AI Agents of Change: Creating, Reflecting, and Monetizing

Speaker 1 [00:10]

You made me quite nervous that it's quite full But I will give an introduction About what we're gonna do. This is a QR for you to download the REPL Everything we will talk here. It's gonna be there so you can just we made the material for you to use a standalone or For you to replicate if you think it's a cool workshop and you want to replicate it and use remix free software and free ideas. So you just remix and share. And while I will talk a little bit about the concepts, it will be nice if everyone got this website for you to create an account here. We will use this platform because it's really easy to use. Yeah, so just create an account will take one second but because internet may have issues.

Speaker 2 [01:07]

Um...

Speaker 1 [01:09]

And this is, wait, wait, oh, sorry. This is the workshop repo. And this is, if everyone still needs a moment, please raise your hand. Okay, I'll wait. This workshop will be in two parts. We have a very nice agenda, and we will try to follow the path. So it will be a little bit of hands-on. I will talk a little bit about what it is an AI agent, if you're not already tired to hear it. I accept complaints and discussions if you don't agree with me. So that will be the practical part. You will create one very simple one. And I want you to pitch yours, if you have something in mind. And then we'll do a break. And we will do some other types of exercises, thinking about what you're doing. Do you still need some time for this particular QR? Then I'm going to take you for the second QR. This is for you to create an account.

Speaker 2 [02:33]

Okay, this is working yeah in the second part of the workshop we are going to do We're going to talk a little bit about ethics and equip you with a framework on how you can self assess How good or harmful your agent can be? In a non black-and-white world of agents

Speaker 1 [03:00]

Everyone good with this cue ball? If not, raise your hand.

Speaker 2 [03:08]

I mean, theoretically, if you got the repo, you should have access to this QR.

Speaker 1 [03:14]

Brilliant point.

Speaker 2 [03:16]

or the link

Speaker 1 [03:19]

Okay.

Speaker 2 [03:22]

and

Speaker 1 [03:25]

Ooh, so this is our, I can, yes. Cool, merci.

Speaker 2 [04:06]

So for a little bit we are going to keep the QR there until everybody's like, okay, we got it

Speaker 1 [04:12]

Is it readable enough for at the end of this room though?

Speaker 2 [04:25]

Just a little check.

Speaker 1 [04:26]

People at the back of the room, can you hear me well enough and can you read well enough? Okay, cool. I'm basically talking whatever is written there. So this was a little bit of introduction of who we are. But we have left our... This is basically a little intro of this workshop is about is what we have spoken AI agents became very hot topic nowadays and we will talk a little bit of what it is and some approaches to create it and then we were reflected because I believe this is I truly believe that this is Almost as a big of a revolution in the technical field as it was an internet and that means that there's a lot of changes coming and I'm inviting everyone to collectively think of what does it mean as an impact for our society and this is why the ethic parts come in and this is what you will learn hopefully at the end of the workshop so what it is an AI agent and how can you create it how can you use the different platforms and tools box for you to create it how do you can critically think about what you have created and how you can reflect on how in a bigger scope this is affecting our society. And this material is reusable, this is a little bit about us.

Speaker 2 [06:02]

to our

Speaker 1 [06:03]

agenda I think I'm just not rely on the internet

Speaker 2 [06:12]

Peach.

Speaker 1 [06:16]

So what it is? I mean, anyone opinionated here that would like to define or have heard or had your own definition of what an AI agent is? I will assume you don't. But again, I will tell you what I think it is, because I have been working with this in a few months. And this is how I think, oh, that's what it is. And I will do that in a very simple word. For me, what an AI agent is, it's sophisticated automation that can do things on your behalf. And you can use different tools to do that. But that's basically what it is, it is automate things.

Speaker 2 [06:59]

and

Speaker 1 [07:01]

