How AI Changed Software Development Forever — Building the Agentic Future with Andre Baltieri [MVP]

0:00 Welcome to the MC65 with our podcast. Artifactil intelligence is changing software development and speed. We have rarely seen before for decades the developers wrote applications by defining exactly what software should do inputs, business rules, updates, database, interfaces, outputs, then generated on iLive and suddenly the developers could generate code, explain antimalosystems, create test, debug applications, interact with development environments using natural language. But that may only be the beginning. We are now moving forward an AI agent software that can reason about goal using tools and AI's main and daily context, interact with other agents and put it to execute complex workflows with much less direct human instructions.

0:57 So what does this mean for the developers? Do they simply make us faster or do it fundamentally change what software, deployment, actually is and where do technology such as.net, Microsoft, agent, framework, AI models, tools and so on, and the architect just fit into the new world. Today I'm joining the Andre Balthary, Microsoft MVP software development specialist with more than two decades of experience in development. Andre has also worked with teams across Brazil, the United States, India, England and the Netherlands. And he also received the Microsoft MVP of our firm, Preston, 40 times for his contribution to the technical community, his areas of expertise, including artifact intelligence, the Microsoft agent, work.net and software development.

1:50 And we welcome to the Microsoft C6-5 of M podcast. Thank you, Miracle. It's pleasure to be here today. I thank you. It's awesome. So, 40 times. Yeah. 40 times. That's a long time. Yeah. Before we talk about AI, take us back to the 2003. What did the software development look like when you started? Yeah, well, that's nice. It's a nice question because I started with dot-ads in 2003 since the beginning of dot-ads basically. And things were completely different.

2:31 So, at that time, we did not have so many information about software development. So, the internet itself was being created. It was very, very hard to learn something new at that time. I remember that I had to purchase some official Microsoft training, travel to another city at the weekends to do some training because it was crazy. You basically did not have access to any kind of information. And then, not sure if you are aware of it, but Microsoft had a very popular training center called MSDN, Microsoft Developer Network.

3:30 Okay. It prevails to Microsoft to learn. It was the big place to go when you want to learn something from Microsoft. That was awesome. So, it all started there. So, since then, I was working with C#.net and it's very cool. It's my passion. Yeah. You have gone from desktop applications to the web, mobile, cloud, containers, Microsoft, serverless, and now AI do is that current AI transition feel different from those previous tests. Yeah.

4:10 It's, it's sound like a new era, you know, because I, when I started, I was moving from desktop to web. This is the first era, they say, from software development. Everything was new. We were talking about Linux servers along. We, we do not ace. Then we have another big movement to mobile development. So, this was the new, nothing to create apps for iPhone and right.

4:51 Then we have the cloud movement. When we move it to, to cloud, but nothing, so meaningful like move to AI, this, this era that we are passing today, because this is, for me, is the biggest change in software since the beginning, at least for my career. It's like, it's, it's like a new language, a new way to interact with, with software. So we basically move it from assembly or a very low language to a brand new language.

5:37 Yeah, that's, that's, that's an amazing shift. Where, where we are today are AI coding tools, most of these sophisticated auto complete or are we already beyond set. Yeah, we're, I think we, we're in, in the middle, because I, I can see that we, we have much more to, to come yet to, to AI. But, AI is able to, to create a lot of things for us today, from the simple components to, high-calplex systems. So, they're evolving so fast because since we, we have a model that can create other models, so AI can evolve very fast.

6:32 And, but, but this doesn't mean that we, we don't need programmers anymore. I think there's a bit of misconception out there, because AI can generate code for us. But, we need to ask the, the code generation in a certain form that it needs to, to understand, we need to understand the business concept, business rules before, uh, talked to AI. And we still need to, to have the technical knowledge, the technical background to ask the good classes to, to AI. So, we, we still need the programmers in, in the middle, or orchestrating agents, AI.

7:18 But, uh, our, our work today is, is, is very, uh, it's, it's changes because, uh, I don't write, for example, uh, much more code as, as I read, uh, in the past, for example, this, this, this, few last months, for example, I wrote, uh, just a few lines of code by hand. Uh, most of my code is written by, AI is written by agents. And I can easily create Python scripts, uh, for, for, to, to automate my boring tests, but they, so it's crazy.

8:00 It's, it's crazy how fast is, is changing and how it is, uh, generating good code day by day. And, and can you, what, what, what I, I found really interesting, what, what is the, the difference between AI coding assistant and AI coding agents? What I, what, yeah, that, that's nice. Uh, there, there are many approaches to, to reach, uh, the, the main goal for your software development. You can simply, uh, but, uh, put the AI assistants in your ID, for example, I was a source studio, from Microsoft. So I, I can simply open a tab, aside me, the studio and say, hey, what's going on here?

8:50 What are suggestions for me to reflect or this method to implement this feature? So they, I can assist me. They can act as, uh, senior developer, um, aside me. And, uh, the, the good points is that I can say to AI, hey, uh, just focus on Microsoft documentation. Don't look anywhere around. Okay, just bring me answers from this source of truth, which is Microsoft learned, by the way, because they have, uh, the, the, the MCP, which was called, uh, the, the, the model context protocol, which is a protocol that allows us to, uh, uh, to scrape information inside, uh, Microsoft to learn website. So this is, uh, a way that AI can assist us.

