us can afford Nutibox. So Speaker 2: No. No. We are not doing a 3 hours workshop at the Nutibox. Speaker 0: 1 of you is doing Speaker 2: No. No. We are doing a tree a a tree or a deep dive about Adobe Compose application. Using tiny models, but, not not not nothing about, sandboxes. We will love the 2 hour deep dive. Speaker 0: So based upon your feedbacks, please ask them to come back. And in time, do you have questions? No questions? Everything was clear? Speaker 2: That was clear enough? Yeah. Oh, cool. Oh, it's useless, maybe. Speaker 0: It's not. It's definitely not because I know somebody here who's been having those exact questions yesterday. So yeah. Perfect. Speaker 2: So Yeah. You you can you can already install it and use Speaker 8: it. Hopefully. Out of the trip? Is it true? Speaker 2: Yes. It is? Okay. Cool. It's controls, but it's free. Speaker 0: Great. So I suggest we take, like, a 10 minutes break for drinks and refreshments before the last 2 lightning talks, if it's time we do. Yeah. I guess it is. Okay. Speaker 0: Guys, I suggest we get back to work. Okay. Let's get back to the presentation. Okay, guys. If you wanted it, let's get back to our seats. Speaker 0: We have, like, 2 lightning talks left. I mean, 1 talk and 1 lightning talk. Are you good to go? Speaker 6: Yes. We try. Yes. I thought that it was in French, but, I don't know if you know. So, I'm Phil. Speaker 6: I'm social architect at Tokyo. So, I'm not an engineer. I'm working with the engineer, with the customers, and with the technical account manager. I am kind of. I love being with a very smart language model. Speaker 6: Who is using smart language model? Okay. You have no choice. I'm saying that because Guillaume and myself, we we give a third on Thursday about how to use very small language Model. I mean, 4,000,000,000 parameters, with Docker Compass, I don't think Docker Compass and Docker Model 1. Speaker 6: I'm doing this with Guillaume during 3 hours in French, but we did it in in English, in Belgium, and it was a nightmare. I will speak about what we learned and what I learned, by preparing this, long presentation. It's the version number 5. And soon, we will do the number of 6. Yes. Speaker 6: Yeah. He did not he did not know, but tomorrow, almost. So, let's let's speak about my Mac. Probably it's a little bit small. I have a MacBook Pro m 2 max. Speaker 6: It's not bad. With 32 gigabytes of RAM. Honestly, I was very happy when the clerk sent me this MacBook. I think you have 128 gigabyte. Okay. Speaker 6: I did not know before. And, if you are playing with small very small model, you probably know this tool name, LLM Fit. Who is it? Okay. So I tried it, and the first model is 23 30 b a 3 gatorshop something. Speaker 6: Sorry. A 3 b. And the fit level on my machine is perfect. In fact, it's bullshit. If I try it, I will be able to say, hello. Speaker 6: I'm Philip for the the model we answer. Hello, Philippe. What what can I do for me? I would like, an explanation about, structure in the long, and, at the second round, my machine will freeze, in fact, and I can't stop it. And I spent weeks, probably Minette, to annoy the Docker Agent team saying, I I don't understand why the card John cannot run with, small and large Model. Speaker 6: And it was because of not this 1. This 1 is too big. It was, because of the, memory pressure. So let me Speaker 8: Yeah. Just say input. Speaker 6: This 1. So, on the Mac, you have the physical memory and the memory used on your application. Speaker 2: Who is using the camera there on earth? Speaker 6: Okay. Next time, I would like that more people use it. You can use it in the terminal. There is an API, and you can use it in the. So I will Staff. Speaker 6: I will search, this 1, when 3 dot 6. I will take the No. That's fine. Yes. You are. Speaker 6: Never mind. So hello? So you can use the command line now inside, the desktop. Please Speaker 9: explain Speaker 6: explain. Speaker 3: First, Speaker 5: strike. Speaker 6: This is, like, yes. It's enough. And the model will work, and I you can see that the memory is this 1 is probably too small. So let's try it in the other 1. 