And it can be as sophisticated and as simple as you want. And most of the times, we use AI agent because we will make use of language models to do so. That's basically this, what I understand. Would anyone would like to compliment or deny that affirmation? I like the beta. I really like it. But then I have been doing a series of interviews with someone who's my mentor and who taught me to code. And he's the brain of this platform that I am currently working on. And this is the platform we use because I think it's really cool and easy and you have free credits. So I think it's cool to share that. And I like his definition about, okay, but what is a good agent and what the heck should I expect? Because the booking the thing, it kind of can give you a little bit of a brain knot. But I think what we aim to do when we're creating one, and I think this is a very nice guide, is that you want that the user, whatever, will use the agent, and that can be you, just for you. You want to expect something, and that needs to be reliably delivered. And the cool thing about working with language models is that different, just so you know, who here comes from typical software development, software engineering field? Yeah, me too. So for me, it was a gigantic shift in the paradigm on how I do things. As software developers, we want to do things the way we want, and we expect that function to behave the way we want it to behave exactly. With language models, it's not exactly how it works. It's more a partnership of some kind. But that changes a little bit of our paradigms, and I do have an interview and a video about that, because that's when I give a little light bulb. There's a part of the work that you do want it to be exactly what you want, like, for example, an API request. You don't want that to devaniate, and instead of going to youtube.com, you're going to youporn.com. Oops, maybe you want, but you want it to be reliable, right? You don't want to... That thing you do want to be reliable. But your agent, you want to expect to achieve the same expectations. So you want that to consistently deliver something that you expect. It doesn't mean that one plus one is two, it will always be two, it's a bit different. But there's a lot of challenges when you're building a good agent, and a lot of complexity. So when you're a software developer, you have like this canvas that you know the steps of the architecture you will build, and this is kind of very finite canvas. You will have from the, if you think from the user's experience, you have, and think about web development, for example, you have a website. You have a button. You will click it. It will take you somewhere. There's a few steps that you can do and you will control it. As AI agent developer, that space became almost infinite because if you want to create it But from the scratch, you will need to think, well, I'm not entirely sure what my user input will be or wishes, so we're very used to chatbots, but maybe a user do not want to deal with text. He wants something else. Maybe he wants a sensor input. Maybe he wants an image. So the world of inputs become not just a little thing, but a gigantic thing, and you as a developer, you will need to create every single part of that. That's one thing. And then there's how you connect all these things. If it is an infinite entry and an infinite output and you, as a developer, you still need to create all these links that become a very heavy work. This is why I will suggest you to when you are starting approaching agent developers, and that's what we will do here, you start with some frameworks. So you start doing that little by little and learning to debug all this infinite parts little by little. Makes sense? Cool. Who here has already created an agent before? Have you run it from the scratch or have you used a framework? You? Cool. Anything else? It's not actually, it's a little raw, no? You kind of have to do some work there already. You? Cool. Yeah. So I think it's a little bit like this. So you have to, but mostly what you want, you want to understand the user intent. And the user intent in the AI world is not that direct input. It's not like go from here to there. It can be very, it's an intention. It's not a full function, but an intention. And that is something that's the first step you need to do. How the thing get in and what do I do with the thing that the person who knows who it is want and who knows what the heck they want and how do I do with it? So understanding the user intent is one part of creating a good agent. The second one is producing the right condition to the desired outcome. For example, a very common use case is that people want to use agents to create content for YouTube or for social media. So the condition for the desire outcome is that you want to push a button, you want to create a thingy, and you want that thingy to go to whatever social media you want and the videos you want to post it right away. And that's the desire output. You have to create an image, you have to post it on my behalf. Making sense? And then you have to make sure that that aligns with the expectations. So again, I don't want to put in the wrong website, and I don't want the image to be not as I expected. I was expecting a fluffy cat, and then it came back like an ugly, microscopic ant. If someone has seen it, it's very nightmarish, I do not recommend. So you don't want that, you really want the fluffy cat. Cool. So that's simple, in simple words, that's how I define it. And I talk a little bit about that, but there is a little bit of mental model shift. And I'm very visual, so for me to understand better how to break down those parts that I need, that's, of course, a little bias from my software developer experience. I try to do some comparison, and I hope that helps you, too, this graphic. So I'm kind of, oh, okay, when I'm creating that, I'm thinking about what it is they're error handling. For example, oh, I need a human to tell me. So it's not only debugging the APIs or whatever developer experience you have. You also need to debug the human. So sometimes you need to have a human in the loop for the systems that you are creating. And so here I left a little bit of text. I want to jump into the practical part, but it's a little text that I talk about the paradigm shift. would love to hear your opinions if you read it, to tell me if that's any helpful or not. And okay, so how do you design these agents? This is also like, not to become so abstract, but you want this, you hear those keywords a lot, which is capabilities and tools. Those are two words that I think you hear a lot when you're talking about agent, and I want to break that down. So tools is a concept that I think OpenAI released, that it's kind of you create a function that the language model can interpret, and it's quite cool. So when you say the capability, you say, oh, my agent has capabilities, and that is changing the hair color, and no, I'm inventing now. Like go back to social media because I deal with that all the time. So create a post and create an image. Those would be my capabilities. And those capabilities you will create as functions inside of what we call those tools. So big words. Doesn't mean much. It's a function. And it's a description of that is using mostly like the description of it. A big thing for agent development, usually people who come from integrations or like to pipe, I like to think about AI agent creators as plumbers, because what you do a lot is plumbing things. And most recently, most precisely in November, Anthropic has released a new protocol to talk with AI agents called MCP, or Model Context Protocol. Anyone familiar with that? It is a pretty cool thing. It is a kind of a very clever API, but it's kind of a big pipe. I like to think of it as a pipe that pipes both ways. So you can... It's like a super power REST API that lets you connect everything with everything. And brilliantly, finally, it is standardized. Because most APIs are not. But the MCP is. That means if you have a server that opens the door, you can talk with anything. With anything. It's pretty cool. But it also opens doors for a lot of security issues. I'm sorry. Oh, thank you. Cool. And then there's a whole question that Laura is not here. But if you want to use it from scratch and create your own platform, for example, your external web app that uses agents, I would recommend you to work with a good designer. Because nowadays we got very used to chat bots, but the design is kind of suck to work with this infinite amount of possibilities that it brings to us. So there's other things that you may want to cross, and that's a really high level. That is memory. There is a lot of people that may, you can connect your agents with databases, most likely probably vector databases or any database if you want to store that. But mostly, agents, when they finish their workspace, their memory will finish there. This is kind of important to know, because people may think that they are learning. They're not quite. They will learn on a very short project. So it's nice when you work with that and you think about the project, and the project will have some scope. But outside of that, things are very limited. And most frameworks and platforms will work like that as well. Oh, shit. Okay. So let's do some hands-on. In the repo. Everyone got the repo? Yeah. So the first thing is for you to create an account in OpenServe. Have you created with the other link? I think you can see better than white. So what OpenServe will help you with? It gives you a lot of straightaway benefits for creating your agent. It helps you with the user intent. There's kind of shadow agents that already figure it out, like we will run an example now for our basic agent that you can say, hi, hello, hola, I mean, use basically lots of languages, and it will respond, kind of understand the user intent. And it gives some other benefits. But before I start that, I know some of you came here with some ideas on your mind. So before we run the base example, I would like three brave humans to pitch your idea about the agent you want to create. No brave humans today. Okay. Stand up and speak your... Thank you.