9:43 So we call this, uh, assistant AI or AI assisted code development. We, I think is the most common features nowadays. Uh, most people that I, I see programming use this, this kind of, uh, assistant. Uh, so we, we don't need to go to websites in more search for an answer. Uh, we can ask to, to AI inside our, uh, our ID, uh, and we can debate about this. So, uh, explain me why are you telling these? Explain me, uh, or show me three more ways to, to do this. So this is good. This is the, uh, assistant AI, the agent development is a little bit different because you can ask AI to do the complete test for you.

10:33 By the way, you can ask AI to plan this before you even start. So you can simply open the same way, open, uh, the, the chat window and say, hey, uh, copilot, uh, by the way, I use copilot from, from Microsoft and it can connect to multiple, uh, AI agents. So cloud, uh, codex, uh, doesn't matter. I like it because it can talk to multiple agents. So say, hey, copilot, I need to do this task, uh, the, my problem is this, uh, what is our plan to do this? Then, uh, it can write a plan, uh, show me this plan, show me how he intend to implement this and then decide implementation. This is a, uh, a different approach because it can trigger multiple agents at the background.

11:32 It split the task in sub-task, micro-task, trigger multiple agents. Uh, this is agents can run in different, uh, AIs like, codex, uh, uh, uh, cloud, uh, term nine and get the job known for me. So this is, this is by, by fire, one of the craze is way. And there one more way with this, the spec driven development, which we can, uh, write specifications and place this is specifications inside, uh, some specific folders inside our software and say, hey, base it on this, uh, specifications, uh, do the system for me.

12:17 And, uh, sorry, and he starts to code it's, it's a little bit, uh, crazy, but it's, it's very fun and he's doing a very good job. Since you have a good specification, you have a good architecture, uh, very, well-defined, it can do, uh, a lot of work for you. Um, but, um, can I only prompt it or is there capabilities that do useful software development agent activity? Right. Uh, is what capabilities do, uh, useful software development agent activity need? No, it's just prompting. Just, you can just prompt it. It's, uh, uh, a natural language. Uh, so even if you, you don't have, uh, the, the, the specification, uh, yourself or even you're in the agent mode and stuck in a test for yourself, you can, uh, you have different modes to interact with.

13:25 It's a component, for example, you have the agent, bold and the ask mode. So if you, if he's doing some task, you can interrupt him and, uh, use the ask mode to say, hey, uh, what is going on here? What are you implementing or why are you implementing this way? So we have a lot of flexibility during the, uh, the implementation. So this is, it's, it's very, very like, uh, a team working for you inside your, your ID. So that's fun. This is, this is awesome. Um, especially I only prompted so that's, that's cool. I don't need, I don't know some context or ask for awesome. I don't know.

14:15 Yeah, that's, that's, that's, that's really, that sounds really simple. Like that. Yeah, yeah. Um, Microsoft has these, um, yeah, agent framework. Can you a little bit say what, what gives Microsoft agent framework is? Nice. That's nice. Uh, we, we talked to here about, uh, how we can code using agents, how we can delegate to, to, to agents, uh, our work, right. But, uh, this is just one part of our, our job. Other part is to integrate AI into our software. So I need to provide capabilities to my users.

14:56 For example, to interact with AI or, uh, to create something we call Ragn, which is retrieval omitted generators to improve, uh, the quality of the response, uh, for our models. So there's a lot of stuff going on behind the AI and it's very difficult for us to manage all the things because we need to, uh, we talked about Ragn, MCP servers. There's a lot of stuff. So Microsoft agent framework is a tool, uh, is a set of tools that is built from Microsoft and, uh, it is provided us to, to do this interaction with AI, sign our software.

15:43 It's very easy, for example, to, it's like plug and play. You can, uh, bring, uh, Microsoft agent AI to your software and connect with multiple AI's, which is, uh, is very common in the, it's, it's a little bit tricky to do because, uh, you talk different with, uh, codex and then you talk to, to call out for, for example. So Microsoft agent framework, uh, it creates a service API. It's created an interface for that. So you can talk directly to, to, to Microsoft agent framework and he talked to codex, he talked to cloud, he provide MCP servers, he provided, uh, uh, access to Rags, uh, he provided dating gesture, which is another process to, um, pick all the information that you have on your company and, uh, but it's either your database, which is, uh, seems to be, to be simple, but it is, is very complicated.

16:47 The last part is that, uh, it's built by Microsoft. So if you work with .NET, for example, you have all the things, uh, ready for you for free. This is, uh, inside and you can speed up your, your development, your AI integration, your, your software, uh, within a click, for example, it's, it's very fun. So this is the, the other approach that we have to, uh, integrate AI into, into our software, which is, uh, I, uh, programming the school. And we, we did, we, we, we factor our, our system, the, the, the, the, this year, and we are integrating everything with AI. So, uh, people can, uh, during the, the, the lectures, people can ask something, say, hey, what is going on here?

17:45 I did not understand this, this concept. So all this integration is, is made by, uh, Microsoft agent framework. So that's very cool. And I'm not, uh, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, what do is the agent framework give me that I don't already have by simply calling in their, I model through, through the app. No, it's more like, uh, a group of, of tools that you have, if you, if you were to, to call this by hand, you have a lot of, a lot of work to do because, uh, the way that we, we talk with models, uh, tend to be different from company company.