2 4. Speaker 6: If I try the 9 1, I can stop the presentation. Explain, but in. And you see, this time, double that is working. It's pretty fast on my machine, but the memory pressure is very high. And if you want to use a second model, not this thing, you will double or more the memory pressure. Speaker 6: And if you ask more question, you will advance the Containers size. And at this moment, your machine will die. It's why I had a lot of problem with Docker Agent. It wasn't the fault of Docker Agent. It was the fault of me. Speaker 6: So before, there was always a DNS issue. Mhmm. And in my case, it's always a memory issue. I'm not sure to be able to do the user demonstration because of the memory question, but it was the conclusion of what we we present, in 2 days. And in 2 days, we will speak about, something we call, compose and dragons. Speaker 6: So how to create, game in text mode. The first version was, Dungeon Call of Game. The new version is changing. Yom and myself, we are not game designer. No. Speaker 6: We are not game designer. And, I have a yes. This is the the the first version of the dungeon. And, every time you move, the application is cracked in your new home with monsters and description. But, Shane, it's totally totally ugly. Speaker 6: So we did a second version. It's not anymore dungeon crawler. The dungeon is, generated in advance and after, an AI agent is looking at the map and add description and monsters. I will try to run the map generation. I'll just look at this because nobody helped me. Speaker 6: I was so long. I do remember that. Late. No. So this is, this 1. Speaker 6: So I'm using Docker compose plus docker model runner. Bigger. Bigger? No. No. Speaker 6: No. I need paste. So I will increase it if needed after. I'm using, yes, Docker model runner, Docker atomic compass. Speaker 8: I need the intro. Speaker 6: Docker atomic compass. This is, an integration of Docker model runner in, compass. The map. It's in the declo. Remember the memory question and and my machine. Speaker 6: So this is the map. Yeah. Thank you. So the map, is fixed. You can select the number of room, The generation of the non player characters is under. Speaker 6: You can move from certain room. And with that, the model the another model will be able to generate all the description. So the entire game has several components. This is a kind of, buzz bingo, VI at Docker. We use, MCP gateway. Speaker 6: The Docker MCP gateway, Adjunctiv, Docker Compose, Docker model runner. What else? And for the version number 6, we we use Docker. So there is a dungeon. It's it's a MCP server, and the MCP server is handling the room, the monsters, the moose, and the fights. Speaker 6: There is a dungeon master. It's, yeah, agent. And we have MPC agents. So, it's kind of small agents, and they expose the rest API. At the beginning, I wanted to use, a to a, but, it was too complicated for me at this moment, And I did not have the product yet to do this. Speaker 2: And we have, Speaker 6: an evil boss. So it's like the NPC agents, but, at the end, you have to ask, you have to give the boss some answers. It's a sphinx, and, to to be able to exit the dungeon. And I will try to show you the dungeon with Docker agents. It's not the last version, of Tokyo Agentic, so you will see the black and white version. Speaker 6: This is this 1 I will start. Again, I use, of course, to get out of the compass. And I need just to Architect to the t u TUI. Okay. So I'm only using small models and local models. Speaker 6: You cannot see all the MPC, but we have several agents. And I will try so the car agent is connected to the MCPC server. So I will try something like yeah. Show me the map. For the next version, I will need you because I have a kind of, issue for the display. Speaker 6: That is my fault, but I don't know how to solve it. So the root agent, the Minette Agentic, called the MCP server to get the map. Directly, it's, it's displayed, in, in smart Model. But, I did it last week. So so, I'm here, and I will try to go I don't know. Speaker 6: Oh, it's not Speaker 2: the same map. Not sure. Speaker 6: So I will try to do to go go to the South twice. Georgie, I'm sorry. I cannot cheat with the character of this. Yeah. So I can ask him to go to the exit. Speaker 6: So you can see here, it's all the tool called the LCP server. Speaker 2: Describe Speaker 6: Describe Speaker 3: the Speaker 6: room. Oh, I was sure I was in the other room, but I want to speak with 1 of the n, NTP. So, go to the north twice. 1 again. K. Speaker 6: Describe. Speaker 7: You can use that. Speaker 6: Okay. So sure. Where I am? Speaker 3: He's lost. Speaker 7: Okay. Speaker 6: So I Yes. I go back. So I need to go north again the last time. So go to 0. Speaker 7: Should we come back tomorrow? I'll decide to try. Speaker 6: Well, I tried to speak in a zone. I think I failed my demonstration. K. So now, I'm using another adjunct. Right. Speaker 6: It's you. Strange normally. Should switch to the agent. Switch to look. Yes. Speaker 6: But it's a small model. Okay. And last question, and I stop. Oh, even with the question. Small Model. Speaker 6: So, so, it's slow. So this is the end, in fact, of the presentation, on Wednesday. Not just before. Thursday? Okay. Speaker 6: You know, it's I need to