Speaker 3 [21:09]

Can we create an agent that will create posts for, you suggested social media, for Twitter or X-profile, for a specific domain, I don't know, machine learning or AI? Yes. Perfect. I'm excited.

Speaker 1 [21:29]

second brave human

Speaker 2 [21:33]

Hi hello

Speaker 1 [21:34]

it would

Speaker 2 [21:34]

it would be super interesting to see

Speaker 1 [21:36]

to see an AI agent that is able to help you with the research for a book and then makes a suggestion for like an how to say that like a table of content and a concept for the book and the topic shall be actually right-wing incidents and Germany. Good agent. Third brave human.

Speaker 3 [22:06]

Hi, I'm not sure if it's like to involve but I wanted to have an agent which does grocery shopping for me So basically it goes on like some recipe website It picks a recipe and then it goes to some supermarket and uses their online ordering thing And then all the groceries just show up and I don't have to deal with people in the supermarket

Speaker 1 [22:28]

Great. I want that. I'm using your agent.

Speaker 2 [22:31]

Yeah, and it's going to be also surprise recipes, right? You have to tell it how complex your recipes should be.

Speaker 3 [22:39]

I'm a pretty good cook, so that's fine.

Speaker 1 [22:45]

Anyone else would like to pitch your agent idea?

Speaker 4 [22:55]

Sorry, I cannot stand up because...

Speaker 1 [23:00]

No, you can sit, it's okay, it's okay.

Speaker 4 [23:04]

want to make an agent like which can the information is really distributed like usually when I'm kind of searching for a recipe maybe it would be on Instagram but I'm searching on YouTube so kind of the real world I'm talking about so an agent I give him in the thing just like I'm making some dish pasta let's say and he picks that information from Twitter Instagram and everything and then gives me up in the one single place

Speaker 1 [23:33]