18:37 For example, if you use, uh, Azure AI Foundry, which is awesome, uh, that you can concentrate on your, your models, your little bits ahead because, uh, Azure AI Foundry provide an API for you to, to call these models and, uh, it all pass through the same service. But as soon as you try to connect, for example, with Google to, to consume these models, you will have to call different API. You have to send different information. So, uh, you, you will need to deal with all this code inside your, your code base. So it's, it's a lot of code just to do simple tasks.

19:19 So, math, math, as we call math, uh, it help us without the code. And, and, um, I think how, how nature is dot that act early for, yeah, for the, for these, in your generation of, of, of any application is, is it well positive? It's not so major, how good, good is dot, dot net, uh, that's a, programming language. That's a good, that's a good question. Because most people, uh, think that, uh, because that was created in back into, to thousand, uh, it is, uh, an old framework, dot net is the framework, C sharp is the language, the programming language.

20:11 They were created together. And, uh, most people, uh, think that, uh, it's a little bit older, but it's not, C sharp is evolving. Now, we, we have major, uh, release from that every year in November. So, it, it's crazy because, uh, every year, they, they bring us new features, they bring us, uh, even more speed and, and, uh, and security side, dot net. Which is awesome. Uh, that's in, there's, uh, a very, very good job. Uh, and on the A.I. side, I have to confess that Python is, as the pioneer for, for AI and the, is, is the first choice because, uh, most of tools that we use today are based on, on, on Python, they were creating a Python, which is a very, uh, good language.

21:07 Uh, especially for, for AI because it's very simple, it's very straightforward. Uh, but we, we have evolved so much with C sharp, uh, in a way that's, uh, C sharp, it's basically, uh, as, as easy to learn as, as Python over the day. You have a lot of features and with Microsoft agent framework, for example, uh, by the way, Microsoft agent framework is a, a, a, a, a, a, a, a, a, a, two previous product that Microsoft build with, which was cement, Carano, and auto-gain. So, when Carano was built to, uh, deal with models the way I, I told you, uh, and auto-gain was, uh, was a project that you, you can, uh, uh, orchestrate workflow, create pipelines using AI.

21:57 So, Microsoft agent framework is, uh, is the combination of these two features. So, uh, is it, I can say that, that, uh, we have a complete workflow, uh, to work with AI sign.net, a complete and a modern workflow to work with AI sign.net. We still have more to come, uh, they are, they are planning more releases over Microsoft agent framework and, uh, I don't expect it to have the same popularity of Langchain, for example, which is, uh, the, the same as still, uh, but in Python. But, uh, I, I don't miss any, any feature, no, no, days, I can code and sign.net, I can use my, my code base. And by the way, not that is, uh, a very broad ecosystem in a very complex system.

22:48 Uh, we can create mobile apps, game development, software development, AI, everything inside of that. So, it's very complex and it's very modern also. Awesome. Um, but how? I say, if an organization has years of business logic implemented in about that applications, how easily can that functionality become available to these AI agents? Good question. For example, uh, here we have a, our entire software is built inside that. So we have a lot of code is sharp here. And with math, for example, with just one line of code, we can export that method and create, uh, a metadata for that, uh, to AI communicate with that method.

23:46 So, like an export a function of my system, to AI, I say that, uh, with just one line of code. So if you have dot that in, you start using math today, you can create an, an MCP server with a flick of the wrist. It's, it's very easy because it's C#, it's native code. So, uh, it's very easy to, to integrate, right? So that means, um, companies may not to, to rebuild their X, their existing application to become an AI application or a, I know, that they can start today, just, uh, bringing Microsoft agent framework, for example, we have other packages like MCP for, for that, which is, you can just add to your, your project and starting export MCP.

24:46 You, you can start today, it doesn't need to, to plan anything, uh, much more complex, for example, you can start simple, uh, and then obviously, uh, to, to create an AI first, uh, on your company is, is, is something that involves, a lot of other things, but if you want to just integrate it with models to push AI to your application, it's very simple straightforward today. And, is there any architecture changes? So how do you build an AI application differ from building a, I say, traditional, domestic application? We tend to separate some sort of things because one thing that concerns me a lot using the AI is the secured, right?

25:45 Because we don't want you to expose anything outside our, our organization, or for our users. And since they use natural language, it's, it's hard to to predict which kind of answer they will ask for our system. So we have to double check to create a lot of tasks. But depending on how your system is built, if you create small functions, if you pass these functions, if you ship these functions, granularly, it's, it's very easy to integrate with AI this way I told you, exporting these functions or wrapping these functions inside a, right?

26:36 So usually we, we just create some separated class, just like I would practice to, to export this, to not export the core function. And in some cases, we create another project just for AI to, to separate the things, say just everything related to AI, you put the science project, but the architecture itself, you don't need to, to change anything in your architecture, you can export this. But if you're thinking to expand this to, to evolve this in a future, probably you need to rethink about your architecture, how you'll, you'll place AI inside your current software.

27:28 And in your, your company in general. Yeah, also, I think when I think from, yeah, the architecture, I say tradition of software, follows logic like this X happened and execute J, which agent, all the, with an agent, we might be instead say, here's the goal that I'm in what actions are necessary. How profound is this, is this architectural change? Yeah, the nice, I got your question.

