Hi, everyone. I'm Bruno. I've worked in data for more than 10 years, and as working in data, AI becomes quite a natural thing to work with, but today's not about me. It's about you, so I want to know about yourself.
So let's start with who never used AI? Put your hand up. Okay, that's a good start. start.
Who never worked with an agent? Put your hands up. All right.
Who never built an agent? Hands up. Okay.
Who never got frustrated with an agent? Never got frustrated. Who never got mad with an agent, for example, like, this is not working, this should be doing this thing, it's not doing, there we go.
There's a lot of hype about agents, and it kind of I feel that it's doing something magic, like it's doing things like it's reasoning, it's doing things all over.
Sometimes I don't even understand how somehow I asked it to do something, but then boom, it's done. Sometimes it doesn't.
But what I'm trying to bring you here today, especially for non -technical people is to, let's try to find one way to say this is an agent and this is not an agent, okay? So if we can start there, then we can move forward with this.
So an agent is actually very simple. We got an LLM, which is the brain, which decides, which is predicting text, and you got Anthropic, OpenAI here, you got DeepSeek, Olam if you run local, and it's just predicting text.
So we put a text, and then it will come with the next word, and then we send all that text plus the new word, and then it will predict the next word, and so on, and so on, and so on.
Agent becomes interesting. we add tools to it. And to add tools to it is actually interesting because the way we do it is not like, hey, throw the tool into agent and he knows what it is. So that's something actually helping the LLM understand what the tool is.
And one of the most important is the loop, and it's actually very often forgotten what the loop is for. And the loop, it will be way more interesting as I move forward, but just keep in mind these four things.
And The trigger is interesting as well, because sometimes we feel, if you've been working with an agent, kind of that the agent is always there, always online, but actually that's not true. There has to be some kind of trigger that actually makes all this happen after.
So a big difference as well, what's the difference between a chatbot and an agent? I think chatbots, so majority of you have been working with it, you put a prompt, you you get another answer, and so on, and so on, and so on. But it actually doesn't do anything. It's just that, it's just generating text.
An agent, fortunately, does a bit more of that. So, you put the prompt and then it takes a tool with it and then does something. It might be reading your email, it might be accessing a database. And then if it reached the goal, it stops there. But if it doesn't, it goes back into a loop.
So, again, same concept, just explaining a bit more if you didn't get the message yet, but I can repeat if you want. Who wants me to repeat it once again? Good. Love it.
So let's dig in a bit more into what's actually there, right? So we got our brain. It can be any of these, and there's plenty more. And then we give it hands.
It can be looking into the internet, researching the internet, reading emails, query databases. it can be connected to MCP. So an MCP is basically just a set of tools that it's already pre -configured for yourself.
Instead of you building the tool yourself, it's already done for you, you just connect to the MCP and the MCP does it for you.
Then we got the memory. And memory, actually there's way more, I couldn't put it all here, but there's two important ones to differentiate. One is the context.
So our little brain or big brain doesn't have any memory actually. All he got is like a window of context, which is all that text that I told you before, sent to him or it or whatever you want to call it, reads it, and then comes up with the next token.
So, our, actually, Arnas, which is all this that is picking up here, has to somehow keep reminding him, hey, we just had this conversation and you have these words, and take it again and tell me the next text. And the problem with this is short memory, and it's limited.
We do have, nowadays, models that goes with one million tokens, which is quite interesting, but you cannot put all your documents there, your thousands and thousands and thousands of documents there.
So one solution is using this rack, which is a vector database that's a way of feeding this context, which then is feeding to this. I don't want to go too deep into this. I'm already seeing some faces go like, what is he talking about?
But bear in mind, this is just basically giving memory to our brain. Our tools just give hands so that it can actually do things. And this loop is just basically to make sure that all this keeps happening. Or else what would happen would go once, and then it's like nothing happens after. So bear with this in mind.
So what kind of agents can you actually have? So, you can, there's multi -planes where you can create or distinguish agents, I chose this one, it's from no code, low code, agent, and full code, and what is basically the difference here?
So, the easiest one is no code, so you basically, you probably already use it, like the Anthropic Schedule or OpenAI Agent Builder, you just go, you describe what you want your agent to do. You have no idea how it's doing it, but it's doing it. It's very convenient, but you have almost zero control.