Cool. Those are actually not super complex agents. You could do it in maybe yours, I don't know, because I don't know any supermarket that has an API. Maybe. But that's the only issue. But there you could accomplish that, like, really fast. I do have to say something, and that's I'm addressing the shame head here. We do have OpenServe do have a very nice SDK, but it is in TypeScript. And I have been working in the past months to trying to translate that to Python, and I have not been so successful. But I have done it for this workshop, a basic example to run, so you can try it out. And I will keep on working on the SDK. So for example, the Twitter, I will show you, everyone has created the account? If someone not, please raise your hand. Are you having problem with the link or you're just creating? An error. I'll go there, check it with you. And the other one that couldn't have it is also an error? Oh, okay. Cool. You could... So, here, you can create... I will do step-by-step how to do it. In this repo, so one of the exercises is... Let me check my agenda. Oh, shit. Okay. one more thing about multi-agents. Multi-agents is now a thing, it's a kind of approach for you. You want very specialized single agents working together and that's another cool thing about the open server. You can just create single things and they will put everything together. You can also use other people's agents together with your agent to create a good team, to create whatever your ideas are. Cool, so here I actually added the whole SDK in here. Just so you know, the SDK is a wrapper for our API, so you're basically sending a message, whatever HTTP requests to some endpoints, but that's what it's doing. But it's way nicer to use the SDK. Can you read it? Here? It's very tiny. But what the SDK does is that this is the main one, just so you understand. It will allow you to create this example, which is a basic one that is kind of easy. It's just like use Pydentic for type check, and we import here is the source, but it's the SDK. Eventually, I will publish to yesterday I learned the UV, I found it cool. We're using logging, so, oh, you need .env for your environment variables, and you can just substitute this one for .env. define classes, this is just a name, and here you define a function. Here's a super basic one, so you will run, just to test it, a really basic agent that what you will do, you say it will give you hello, goodbye, and if you say help me, you say well I can only say hello or goodbye. It's a very silly one, but this really basic agent, help you understand all the steps. This is why I have created. And for you to start, this is another part that I, it's important. So it's just this file and the system MD. And the system, it's kind of a, it tells your language model, which mindset you should have. This is like a kickstart for a language model. So you are a helpful assistant that helps me tell the workshop about agent, whatever, they that's what the system is kind of important and it's part of what you need to do. In the readme you have step by step about what you need to do. Shit there's one more thing there I forgot. The way you communicate from your local machine to open serve it's creating a tunnel meaning you make your computer you will create URL on our computer that will from your local thing communicate with the platform you need to open this tunnel I recommend using ng-rock and your rock if you download it it's free takes one second and you can just download it so sorry I forgot about the tunnel it's in the readme you can do it but if you want to run the agent now you kind of need it forgot and then it's It's just a simple Python, and I will do it with you, so we can do it together. So first, you need... After you download the ng-rock, you literally just say ng-rock http, and the door, you want to use it. We typically use 7378 for the examples. And that creates this nice URL, which I find is super cool. Then now my computer is open to the world. And that's just the local part. Then you go to the OpenServe platform, you say, I want to add an agent, and we will create my first workshop agent. The end point is your tunnel. This is why you need it, because this is how they will communicate. And you add the capabilities. The capabilities is this that is in your example. Create agent. So basically what you're doing, you're instantiating a class of the agent SDK. And you're just saying, I have a greet capability and a farewell capability. It also has a help. And when I do this capability description, my shadow agents will find the agents I need to accomplish my idea by this capabilities definition that I have created. So it's kind of clever in that way. Like I want to create an image, but you didn't choose an agent that creates an image. It's not magic. You need an agent that creates an image. Then the project manager or the shadow assistant will call it. And this is for the integrations part. This is super cool. You just add a scope. We have some native ones. There's already the MCP. You would just literally say, add the scope. So I want to get account info and post to Twitter, post it. You just add the scope, and that's it. It's kind of magic. There's probably a PR right now that any agent will be able to use any capabilities. But right now, you need to add it to the agent, and then That's it, it's very simple. You need to connect here in integrations. For the basic example, we won't need it. I'm just showing because one of the ideas for social media, you just authorize and that's how you do the integration. It's super simple. Am I making everyone confused already? Yes. Oh, damn it. You're NG rocking.

Speaker 2 [33:07]

Holy shit

Speaker 1 [33:09]

I'm sorry about that. With what? No, I don't think so. I have to log in. There's some others, but I think that's the easiest one. I have a password management tool. If you want, I can do the two-factor authentication for you. Total privacy. You just link it to my phone. All fine. Yes, cool. So for you that has already done the ng-roc, again, all the steps are described. I would just show you for us, sometimes it's just nice when someone shows you and you say, ah, that was it, it was so simple, most of the time. So here's my first agent, first agent. You need the API keys per agent, so you need to create your agent first. Agent. So your endpoint is this forward in here. The read, farewell, help. In this case, I don't need any integration because not actually needed to say hi. And then you have this manage this agent. And here you can create a secret key. Yeah, so you will use the platform as your language model, and you will use it as well to use the benefits of the shadow agents and the whole structure, and to connect with other agents. So you're already creating a project. So first we're creating our agents, right? So you create this API key, you copy, and you add that to your environment variable. me very securely showing and being recorded to the world, and they will add to your .env, save it. And then you run, so you have the ng-rock running in one terminal, and now you go to to another terminal, and you will run your example, use a Python, in my case Python 3, examples basic agent.py, cool, your agent has created. As soon as it's approved you go to a marketplace, everyone can use it, but before that you can use as much as you want to test it out, and then now you use your agent. For use your agent, you create a project, and you say, hey, or whatever you want, like a hi, I'm Brutus. Say hi and bye, because there are the two capabilities that it has. I'm not sure if it doesn't talk in language, project, workshop, agent. So the way you test your agent is through your project, creating a project. Where are you so far? Has anyone could have run the ng-roc on your computer? Okay. Should I cut it? Should I cut the tutorial? Oh, okay. Frustration level, how is it going? That's a high level frustration? No. This part? Perfect. Is anybody still doing ng rocks? OK.