28:10 It's a, it's a kind of a part of the shift, right? Because we, in software development, we tend to do things very strict. So you have one way, other way, you have to pass from a to b, this certain kind of information. And for example, exposing your, your, your functions, using Microsoft agent framework, for example, which will create an, an MCP server for, for your code, you can expose this function to AI. And you say basically, hey, this is the function to turn on the lights. Okay, this is the function to turn on the camera.

28:53 This is the function to record some podcasts. Then with this information, you can use natural language to interact with it. And that, that is the, the cool part of AI, because we, we interact with AI today using natural language, but inside chat, GPT, inside a cloud, things like that. So we tend to, to do some prompt, this, but to write some text, to write some code. But no, they, we, we can create a prompt, for example, to, to create a product inside our database.

29:36 So you can simply say, hey, register this product for me, the ID is one, the prices. There, and that's, that's fine. It's, if we expose the, the, the function to, to register a product, to create a product, we can interact with AI that way. Another cool thing is that if you're using a multi model, model, a multi model, model, you can simply take a picture, for example, from a product sent to that and say, register this product. It can bring information for you and try to register this product.

30:18 So we stopped to, to do in some areas, right? We stopped to do all the methodic way that we used to. So you have the fields to feel you have the, all that stuff. On the screen and we can start to interact more with prompts with natural English say, hey, do this for me and it will find a way to do that. This, I classify this into two, two pieces. The first part is the IDA potent way, which must be that way it can change.

30:59 You can, you should expect the same result always that, that seems. And then we have the, this new way that you can simply send a prompt to, to, to AI and that will follow the prompt. So we can interact with this way. That's really amazing. What's possible. So, yeah. What's happened was this traditional design patterns in, in, in an agent architecture, will they stay there or is it, is there, is completely new?

31:42 No, I think the, the, the good practice of software development is still alive. So, right code one time is, is very easy to maintain, right? A maintainable code is very, is very hard as Uncle Bob says, because even AI, looking for a massive code will do a massive code, right? They follow the pattern. So if you say, hey, implement this, it will look to our code base and say, well, I implement this the way it is implemented on, on your code base. So we still need to think about patterns.

32:24 We still think about to, uh, uh, uh, clean code, which is something we, we, we talk every day. But it, it's way easier to implement, uh, no days using AI. You can, uh, you can even forget about it for, for a second, when, when you're planning, for example, when I, I'm planning, uh, I, I don't specify any design pattern for AI. I, I, I live a blank canvas at least for, for the beginning and say, uh, let's do this plan. They will come with a plan for me. And, uh, if the implementation is way different than, then I think you say, hey, but what about using the strategy pattern here?

33:12 What about using the prototype design pattern here? And I, I can suggest you, to, to then, uh, and, and do the, this kind of implementation because it's, uh, create some, maintainable code, uh, two, two things that, uh, you must achieve in software development is a testable code because, uh, when everything changes and, uh, everything will change, uh, you need to pass everything. And I'm good looking and maintainable code, right? So that these two, two things are negotiable. Uh, you, you always need to achieve this, just two things.

33:54 So, uh, design patterns, best practices, still present. Uh, they, they, a little bits, uh, on the depth layer now because AI, we interact more with the surface, uh, the, the code itself, but, but they need to be there. Uh, you, you, you, you still need them. Yeah. Yeah, I think one of the biggest conceptual change is that developers are used to, yeah, deterministic systems and AI models are probabilistic. So, um, how difficult is this transition, especially when I think, um, when you solve, when you test software and, uh, yeah, the same input, not necessarily, necessarily produces the exact same output.

34:44 Okay. Nice. Uh, we call this stochastic because, uh, every, every AI generation is, is different from the, the previous one. So, uh, that, this is, I think this is the, uh, the most, uh, difficult part of interacting with AI being a programmer because most of, uh, programmers that I, I, I know today, expect the same result because we were built to that. So, uh, we, we do this when they expect the, the same, the same result, but this is not, uh, this won't happen with AI even generating a code. If you ask the, the same code generation to, twice in a, in a minute, it will generate complete, different codes.

35:31 So we have to, to, to keep that in mind when we're at your specifications. So we, we are focused, we're focusing more on algorithms on, uh, how things should work rather than how the code is written. So, uh, basically what I try to do is, uh, understand that, uh, there are many ways to do the same thing. So, I don't have to, uh, to do that, uh, the way I intend to do, I can leave it to, to, to, away, I, I just have, I, I just need to have this don't. So, I don't, I don't try to, to, uh, get to read to my, to my code to, to, to, the, the generated code I prefer to, uh, line it to, to, to, A.I. Just, uh, validate this code.

36:25 If this, this code is, is good. If, if, if he, he written the, uh, I, I go to approach, uh, I, I don't mind. So, that, that's fine. Let's, let's do this way. It's like, uh, you're, you're delegating the, the, the creation to you, another person, uh, barely, uh, it's very hard to find a, uh, another person who, which will code like you code. So it, it will code different from you. Uh, they, both of you can, can achieve the same result, but, uh, the, the approach will, will be different. So, uh, using the, uh, is that way. I think the, the most challenging, uh, things behind it, uh, is more to, to test the software.

37:10 Because when we interact with, with, with A.I., for example, uh, we have millions, uh, uh, ways to, to do the things. So, uh, when, when you expose this to, to million, user, you have million users with medium ways to do the, the, the, the same thing. So you need to, to create a very good, hard, award, rails, uh, to, to that. We had so, so many cases of prompt injection, which is a, a technique to, to, grabbing information and to, uh, hack your, your system using A.I. So I think this is the most cool, certain part.