Then on the other side, completely on the other side of it, you've got the full code. You literally code line by line how you're going to interact with the API, how the memory is going to best store, when the LLM answers, how you grab that conversation, how you send it back to the LLM, how you connect the tools, how you develop all that. that, of course, it's way less convenient, but you have full control.
And then there's stuff in the middle, like for example, low code, you have an A10, which is, most of you probably already use this one. It gives you a really nice workflow to work with.
And then we got agent harness already. Basically, they are already ready agents, which your goal here is more to customize them rather than building them. them.
So I'm going to just give you this image very easily. So you got an idea how it works, which one of them, but it's like how you plan.
So this, this is the same agent from everything. But one is just like a bit of a single simple text. This was an attain when you kind of block by block, you build your, your agent. This is Hermes. And this this Python, this, of course, is not the full code is just a bit of the code.
So any coders out there please don't kill me and a little message from my agent he make sure that I sent this for you before I show you the next thing which is one agent that I created how are we on time still a good time awesome so let me show you this so
So this is my daily task planner. So basically, I don't know about you, but I have an app, which is Todoist. And I just tend to put my tasks there, like I have to buy my wife a gift. And then I forgot about it, and then I have a new task, ask forgiveness because you forgot to buy her a gift. And then I have another task. And you know the drill. So you end up having loads of tasks, but you never do them. them.
So, I thought, well, we are in an AI time, so maybe an agent can actually help me with that. So, this is how it's kind of planned. So, I talked to my phone using Whisper Flow. Actually, if you were here on the last event, Benny talked about it. He's right there. How to use Whisper Flow. So, it kind of inspired me to do this as well.
And what it does with with Whisperflow and with Todoist, I send my tasks to Todoist. And then my agent reads whatever tasks are there, checks also my calendar, and then basically fills up my daily, my day. It's gonna say at 9 a .m. you should be doing this, at 9 .30 you should be doing that. I'm gonna show you how it looks like, but it's just to give an idea.
And of course, this could be all in one agent. Why I chose this approach? First of all, Todoist is already very connected. You can use it on your phone. you can connect your email to that, so it's very centralized there which allows me to be, I don't have to code it, I don't have to program the agent, everything lives there and then there's an MCP already connected to, Cloud already knows to list so I didn't have to program anything else on Cloud and then yeah also Google Calendar which is interesting but let me show you how
actually that looks like let's see if I have it here okay so the schedule is easy so basically it's this was the easiest thing I could do I just give it a really really long prompt and so it does exactly what I want has they want it but in the end what it does is something like this I work very well with 90 -minute blocks, usually my energy goes down, so it kind of starts putting all my tasks that I have on Todoist and puts this in each point of it. And it might not sound a lot, but actually it's still taking the priority, how important that is to me, how long it's gonna take, and it takes all this from the information that is on the name.
Sometimes it asks me, hey, I'm not sure about this, This is a 45 minute, it's 30 minutes, what is it? And yeah, this is pretty much what it's doing and it's been helping. So first of all, my wife has been quite happy. I've never forgot about our gifts again. And also I became way more productive, which for me, it's quite interesting because I was kind of all over the place with the tasks.
Another example of an agent, so this is running every day.
So that's all loop that I've talked about, but there are some times that we don't need a loop.
The loop was, has anyone joined Geneva recently? Like this year, two years? Okay, so maybe this is useful for you.
I joined Geneva, I came in April, and one thing that Geneva has really cool is a lot of information in the websites. However, they don't tell you what is in those websites, they don't tell you how to research.
There's a lot of things that you need to do once you arrive but they don't tell you what so I asked my agent to go through all the sites that exist in Geneva from the GE dot CH from other other legal not legal other with events and other sites that and asked and asked it hey please give me all the things that necessary for me to do during this day so it
It starts telling me like, well, you've got to have your insurance. You've got to have your basic insurance. You've got to do this. You've got to do that. And I've been finding a lot of things just based on this, and it's been quite useful.
So it read 370 pages. And yeah, if I map, if I show you how much it went for it, it's like it's really huge amount of websites that read to give me all this information. information.
We can also show you that information is online. Just quite scared.
Anyway, this was what I wanted to present you just to give you an idea what an agent is and what you can build with it. And yeah, I think we can move to the questions.