Speaker 2 [39:03]

You can also just look at the laptop of your neighbor that might have done the ng-rog already and pair up Like yeah, the steps are all documented And you can always like follow up later

Speaker 1 [39:24]

Okay, where are you at? Cool did you add it to your environment variable the secret key? Perfect, then you run your agent and you create a project Like I say here your name and choose your agent here like a your select your team. You're choosing your agent My and you run it How do I interact with it? My first. So this is my first I have created. I select it. And I just, OK, start the project. And then you do it on its own. As a developer, you want to see the, oops, it will always first ask for a human interaction here in the loop to say, oh, OK, this is the part of understanding the user intent. Oh, so you mean that I should, it breakdowns in tasks. So you should say, hi, Brutus, is that what you want? If it's not that, you can fix it, or you can just approve a plan, the agent should say bye to Brutus. Yeah, that's what I want, cool, so approve it, and then it called my agent. You will see that it's in progress, you follow the steps, and you can follow from your terminal as well everything that's going on okay receive 200 HTTP requests so successful you will follow all the steps with the logging and you say even the what he returned and to you bye Brutus have a nice day and here it often creates an MD file with your research in your case would it be like for example the book research here you can see the output you can also ask to create a PDF or a dynamic HTML HTML for example let me see one by one now where you're at you're having issues can I see your end

Speaker 3 [42:02]

in the project

Speaker 1 [42:04]

Are you in a, did you do the VA and you installed the requirement? Requirement, so if you come here to the readme, you have to, is all the Python stuff, create of Yen when you style the requirements. Yeah, yeah, yeah, I did. Yeah? Yeah. But you're using the .env and it uses the .venv. You should not do that. Well, not really, but the .env. Can you maybe delete and create a new one? Yes. Okay. I will give you more like ten minutes for us to deal with this frustration. But again, it is meant for you to do in your home, and the basic agent is for you to understand that fun, like, oh, break it down to see how simple it is, and then you can create things as complex as you want. Because shortly, we will, yes. I just have, like, a small question. Yeah. When you created an agent, you wrote, like, the name of the functions that you defined in Python with this grid, this grid capability for the world capability. So, can we, like, create any function that we want? Any function? Yeah. Any function? Yeah. Any function? Yeah. Any function? Yeah. Any function? Yeah. Any function? Yeah. Any function? Yeah. Any function. That's why it's like this infinite amount of possibilities. And then we have to just copy-paste the name of the function because we know that these functions are available and we use them. Or you can also describe it. But for me, I'm more lazy. I just describe it. You can also use natural language to say, this agent do this, this, and this. But for me, as a developer, I like to debug it. So when I put the capability, it calls this name. So I'm like, ah, OK, those are when I remember what it is. But if you think that a user will be a user agent, then describing would be better. You cannot create tunnel on your computer?

Speaker 3 [44:43]

and it doesn't exit.

Speaker 1 [44:49]

There's also a local tunnel, which is an open source project that is tunneling But I just never actually use it This is just a silly you can try it again you can try it Sorry, I don't know how to go over the security any person good with securities and No How are you doing? You're chatting with him? Oh, perfect. Yes, this is another thing. He's not responding. Rude.

Speaker 3 [45:39]

When I chat with it.

Speaker 2 [45:46]

it right we have this

Speaker 3 [45:47]

Farewell and health capabilities. How does it decide?

Speaker 1 [46:04]

So, two things, there's this one, you have to, it's waiting for your human input for you to approve the plan. And you will call it through the capabilities that you have defined. It will, how does it know? It's a kind of magic, to be honest, it's a kind of a black box magic that it reads the capability and, oh, these agents do that, and it calls these agents for this function. So if you'd say, it understands that, and that is the reasoning part that the platform adds, that's kind of hard when you do it on your own. If you say, hello, hola, , it will understand the context and understand that the user is saying hello. But that's why we call it reasoning, that it's harder when you build it from the scratch. Okay, so with this, it figures out on its own which of the capabilities it wants to call. Yeah, but it basically reads the function and see what interprets what it does and try to match what the user is trying to do makes sense but really really it's kind of black box

Speaker 3 [47:11]

Mm-hmm.

Speaker 2 [47:12]

after adding the agent?

Speaker 1 [47:14]

You create a project Projects create a project and then I say say hi. What's your name? Hi, I'm hey You would like up say hi to me and

Speaker 2 [47:35]

Can you specify that in?

Speaker 3 [47:36]

agent

Speaker 1 [47:42]