37:52 And, and as, as you may imagine, uh, it's very hard to predict everything that, uh, people will say, uh, there, there was a, uh, a fun case with, uh, a guy, uh, asked the A.I. to sing, uh, some Windows keys for, for him. So, he told A.I. a story that he was, uh, his grammar, sing, uh, song, uh, some Windows keys for him when he, he was young and A.I. gave him some, or he told me those keys. So, uh, that, that was for, but, uh, it was for us, but not so fun for the, the company running.

38:36 They, they, I, so, but we, we have to, to be aware of that. So if you're exposing the, uh, a point for the, the people to, uh, write something inside our system to interact with, with our system, you, you, you should have the, the, the guard array is in, in, in, in mine. And if you're coding is the, the same way, if you're coding, if you're asking A.I. to write your code, you have to, you need to have the, the guard array is to say, Hey, don't do that. Uh, try to implement this pattern, try to do this way. Right. Yeah. Well, when I look a little bit into the, the developers groups are in, in, in, in YouTube, I have the feeling, uh, that's often focused on, on, on the model to use, uh, but the, how important is this, is context compared, uh, with, with model selection from your perspective.

39:34 Uh, I think the context for A.I. is everything, you know, we, we need to, uh, understand and manage your content, context, uh, uh, very carefully, because, uh, if you, if you write a simple context, A.I. Won't do the, the job and few write a complex, uh, context, A.I. We'll lose itself. So, yeah, you need to, to write, uh, the, just, to stuck in the middle of this. And sometimes it's very hard because sometimes we think we know about the context and we, we, we, we, we, doesn't know. So as, as a programmer, what, what I usually do is basically, uh, implement some business rules.

40:21 So if I don't know the business rules, how can I ask for a item, implement for me? How can I implement myself? So more than never, uh, we, we know they need to get into the business. So we need to understand the business, understand the needs and another thing that I, I, I, I always, uh, talk to my students, for example, uh, you need to question, you need to, to ask questions to your, your business and say, hey, what is going to hear? What are implementing? Why are we implementing here? Because, uh, with this question in, in mind, we can, uh, provide this to, to A.I.

41:06 into, into have a, uh, good context. And this is the hardest part to, to, to be true because, uh, software development is, is a little bit straightforward. So you have the good practices, you have, uh, the, the concept, the base content, the fundamental concepts of computing, um, which sends, uh, CSI, I was born and from, from now, but, uh, the context itself is, it, it's hard. When I, I talk to you, you, you have something in mind, you say to me that when I talk to another person, they have a little bit different context, a little bit, uh, point of view.

41:49 So merge, uh, everything of that and pass this to, to, to A.I. It's a, it's a, a very, very challenge. And you have to mix that this, this business, uh, context with the technology context, because A.I. also lead, uh, the directions on the technical way. So which architecture we are using, which stack, we, we are using, uh, what comes next for, for, for our software. And so, but, uh, all this together is, uh, is a, is a very, is very challenging. So, we are, we are still learning from, from that because A.I. is relatively new. We have a few years, but, uh, it's relatively new.

42:35 Yeah, I think when we talk about context, we have also talked about retrieval, augmented, generation, right, um, our agents, essential turning rack from receive information and answer into receive information decided act. Yeah, sort of, uh, uh, uh, rugged itself is not just built for, for A.I. And, uh, I think, uh, most of, uh, companies that are doing AI first, uh, the first thing I recommend is to create a good rag because create a rag is easy, create a good rag is, is, is, is other.

43:17 It's a little bit complicated because, uh, is, is not just, uh, pool information inside. You have to, uh, to verify this information. You have to cut the blocks of this, this information. For example, uh, big challenge we had this, this year was to, uh, take all lessons that we, we have on our platform, it transcribe all these lessons and for each, uh, phrase that, that we have, we create an, uh, an entry point inside our, our, our, so the user can search for a specific word and go to straight to that point at the, the video, uh, within that, that we have this, this day, so it can be, uh, very, very complex.

44:04 And at, at the other hand, it provides a lot of context, uh, to, to AI. So when you prompt something, uh, you, you can, you just can achieve a good response to, to our customer when you, you have data. Uh, the, the core, uh, behind AI is the data, the data you capture, the data you, you trade, the data, you, you pull, uh, into your, your, your existing, your, your database. Uh, this is the most important thing. And, uh, when you plug this together, uh, rag is the way that you, you have to, to provide this, this, this information.

44:44 So you have to, to make the right cuts on this data, uh, to, to, to, to do this context. We all know that we have a context swing though. So the context is, is limited to, to, to an amount of torque for, for example, and we, we also know that, uh, there's a sweet spot on the, the context is, uh, two thousand, uh, two hundred thousand, uh, tokens. For example, even the context, uh, the, the model have the capability to one medium tokens. We know that, uh, it starts to, to, to, to, to, to, uh, so, uh, we need to do the right cuts on, on our data to provide the, the, the right amount of data when, uh, interactive with, uh, rag.