not forget to quit school. Speaker 0: It's good to work. So, if you have question, snap. Speaker 6: And you can use the guidelines. Speaker 2: Is it I mean, what's on it? Speaker 0: Yes. 1 of the questions in the back, Speaker 6: and then I'll come back to you. Yeah. Speaker 2: It seem like this. Speaker 6: In Spanish? Of course. Speaker 9: Thank you so much. This was very fun. Having, lost the first laptop, motherboard permanently to trying to set up the inference. I'm using it quite often, but, I have a question about the underlying local models and how does Docker model Runner. use it. Speaker 9: What does it use behind the scene to run the models? Okay. And, also, I noticed a lot of agents, they use the same Model, but I imagine, you won't be creating a several instance of the same Model to run the different agents. So 1 thing I've noticed a lot, with local LLM usage is it's very tricky to run Serverless requests at the same time because it's a limitation of llamas or llamas that uses it. Speaker 5: But, if you want Speaker 9: to use something more capable like VLM, it's very complicated. And often, in the enterprise, they're also using a Docker container to run their tool properly. So I like some more insight about how, the Docker model Runner. today is running these Model, and, what should we expect in terms of Speaker 6: Okay. Your question is very long. So I'm I'm glad I ordered to no memory. I'm like a a good fish. I will thank you, Mala. Speaker 6: Yes. I I give Speaker 0: you I give you the summary. I give you 1 I'll Speaker 3: give you I'll give you the summary. I'll give Speaker 6: you 1 summary. I'm joking. It's, it's so about, the trait of my machine is to use the same model. So the model is loaded only 1 time. So you can use several Agentic, but with the same model. Speaker 6: If you use several model on my machine, you are free. You you need something better. Yeah. A bigger thing. What we are using to, from an engine perspective, with Docker model, we are using a wrapper around LMS DP, VLLM, and probably soon, with, MLX, but I'm not sure. Speaker 6: And right now, I think that's that's it. But we we didn't invent something more. We didn't try to reinvent the NEMA CP. NEMA CP is doing the job and is doing quite well. So we only use it, and we're just adding some end first. Speaker 6: And, VLLM probably is better for production than LAMA CPP. But, luckily, LAMA CPP is perfect. There was another question in your own question. Speaker 5: No. It's just, it does some performance. Speaker 0: But in terms of Speaker 6: performance, I don't know more than what I experimented in my machine, but, I'm sure. If you have 32 gigabyte of hand, use only a model of 4,000,000,000 parameters with the quantity utilization around q 4, I think. I was able to use, correctly, bigger model, but, I choose, higher quantization. So I lose some What? Yes. Speaker 6: Precision. Accuracy. Yes. Perfect. But at the end, even on my machine, I am with the guardian and with, I'm I'm using the new model named, OneCode. Speaker 6: It's a 4,000,000,000 parameter Model, and I'm able to do some efforts to Staff simple applications Speaker 2: to start by myself, my own development. Speaker 6: So, yes, I think it's useful. I use it to add the news window to. Speaker 0: Thank you. We really have time for 1 last Speaker 7: question, and then Speaker 8: So the last question Speaker 2: was in the end. Right? The last No. Speaker 4: It was about, 1 of the the Agentic you you had was only providing a small API. Why not using, Sandboxes? Speaker 6: Because, Guillaume was no. You you didn't speak my last blog post. Speaker 0: Yes. I do. Yes. I do. Speaker 6: So Guillaume was speaking about Sandbox's season today. I prepared the demo. It's in fact, it's a part of the demonstration of Composer and Dragons. We did it some months ago. But I have some demonstration of Docker Agent inside sandbox with small, model owner, with the model owner with a small, language model. Speaker 6: You can't get that. And, just the trick is to use the the policy to open the appropriate port on local host to be able to join the Carmel and Ronar because this this this is not the same, the car engine. The Carmel and Ronar is not inside the the sandbox. Okay. But it works perfectly like a child. Speaker 6: Of course. Speaker 0: Never mind very much. Speaker 3: And Speaker 0: last but not least, before Woo. Make some noise for her. Figure out how to do that? If you wanna know more information about the build, you can click on the presenters page. There you will have a view project page link. Speaker 0: So I Speaker 8: Check your meeting code. Make sure you enter the correct. Exactly. Speaker 0: So in case Sebastian doesn't present