Well here you say you tell what the agent is to use like what you wanted to do I want to say hi and bye to me Okay, and you give the name Next and choose your agent no you have to choose your agent it will give you some oh you maybe those are good for you but then you want to choose your agent cool next next now it will put you in the loop to approve the plan and you can either say yeah it's a good plan or deny it yes We have a chance here. So we didn't know where we were going, so we created the engine. And now we're trying to find information about . And you're up. You installed it? You would brew? Cool. We're not yet . Yes. We have an engine. And we're trying to find information about . okay so close okay go back so this is a generic agent if you close it you this is you don't need you want to create an agent you don't the agent builder you don't need code at all. There's no code. Did you put your end point here? The build agent is no code and add agent is when you need to add your agent. Welcome. Anyone else has an emergent question? Otherwise we're going to the second part okay okay approve the plan or not restart the task Is it giving the server a lot? Can it restart the task? Oops. Maybe engineers are doing some push. We're in beta still. How are you doing? Okay. So BuildAgent is a no-code agent builder. And ad agent, it's where you send your code to the platform. Yeah. Yeah, with no code, you build directly on the platform. No, that's just if you want to just play with natural language to create your own agent. And yeah, it's good to play. If you think it's good, you just approve the plan. And that's it. And then you will see the thing going, and you see the message on your terminal of the process. Question issues? Not sure if it does what you ask, but cool. Anyone else, Merge and yeah. Approve the plan, that's it. See in your Python, you should return at least a 200. You have to run your agent. From the Python side, yes. agent.py. So you go there. But it's interesting because I forget that all the time. And if you go back to the platform, it will tell you that it's a... I'm trying, but I cannot find this. It's not there. You can build agents without code at all, if you want. You need these connections because you need to open your code to talk to the platform. code because that's a very simple agent but you may create an agent that has an API call to miss hundred of other places and do all the things not quite we have a secret management that you can use for others to authenticate You can just, just for you to test the platform itself, you can just use the no code function because you test just with natural language how to create an agent. But in the deployment mode, you would deploy your agent to a server. You cannot upload the code. It will not read your code as you were expecting to do. Yeah, you can drop files. You can, but it's not going to read your code as an agent and make it work, your agent. That's usually what I like. I use that as a context. Or you can try it out, see if it works. basic PI? The basic agent? But it's... How did you build your agent? Ah, it would be interesting to see how it would work, what we would do. You don't do anything. It would do on its own. The idea is that you use the minimum of human interaction. I honestly don't know what we would do with the coder. We would probably need to have a coder or some other agent read it. But for now, I would just use the, if you go to agents, build agent, and then you use the agent builder and you create just with natural language just for you to try it out creating agents. But if you want to create a specific agent, then you would need to connect that to a server. It's not an agent, yeah, let's do it, yeah, okay, whew, oh, shit, oh, shit, okay, we're going to the second part of the workshop because that part you can start

Speaker 2 [55:58]

I remember.

Speaker 1 [56:01]

Should I think should I sing Star shining bright above Okay. Hey everyone, we're going to the second phase You now have access you have credits. You have a detailed tutorial You can go on your house and do it on your own. Um, I Hate going to workshops when if you're frustrated because I cannot finish the tutorial and I became that person I'm sorry about that, technology as it is. But now we have the very cool part of the project of this workshop that is reflecting about what we are doing.

Speaker 2 [56:48]

So, coming into the second part when we are talking, before we even start, I want everybody to open a new markdown file on your terminal, on your computer, and just write down the idea of the agent that you would want to build. Take a minute. It doesn't have to be extremely sophisticated. Just write it for yourself. You know, it's in writing, and then it's happening, right? I mean, if you put it in writing. No, I'm kidding. Just write it be as creative as you want. You know, you want to destroy the world. That's fine It's you're able to do it. You want to save the world, you know, you have all the spectrum there Just write down for yourself Ignore what's on the screen here. It's not important What would you build if you know everything about agent building? So you're getting there Thank you. Is everybody done with their amazing idea? You are allowed to change it in the future, so it's not set in stone. So the thing about agents is that it's not just like you're not just writing code. It's also about the consequences. And, you know, a lot of people are scared and, you know, people are talking about AI taking our jobs and so on. But, I mean, there is maybe half-truth in that. But every agent that we build could impact jobs, perpetuate biases, or create or destroy social values. And now you actually are responsible for thinking about all of these things before you even start writing code. That's why I said, like, you write down your idea of the agent and you can already start thinking about what are the implications. that before you actually start implementing the agent because after you start implementing who here has experienced falling in love with their code yeah so it's really hard to throw away code that you wrote and invested so much time just because it turns out to be bad or written for the wrong reasons or you know you didn't like think it through properly so and you can all read this document on this a little bit of a snippet on the ethics considerations the idea is that these agents that we're building are existing on a spectrum it would be so easy to say that my agent is bad is an autonomous weapon or my agent is good it's just you know helping with diagnosis in a medical world so that's like super extreme but to be honest i'm pretty sure even the cooking assistant, we could use it in an evil way if we really tried. Like, for example, introducing a little bit of poison.

Speaker 1 [60:20]

Right?