45:34 So we pick up the, uh, for example, an entire article with, uh, two thousand words, but if you inject two thousand words to, to our context, uh, it, it can be bad instead of good because it's a lot of information that it doesn't need. So you have to chunk this, this data, put it into, uh, into a rag, for example, and when the user, uh, asked something, you can just go to your, your, your, your rag, pick just that paragraph and send it back to, yeah, so this is the importance of ragging. This is how it improves the, the, the, the context and the, uh, the, the answer to your customer.

46:19 I think one of the, uh, candidate rack have, it's, it's the memory in that coming, I don't know, new, new ideas like Vicky, LMM, I don't think he, I think that's the, what, what should develop us think about the memory topic? Yeah. The memory itself is, uh, is the trickiest part when you say, uh, the context, uh, we, we are talking, uh, to the, the, the, talking about the memory itself, uh, you need to understand, uh, a little bit about, uh, how AI works, for example, uh, because this is, uh, this is the, the fundamentals of, of AI and this can help you to gather better answers from AI.

47:12 For example, when you start a chat with Copilot, for example, and you say, Hey, Copilot, uh, do implement this for me. Okay. It will try to read your code base. It's right to read your code. But maybe, uh, a lot of code that is, that is your code base is meaningful for, for that action. So it will update its memory for that chat with a lot of, uh, a useful information. Instead of this, you can say, uh, Hey, Copilot, can you implement this code for me, base it on, and you can specify this fire, disfile this fire.

47:56 So you, you gave Copilot, uh, uh, a, a, a very, a very good context with examples, with simple and straightforward, uh, example, with good example, because I, I, I, I, up to your, your entire code basis is written on the, uh, at the same way and the same, uh, uh, uh, the same way. So there, there are a piece of code that you wrote with more passion than, than others. So you can point this to, to, to Copilot, for example, uh, and all of this information when you start a chat is going to a memory is going to, to the context itself and it have a limit.

48:39 So what's the average hat we do is to compact this conversation since you reach, uh, on a certain amount of, of tokens reach it there. And this kind of, the, the trick is part because when, uh, Copilot, uh, for example, compacted by itself, it will just do, uh, compact, it will say, Hey, uh, compact is for me or, uh, summarize this for me. That's fine. This is okay. It can work, but you can compact it to yourself. If the conversation is, it's getting a bit longer. You say, Hey, compact this conversation.

49:19 Uh, the main points here are these, these, these, and this. So you'll have a, a, a very different kind of, of memory update at that point because you point and then to, to, to, the, the important topics of, of this conversation. And another thing that I do is every, every time that I finish, uh, something that I say, I achieve that, that what I was looking for, uh, I close the conversation and start another conversation because it starts with a clean context that, that would be say, you can clear the context, but I prefer to start a new, a new context.

50:06 And, uh, regarding memory, uh, we can also have a, a, Markdown files, uh, with this, the summaries that I, I told you, inside our folders. So you, you, you can, uh, summarize our conversation with your specification and other things, place inside a, a folder. And every time that copilot reads that, that, uh, I'm saying copilot, but it's the same behavior for, uh, in AI that, that, that you use. So it, it will reach this, this Markdown file inside our folder and update its memory with, uh, with all the, this stuff. So, uh, this is cool, but you, you need to understand how, uh, uh, a lot of works, how the memory works because this is, is very important for every human being.

50:55 Not as, uh, I think in a, not a topic, when we talk, especially AI, uh, agents, a company, company is really, yeah, enthusiastic on the one part, but on the other part, there's the security part. And, yeah, um, And I also see it in a lot of companies, they say, okay, we make it and then, uh, we have to train it, uh, and, and then they give, I don't know, access to reach. Uh, rights, uh, the world exchange show us and so on. We have to train it and so on.

51:36 Um, what, what did you, you know, shoulder for me for the start, I think a little bit, we should give, uh, agents, the same principle of last privilege, like we do it for humans, or, uh, is it, uh, how should we think? Yeah. Uh, it, I'm a little bit concerned with security and we, we had so many leaks, cases, data leaks, cases in the last few years. So, uh, I tend to close all the doors and just open the door that you really need to.

52:22 And I think, uh, Azure AI founder is doing a very good job for that because you can, uh, simply plug the models, you can supply, uh, create an AI infrastructure for, for your, for your scenario. And you can pass by a, best use in the Android ID, for example, which is the Microsoft, uh, looking at the previos active director. So, uh, you can write the permissions, uh, on, on the portal is, is very easy. I think it is a very good approach to have a security layer, uh, in front of your AI models.

53:05 This is the, I think the main thing you, you need to do it, uh, the, the very first thing you need to do because we are not just concerned about the users outside, but the users inside our company, right? Because, uh, it can, can be massive. Someone can ask for this. Hey, give me the salary base from, from my colleagues and issue. You want to, you don't want to, to have this, this problem. Uh, and inside our applications, we also need to, to create guard, raise, uh, for, for that, for example, uh, I told you that we can expose some functions, uh, inside our system, right?

53:52 And that is true. And that is one, you, you can, with one line of code, expose, uh, uh, uh, a function, for example, to list our products, right? But, uh, let's say that is some user interact with that and say, hey, bring me all products, including, uh, the ones that were not launched yet. So you have a problem because if your function, uh, are exposing the products, which was not launched yet, you have, uh, a leak on your, on your code. So, in some, in some cases, we need to put some, some, uh, wide layers on this, this function is a, uh, instead of exporting this function, let's create a new function, which exposed just the products, uh, itself, without, uh, uh, release date, for example, uh, does, does the release it products and, uh, let's expose this function instead of the function that's supposed all the, the, the products, uh, for the customers.