tonight, you have the presentation page just right here with all of the information, all of the technologies that he's using, all the related projects that he might have. Speaker 8: Great. Join now. Speaker 0: Okay. So everybody please mute your mic. And, Sebastian, you can mute your mic. Speaker 8: Okay. I'm good to go now. You're good to go. Oh, okay. Awesome. Speaker 8: Right. Oh, I probably need to share my screen as well. No? Yes. It's definitely not sharing. Speaker 8: It'll be better. I I can just talk if you want, but it will be a bit boring. Too many call. That's going to be interesting. I probably have to restart and rejoin again because it's probably the first time I join, Meet on this laptop because it's a new laptop. Speaker 0: Okay. Why are you doing all of this? Is your company recruiting, Sebastian, or Minette technical director? Speaker 8: Oh, yeah. Sure. We are recruiting really smart people. Okay. Only ones. Speaker 8: Only smart people. Speaker 3: You're in Speaker 0: the right room. Yeah. Looks well, Speaker 8: looks like yeah. Yeah. Yeah. Sorry. I was making oh, there we go. Speaker 8: Joan no. Cloud Knative course. Not at all. But, just yeah. Sorry. Speaker 8: But people I was informed this morning that I was accepted for this talk, so I was not really prepared for that. Sorry. I will switch to English. I'm just writing now. Okay. Speaker 8: There we go. Yeah. Awesome. Oh, yeah. There we go. Speaker 8: Hey. Yeah. Yeah. Okay. I'm sure you're hungry because you are smelling the pizza. Speaker 8: I'm smelling that too. But Speaker 7: don't worry. Speaker 8: In 10 minutes, it will be ready. Okay. So, we had some really, really interesting talks tonight, but there are 2 topics that were not covered and which, in my opinion, are the future of agentic workflows. The first 1 is Java. Yeah. Speaker 8: No no 1 yeah. I'm a Java champion. But, believe me, I think in 2 years from now, 90% of the workflows of Agentic Workflows will be running on the GBM. Challenge accepted. Challenge accepted. Speaker 8: You can mark my words. Okay. Second thing, where will be the those workloads be running? Well, you can have a guest looking at my cap. I'm probably sure they will all be running on Kubernetes because it's already the operating system of the cloud native, and it's there to Staff. Speaker 8: Okay? So, again, I was not supposed to give this demo, but you're on, and I love giving live coding demos. And, usually, I'm also a big a big fan of running domain, elements locally. I was a big fan of Olava, but, then someone that gave a talk just before me and we were just sitting right here, hey. Have you tried Docker Moza Werner? Speaker 8: I said no. Yeah. Give it a try. And then I read an article about Ollama that is structuring around with the the the open source community with Llamathcpp, and, apparently, they are doing really bad things. So I said, okay. Speaker 8: Let's let's give it a try. And, there we go because, I say, And and and, Felice, tell me, you know, they also expose if you use the docker domain, docker Model Runner., it's exposed just an open EA compliance API. So you just can use any of your libraries that use OpenAI, and it will just work. And you can say you can see here. So, I'm just using an OpenAI library using Quercus, the best Java Staff ever. Speaker 8: If you want to be in the future, use Quercus. Okay? Don't use Python or all those things. That's that's funny for when you are at school, but when you are in the industry, use real stuff. And you can just go into a domain, to a model. Speaker 8: Okay? And it's just like, you you can do, like, something like Docker, and I'm doing that with just 1 hand. Speaker 2: You want some hand? Speaker 8: No. No. I'm I'm good. Docker Docker Docker Docker model list. So if you're used to Ollama, you can do just do something like that. Speaker 8: Okay? Docker model list. You get some models. So you can see I have some cryptic domains, models there to to to embed some stuff. But I guess this 1 here, llama3.two, I'm used to use this 1. Speaker 8: And I said, okay. Let's try this with this model. I got a really simple application, and I can show you this application. It's a chatbot. Okay? Speaker 8: Sorry. Back 02/2023. But agents chatbot, it's the same thing. It's just best word marketing shit. So it's the same thing. Speaker 8: Anyway, here, I got a chatbot here. You are a helpful assistant that provides information about countries. Wow. Best chatbot ever. And and you get a user message, and you give it a country name. Speaker 8: Okay? And then it will ping a model. In this case, it will hit the llama 3 dot 2 running on Docker. And, yeah, before I deploy it on Kubernetes, let's give it a try locally. So