Speaker 2 [60:22]

Like, if your wife wants to kill you and she hacks your agent, she could, you know. So, anyways, just, you know, hypothesis out there. Be careful what you're building in your house and making available for everyone. And so, we all know from building models, working with data, that we have the topics of bias. The data is biased. The models are biased. That we have to care about. and there's all those considerations that we have but one new consideration that is more prevalent with agents is accountability who is responsible for the decisions the agent is doing once the agent is allowed to do decision making the developer the people who build the llm the people who put it in production so it becomes a little bit like you know you have all this conversation on autonomous driving you know tesla doesn't want to be responsible but is it and and we also have regarding costs these agents are using a lot of energy and a lot of resources so it is nice to play around and put toy examples out there and for letting everybody to use them but in the end it can be quite expensive do you want to be responsible for that environmental cost. Just a couple of days ago, I don't know if you've noticed in the news, there was like one user asked themselves on X how much electricity is burned by people who are saying please and thank you on chat GPT. And the CEO said that, well, it's just, you know, millions, Just being polite But he also said that It's worth it You never know Maybe it pays off to be polite To the AI chatbot So that is like Maybe what is Europe It's not like really I think who here is agreeing It's worth the billions In environmental cost To be polite to chat GPT yeah so we have the devil's advocates here of course some people so who here disagrees with this we should all be like impolite on purpose just to make it cheaper that's a that's a different direction of course yeah so the so the opinions are you know it's not really clear what is like good or right and so on there is a very nice blog about this you can read it it's on AI agents best practices so it has like what are the building the places that you have to pay attention and think about like pricing models in reducing hallucination having use cases human oversight you saw it also earlier that you have to say yeah that is exactly what I meant but later on do you want to have human oversight. So today, we are introducing this framework that you will have a little bit of time to read. It's the AI Ethics Canvas that is written by Tristan Goetze, which is also an adaptation from the Business Canvas. And the idea is, with that idea that you have for your agent, is to open this canvas and you pair up. You're not going to work on this alone. So you turn around to your favorite neighbor, pick a neighbor, yeah, and you open the, you go to resources, and here you have the canvas. There's the guidebook, so here, no, that's the picture, so this is kind of how the canvas is, but we also have it in Markdown, so you can, and you kind of start thinking about Like, what can be, what is the purpose of my agent? What data I'm going to use? What are the vulnerable groups that could be targeted by this? What are possible harms, possible benefits? And I think the exercise that, the most important exercise for you to do today is really think what are, like, critically, what are the possible harms? I think we normally are very romantic about the code that we write, and we think nobody's going to use it. Who's going to know? Nobody's going to do anything evil with it. You know, image recognition. Nobody's going to do anything evil with it, said a lot of people. But then they got proven wrong, right? So I think you can read the canvas, the PDF. There's a whole PDF. No, this is the same one.

Speaker 1 [65:22]

exercises.

Speaker 2 [65:28]

It's in exercises So in exercises you have it as a markdown that you can interact with And You can read up on it you can but I think today we don't really have time to go in a lot of details about everything so you go to exercises ethics and there's the markdown and just ask yourself some of these questions don't answer all of them, but try to think critically about like the more negative ones and you have how many minutes do we have okay so you have for this five minutes don't answer all the questions just answer some of the questions and pair up and discuss this with your partner let tell your partner what you think are positive negative things so let yourself get challenged

Speaker 1 [68:10]

He didn't like it. No one likes being poisoned. I think you can finish them. Did something happen? Did something happen when you upload the file? Did something happen when you uploaded the code? Yeah, I don't think it's going to happen because it's not expecting to do it though. Just as a code for instruction, how to install things without a code, if people have a chance to spend an hour before they install, otherwise they also have a short time to try to install things that might not work. That's a very important thing. Everything that needs to be installed, you need to know. Before. Most people don't do it anyway, but at least they had the chance. So I spent an hour trying to do all the things that I had to do to create an account, and then in fact I got a notification. Then I tried to do all the things that I had to do to find the security, and then I just realized that you were the one who deserved this. I needed this opportunity. Oh, I'm sorry. Because I didn't think that anything was going to help me. It's just... But then I decided. And that's the thing, they never face the security part. They go, oh, I was not counting. That then takes longer even. Yeah, I'm sorry. Learn I hope you can do it later, though, because it is so fun when you start doing it. Okay. Good? Time. OK. OK. Get out of there. Get out of there. So... I'm singing in the rain.

Speaker 2 [72:02]

cannot think of anyone okay if you're happy and you know it clap your hands okay so this was the part of thinking about the risks and i now want you to think about what how can you flip that risk what is a countermeasure that you could implement you can think about implementing that you would turn that risk into an you know opportunity don't we don't want that so you don't have to solve all the risks but pick one and again discuss how could and then we can do sharing is caring whoever wants to share about a risk that they identified and like a problem that their agent could be misused and how they would have mitigated that and then we can have a sharing session at the end so can you flip that is the question the problems that you thought your agent would have five minutes

Speaker 1 [73:37]

Yeah, I think it was enough. Good, good. And you smile. Kidding. Thank you. No. No. I want that. It's good, yeah? No. It's a case? What is it? It's from the computer? Yeah, yeah, it's from the computer. What? But is it because your computer let you do that? I can just put everything in there. I thought it was like an external thing that you could have put it. Oh, that's so cool. That is so cool. Oh, that is cool. So again, you can put in logos, you can make it stamp, or you just put in text. I love it! It's good, yeah? It's good! Is it time? No. No? Sorry. Maybe to post it. The post, ah. I can't speak out. Oh, but they're not seeing. I'm just, I was checking if I should do this, yeah. Wait, are we in the...