54:57 So as we discussed previously here, uh, this is the item, the way. So every time that you ask for the products, you got, you get the products, there's a matter of which prompt you, you write, you won't get the launch and product. So that, that's the way we have to, to leave it. So I, I don't like to expose everything to, to AI. I prefer to create small functions for, uh, the, the money situations and to, uh, and expose these functions because I can control inside the code, which, uh, which can be done inside my, my code base. Right.

55:38 Uh, uh, so it's the same principle to, when you connect things because, uh, I know that it is, is, is very fun and, and very straightforward to, uh, to connect to your AI to your SQL database, for example. It's very powerful. Uh, you go, you connect to your SQL database and say, Hey, this is my database and, and now, uh, let's do some, some magic here. Bring me this information that, that's fine. That's, that, that's cool. If that's, if you, just, if you're working alone, for example, that's very fine.

56:19 But if you work with multiple people, uh, this is, uh, this is likely, uh, insane because you're given access to your entire database or to, for an entire table, uh, for people asking whatever they want. So they, they can ask if I want to build a better product for days and start a new component. How should I do? Okay. You have access to all the information. Data is the new oil. I think what, what, or the older topic, uh, a lot of companies talk about who was the, the topic prompt injection, but it's especially with MCPs, uh, the, the, the, the RAC pull attacks.

57:04 Uh, the new, new chapter, can you a little bit explain how, how to prevent and I think for the people, they don't understand what, what, what, what's the different. Wow. Nice. Uh, property injection is the way that we, since we have an entry point that we can, uh, talk to AI models, we can ask whatever we want. So there is no limitation for that. Uh, And then, uh, since we have no limitations, we are open to the people creativity. So, and they are very creative. Uh, they can ask everything about your system.

57:44 So if, if you don't, do not have any kind of grad radius for, for that, uh, it will expose data. It will even worth let people write data inside your, your system. So we have, uh, so many cases that people were using, uh, a chat box around the word to, to, to, uh, solve programming questions. For example, uh, instead of paying for a clause to, to ask, to ask my questions, I can go to a famous, uh, website out there, won't say the name because of a copy writing.

58:24 Uh, but I can go to the website, start a chat, a support chat and ask the same questions and how can I, uh, do this function in Python and the chat will, will answer. It's, it's a very simple thing. It's the same, but it, it's, uh, spend your tokens. Yeah. It's, it's, it's costing for, for your component to transfer that question. It, it's the, the, uh, prompt injection is a little bit like, uh, in the old times where I can write into, uh, that forms, uh, um, database code and it takes it, uh, and say, okay, uh, I'm, uh, I'm, uh, I'm, I'm, I'm, I'm, um, and, uh, I was on the database.

59:09 It's, it's all that, like, like this, but what, what, what is the, what's the right pull attack the people now? Yes, it's the same, the, the term that you're, you're saying is the SQL injection. We, we can, uh, write, uh, some, some prompt is, uh, this is the same. It's basically the same. So now we, we can, instead of injecting SQL, we can inject prompts and looks through the Trilett database, uh, ask questions about the things that is not, uh, in that context. So, uh, to, to prevent that, the, uh, I really recommend you to, to use something like, uh, Azure AI Foundry, for example, if you're doing a very, uh, complex, uh, or, um, AI first thing, uh, because it, it brings you, uh, a lot of Guadrass, do that.

1:00:06 Uh, when you, you create a, when you start interacting with a model, you can create some, uh, system prompts, which we, we call it some kind of some sort of memory that you have that you can say, for example, do not answer questions about and say some talk. Even, uh, when you customer, every time that customer asks, uh, about something that is not related to hamburgers or to some sort of products, do not answer this, this question. So you can specify this, uh, when I say, create a word array, it seems to be something very complex is not, it's just a pure text that you, you can write a sign or a model that you can, uh, bring sign or model.

1:00:57 But it's, it's hard to predict because, uh, it's, it's very hard to, to, to predict what users can, can think and can ask you to your model. So if you do not have anything, uh, prepared or anything, uh, how can I say, uh, trade it to, to, to, to that. Uh, you will, you end up having to write so many business rules that it's not, uh, it's not, it's painful. So, uh, I recommend you to, to, to bring, uh, something like Azure AI founder.

1:01:41 I think I said, but basically Azure here because it's them, uh, the two that I, that I work, but there are many tools outside, you don't have to, to be strong with, with Azure, for example. But you bring something, something closer to your, to your business, uh, with business rules, uh, or an inside it, uh, with blocking your context, for example, uh, to prevent that. So, um, we have all, all, all these, these cool things we can now do with, with AI, especially encoding. Um, but, but I think, um, yeah, we have this, the tokens and the tokens are actually about getting cheaper for my feelings.

1:02:27 Um, is it a show, developers, uh, also focused on architecture and the look, especially, I think fine ops can be find a sheep, a solution, or is, is it, uh, the time invest for the sheep, a solution. It makes it sense. So I, I don't know, I, I, I can use, uh, AI to, uh, uh, make me, uh, PF to document or show that use OCR because it's cheaper, like, something like this, as example, is it, what would you think?