I do n b n no. Speaker 8: Not this 1. K. Because it's a a Java project, so I can just run it with Maven. K. Let's give it a try. Speaker 8: Okay. It's Runner., and I just open a terminal, and I do Speaker 7: a curl. Speaker 8: Okay. And now let's hope on my history because I am really lazy. Yeah. Local host, chat. Tell me more about France, and I think I changed just the port because okay. Speaker 8: Tell me more about France. No. Okay. There we go. 8080. Speaker 8: Oh, sorry. Oh, I didn't save it. Okay. Speaker 3: Come on. Speaker 8: Save. Okay. There we go. I just stop it. I should not have to stop it, but just to be sure. Speaker 8: Because AT and T is used in my Kubernetes cluster because that's the follow-up of my my my demo. Okay. There we go again. I do my curl again. And it's fucking around. Speaker 6: So h t tp. Port. Speaker 8: Oh, it's http.port here? Yeah. Okay. I changed that. Oh, you're right. Speaker 8: I just changed that, of course, because I'm the kind of guy that changed his Workflows just before the talk. There we go again. Thank you, Clement. Well, just to check if you are paying attention. This Speaker 2: time is good. Speaker 8: This time is good. Okay. Here we go. Okay. Jet resource is special. Speaker 8: Low. Okay. It's not helping me. $19.91 for the English speakers. 191991. Speaker 8: Okay. Tell me more about France. Okay. There we go. And where are you not happy? Speaker 8: Slash hello. Slash hello? Did I miss the slash hello? Oh, yeah. Yeah. Speaker 8: Yeah. Yeah. Yeah. Because again, thank you, Clement. There we go. Speaker 8: Okay. So it's running locally, and this will use a local domain running on my laptop. Okay. Oh, 0, France. Okay. Speaker 8: That's so that's running locally. I did this demo this morning in the in the plane. Oh, no network. So it's using you can check it here. If I go to the Docker desktop and you go to my models, you can see it's running there. Speaker 8: If I go there, I should probably see all the requests tell me more about files. You can see it here. That's cool. I'm I'm I'm I I I love this. Okay? Speaker 8: So but the next step is, okay, I got this. What do I do with my application? Well, I make a container with it, with Docker. Okay? I push it to the Docker hub, and now I want to run it on Kubernetes. Speaker 8: K? So if you're using Quarkus, it's pretty simple. You just have the Kubernetes extension. And in the target, you will have something like that that generates for you a deployment, a service, and probably something else. I don't know. Speaker 8: Let's try it. Let's see if we are connected to a Kubernetes cluster. Still cluster. Cluster info. Sorry. Speaker 8: It's really hard with 1 hand. No. It's a bad day, Sala. Cube CTL, cube cube control. How do you say it, Aurelie? Speaker 8: Please explain to me. Cuba controller. I'm from the South, so I have to say a Cuba controller. Cuba controller. Create a namespace date namespace namespace namespace live live box, something like that. Speaker 8: Okay. There we go. And let me go into this context. Solutions live box. Okay. Speaker 8: I'm in the context. If I do a cube TTL really hard with 1 and get pod, I should have Everybody's no pod? Okay. And let me apply what I just had. Apply dot f target, which target Kubernetes. Speaker 8: There we go. And that is the 1. Okay. K. So now I should have a path that get deployed. Speaker 8: There it is. It's already waiting. You can see there. And if I do a cube c t l exec, I go into this spot. Okay. Speaker 8: I'm connected to this spot. I'm realizing it's really at the bottom, but if I do a curl if I do the curl, the same curl that I did here no. I'm so lazy. So lazy. Hello. Speaker 8: If I take sorry. Sorry. Okay. Okay. I'm inside the pod. Speaker 8: Okay? I'm inside the pod, and this pod is running in Kubernetes, and I can still ping my Docker model Runner.. Okay? That's, that's pretty cool. Okay. Speaker 8: So that was Staff first demo. Okay. Really simple. I can just you saw it. It's pretty simple code. Speaker 8: I got here my my Serverless. It's a helpful chatbot. I package it as a I didn't show you the the command, but I can show you if you want just to be sure. Let me know I'm not joking around. I just did a Docker probably go to my history. Speaker 8: Yeah. That's probably 1 Docker build. Okay. And push it. I did a main thing package before, of course, to have the drawer, and then I just push it. Speaker 8: And then I just can deploy it on Kubernetes. So that's great. That's the first step. Okay. But, again, I'm a big fan of let me go for my all my windows. Speaker 8: I'm a big fan of Kubernetes, but also all the ecosystem around it. Does anyone knows about Knative? Raise your hands. Yeah. I know your hand, but