Speaker 2 [77:10]

So brave humans who want to share who wants to share a Risk that they identified with their agent and a way that they identify how to mitigate that it that risk

Speaker 3 [77:48]

Is it on? Yeah. Okay. So earlier I had the idea of automating posting on Twitter. And I just now was thinking about an agent that would look through other people's tweets, identify the ones that have a need for, I don't know, if I have my own product, a product that serves their need. and for the agent to write a helpful reply to them and then casually sprinkle the product in there. And the risk we identified is what if the agent misclassifies what the person is trying to say. Maybe they bring up a problem, but then they are like, and this led me to develop depression or something. And then the agent could reply and, I don't know, give like a nice recipe for cooking a cake, which might be appropriate, but not in that context. And mitigation strategy might be to have like a human in the loop, for example, that the agent identifies the original tweet, prepares a reply, but a human in the loop would need to check if this is really appropriate.

Speaker 2 [79:06]

Thank you Anybody else wants to share

Speaker 3 [79:22]

So we talked about building an agent that would scour different APIs, catalog them, write kind of a user documentation and remember how they worked so that you could easily find out how to interact with this system. And it could offer it to the end user to be like, hey, would you like to make a query towards this or this or that? And then quickly we realized that that's probably really dangerous for cybersecurity purposes. And I don't think we found a really good way of curtailing that system, except maybe only use it as a cybersecurity tool on internal systems or similar. Debate is welcome.

Speaker 2 [80:03]

Thank you.

Speaker 4 [80:09]

So, we did think about an agent that can apply to jobs, making the resume tailored on the job descriptions, but the risk was very high, like people could misuse it for like they don't have the skills and they still apply for the jobs, so we did not find some valid mitigation of the risk for this agent.

Speaker 2 [80:42]

Anybody else wants to?

Speaker 3 [80:56]

Thanks, we expanded on the food idea, must be hungry, and we thought that this agent could take in a photo of ingredients that you have at home and then suggest some recipes based on just what you had. A risk here is that you might put some fish or some meat into the image and it may not be fit for eating anymore because it's out of use by date so the risk there is you know you might give yourself food poisoning benefit is that maybe no one will ask you to cook again but to deal with the risk of food poisoning we thought maybe the agent could also then just do a check to make sure that the food is still suitable for eating

Speaker 2 [81:52]

Nice. Thank you

Speaker 1 [81:54]

Do you want to give the example, the original example?

Speaker 2 [81:58]

Yeah, so we actually tried really hard to come up with an example of an evil agent on purpose and it's kind of hard to get that, you know, like defined, like apparently, you know, nobody wants to build any evil stuff and we are not evil enough. Anyways, so one agent that we thought about would be someone to help people with job interviews and to act as a, well, actually I asked it to act like an evil interviewer, and perplexity translated that to a skeptical interviewer. So this, because, you know, nobody wants to do evil stuff. So the idea of this agent was to be used by people who apply for jobs to prepare them for those unexpected questions, like, what makes you think you can apply for this job, that you're qualified for it, you know, some questions that get asked in job interviews. And of course, this can get misused by people preparing interviews that don't want to prepare themselves for the interviews to come up with questions to lock people in states of high uncertainty. so yeah and I think you built the agent right

Speaker 1 [83:15]

I'm having issues with SDK, it would be my personal nightmare. Do you want to close? Yeah. Good. So as a closing, like for us, the takeaway that we wanted to take, I hope that was somehow helpful for you. But first, that you have a cool tool now that you can use at home and start playing with. It's a really easy entry to start dealing with the agents that can get fairly complicated. But that you can still do a few cool things. But the second one is to really reflect and Laura Summers gave a brilliant talk yesterday talking about the aesthetics of AI and how violent it is, but it is kind of invisible and we're all buying in a very fascist aesthetics and this is quite dangerous. So when we create awareness about what we're doing and conscious, that can be a very powerful tool. And one of the things that is shifting a lot, as we have mentioned, it is the job market and how much we're seeing our own jobs being affected. That can be as positive as as negative as we want. But when we create things in the world that can be really helpful, for example, the same agent that could be evil could actually prepare people or create conscious for recruiters. We're actually putting something out there that it is your direct decode and creation into the world. And a possibility for, and that was the part of earning with that, there's a lot of platforms just like OpenServe, but many others as well, that are allowing you to directly sell your code to someone. So there's a lot of responsibility that comes with it, but a lot of also market shift possibilities for us as individuals. And I think that can be something cool to try it out by selling your agents. And if you have an agent that is ethical, that has a purpose, and can do some good to the world, why not earning out of it? And that's it. We deeply appreciate everyone for staying, trying out. Your feedbacks are pretty welcome for us to improve next time, and thank you.