1:03:09 Yeah, I think and I, uh, and I use here is, is a mix of things, uh, AI is, uh, is a, is a broad term and it's been here since 1996. So it is, it's nothing new, but what we have new here is the generative AI. So the words say itself is generative, is to generate. So every time that I need to generate something new, I delegate this to AI, this is clear for me. If I need to do a repat it, task, this is not AI for me. For example, uh, every video that I do for YouTube, I have to extract the audio to do a transcription of this, this audio, uh, generate the, uh, the SR, SRT files to upload to YouTube.

1:03:59 And translate this SRT to other languages, so I have the, uh, I have the, the translations, uh, on, on YouTube. So this is a process. This is a workflow. So I can send my video to Claude, for example, say, hey, do the transcription, uh, create the SRT portuguese, uh, anguish and Spanish for me. It will, it will do a very good job. It's, it's fine. It's pretty, it's fine. But it will, it will cost me a lot of tokens because it will do the same process that I do locally.

1:04:39 So every time that I've, that I get a, uh, a video file here, I run a Python script, which was created by Claude. And because I don't program it in Python, so I said, hey, create a, a Python script for me that extracted the audio for, for the video. That's fine. We extracted the, the, the, the, the wave file. Now, then I have another file that say, hey, subscribe this wave file and create an SRT. And then I use AI just to translate this SRT to, to, to, to, on other languages. So instead of using a lot of token to do all the process, uh, using AI, I just use a small token just to, to, to, to translate the, the, the SRT files and the rest of the process I can run directing on, on, on my machine.

1:05:30 I think, uh, having this, this mix, uh, it's very important because, uh, yeah, say, I can do everything, but doesn't mean you, you have to do everything using AI. Yeah, yeah, yeah, uh, I actually look for, uh, I like to build and voice depth tool for, for my YouTube videos. Yeah, that's fine. I have some good, good Python scripts with, uh, for example, uh, this podcast, we can, uh, cut all the, the silences from, from the video, we can analyze the, the, the audio layers and, uh, take off your, your language, uh, uh, accents, for example, everything like that. And this, this is a, uh, a, a systematic matter. This need to happen every time the same way.

1:06:26 So this, for me, is programming. If you need to generate something new, if you'll need to have, uh, different approach from, from something, this is AI, this is the generative AI. Uh, some, yeah, we're a little bit over time, but I, uh, if it's okay, I have a rapid fire around. No, that's fine. Yeah. Um, single agent or multi agent. Uh, multi agents, uh, for, uh, I think for the most of complex job, uh, multi agents with, so be agents are, are the way to have, um, uh, prompt engineer context engineering.

1:07:07 Uh, context engineering because it's way more complex and can deal with multiple, uh, stuff. Uh, loops are the future. Uh, in a way that, what is coming up for, yeah. Uh, you may know what's coming up from, from AI. Uh, I think we, we have more powerful agents, uh, more powerful models. And, uh, writing better code, writing better designs, for example, which one, something that I use in real, continue to evolve.

1:07:53 We are, uh, in the beginning of the generative AI. So we can expect much more if you're impressive with AI nowadays, uh, in the next few years, you get more, uh, double time more impressive. Yeah, uh, under, if satan and dahlia call you and say you get all the money and resources to make dot net better, what feature will you develop? Uh, feature, uh, I don't know because, uh, the dot net is, uh, is a very good framework in my opinion that the team is, is doing a very good job. And we are increasing performance, increasing security. It's a very major, uh, for a framework.

1:08:39 Uh, we don't have, uh, if you look to other frameworks, you'll see a lot of issues with security, for example, in that fact, we basically, doesn't have this, this kind of, uh, of things. So, uh, security and speed is, is always good. So I will invest in marketing to, to show the world how good is dot net. So I will pick a lot of money and say, hey, I want to, I want the people to know about that. And, uh, know how good, uh, stuff can be done in dot net. So let's reflect on some products that we have, uh, right, rewrite it in dot net using blazer using all the technology that we have.

1:09:27 And, uh, show the word how good it is. Okay, and then my last rapid-fired route question is, uh, if I showed invite, uh, who, who shall I invite as next guest and what, uh, question should I ask? Wow. Uh, programmer or not programming for business? Uh, well, I'm very excited to, to, uh, programming. And, uh, one person that I, I really into the, the videos, the other is Mac, it just, uh, he created some good skills.

1:10:09 And, uh, a, uh, an AI engineering with high level skills, uh, he talked, a lot of, uh, a, AI. So this is the, the, the person that I most watch the videos, not a. Yeah. Then, uh, yeah. And then, uh, yeah. And then, uh, thank you for, for joining me. Um, I'm talking, um, taking us from traditional software, development, all the way into AI agents, Microsoft, agent, product, world dot net, and the future of, uh, and gender software. Uh, yeah. One thing is, is, from years clear, AI isn't simply another library developer, um, add to an application. We are moving from where developers explicitly define every step forward, we are developed increased, we define goals, tools, context, emissions, and boundaries.

1:11:00 AI deterministic, how, yeah, to accomplish the past, um, that change how we build software, how we test software, how we secure it and the potential. The, uh, why being a software development means for, for the first place. So, yeah, um, I say to all the guests, if you enjoy the conversion, follow the MC65 podcast, you find all the info, uh, under it in show notes. And, yeah, um, thank you for, for, for being here when you're one hour. Thank you very much, Verica was a pleasure for me. And congratulations for your.