no 1 else knows about Knative. Speaker 8: Okay? That that that's a shame. Knative is let me go there. Knative is a native solution to Kubernetes to have even more serverless. For me, Kubernetes is already serverless. Speaker 8: Okay? You don't deal with machine infrastructure. You say, hey. I have this Workflows, this container. I want to run it. Speaker 8: I have to create a deployment, a service, maybe a conflict map, maybe a gateway. Okay? With Kinetics, it's the same, but you say, okay. I just define 1 resource. It's called, service, a Kinetics Serverless, and I just point to a container, and it gets deployed on Kubernetes. Speaker 8: And the great thing is if I don't get any request, I just scale back to 0 and I wait. When I get an HTTP request, I just wake up. If I get thousands of requests, I can just scale to maybe 10, 50000 to pass, and then I just go back to 0. Okay? So creative is really the way to go if you want to do some Serverless Kubernetes. Speaker 8: And since I already have a, let's say, a container, I just can let me go back here, and let me show you how it looks like. I installed Knative, and there we go. Nope. That's the sorry. That's a incorrect 1 again. Speaker 8: That's there we go. Look at this. 11 lines of yellow. I said I want the service of the kind of candidate for this Docker image. Okay? Speaker 8: And let's deploy that. So I already deployed it. And if we go here and I go just I do a watch on guest pause. Okay. I'm on the incorrect. Speaker 8: Cubans. Cubans, k, native serving. Okay. I have some pause here, but I don't have my my chatbot running here. Okay? Speaker 8: And let me do a crawl, and here I go shamelessly around my history because I'm really lazy. Okay. So here I say, tell me more about Sprint and keep attention. Oh, look at this. I got here a new pod getting deployed, giving me the answer. Speaker 8: So that's awesome. Okay. And, and that's the funny part with connected because after 1 minute, it will still back to 0. That means that I have to talk to you during 1 minute. I can also take a pause and drink some water. Speaker 8: Okay. But I got my answer here. And look at this part, this 1 here, the the Amir for Docker model runner Kubernetes test. I'm really good at naming stuff. 55 seconds. Speaker 8: I think I put the the the natural limit is around 1 Minette, but because you give an answer, it should be a bit Model. And let me just Speaker 2: be 5, Speaker 8: 4. 3. 3, 2. 1. 1. Speaker 8: Yay. And it's just terminating. And now my pod is just my deployment is at 0 point. So okay? All of this, just to finish, and then we can have pizza. Speaker 8: That was a starting point for me. To be completely honest, I was fired from my laid off from my startup by the piece of shit, end of last year. And I have some time left, so I worked on some cool stuff. And the result of that is something called why I I prepared all my my my my stuff here. I got something. Speaker 8: If you go to siltech.org, that's the ultimate result. Okay? Because I just show you here in the beginning, Canadian, some stuff. But with Canadian, you can wake up a workload with some Kafka messages. And what I created is also an operator, some CRDs. Speaker 8: And, basically, it's a bit I'm a bit competing with you guys at Docker, and I'm just on my own. But just with a few lines of channel, you can deploy a lot of Serverless agents. Okay? They can communicate together. They are persistent. Speaker 8: They have a a shared memory using Redis because memory, when you scale back to 0, what's happening with the memory? Well, we are using Redis for that. So, go there, siltech.org. It's just an open source, completely side project Staff that I'm doing on my own. But I just wanted to show you the beginning of my reflection. Speaker 8: Hey. I love Java. I love Kubernetes. I love Kinetic. I love agentic workflows. Speaker 8: Let's see what we can do. And slowly, I came to this, kind of, framework. So, yeah, that's it. And, yeah, I don't want to bore you too much longer. I don't I want you to enjoy pizza. Speaker 8: And for those that are at the box, we can use the box. If you have questions, let's have questions at 100 piece a slice of pizza. Thank you so much. Speaker 0: Thank you. Do you have questions? No questions? Okay. I guess we all know what it means. Speaker 0: You have some beers left. And, so go on our page, terris.aitakers.org, if you wanna know about our upcoming events. And, yeah, that's it. Enjoy the rest of your night. Talk to everybody. Speaker 0: Thank you. Keep caution. And thank you, Docker, and thank you guys for making it this late as well. And thanks to our volunteers on and and and, and, and Mario and Lavin and everybody Speaker 6: who make it.