Stop Building AI Wrappers Build Agents | Google IO Connect India

[00:00:00]AI is here to stay, right? So please accept that firstly, right? If that is if that is a question in your mind, go ahead ask whatever you want to the AI.

[00:00:06]It can be a dumb question. It doesn't matter. But let's say if you want to start a new project today, irrespective of a programming language, what will be your guess the AI language?

[00:00:16]>> Please do use AI.

[00:00:19]>> Welcome back aliens. My name is Davin Reddi and today I am at Google IO connect which is happening in Bangalore and I got this opportunity to have a podcast with one of the speaker which I wanted to talk to from a long time and yesterday when we were discussing I thought podcast will be a good opportunity. So we got Prashant here.

[00:00:37]Thank you so much Brashant for coming on the channel in the Google Google studio.

[00:00:42]Uh so would you like to introduce yourself first and then we'll start.

[00:00:44]>> Absolutely. Thank you Navin. Thank you for having me here. It's my pleasure to be here talking to you. Um my name is Prashant. I lead developer adoption for APAC at Google cloud and um what does developer adoption mean right so we build a whole bunch of products at Google cloud there is AI there is Google cloud there are a whole bunch of services products platforms it's in my interest to ensure that developers are able to adopt and use this in the fastest way possible and understanding where to use it so that's my role that's what I do I've been doing this particular role for about 5 years been in the industry for a a lot longer than that.

[00:01:21]>> Oh, that's awesome. The way you said said about it, it feels like you love your job.

[00:01:26]>> Absolutely. I do. I do. Yeah.

[00:01:27]>> Perfect. Now, based on this, since my audience is more of developers, architects, since they are following from a long time, uh they have a lot of questions which I get in the comment sections and I thought this is a best opportunity for me to take those questions from the comment section and shoot to you. I will start with the first one which is when you talk about MCP you know two years ago everyone started talking about MCP and then they thought it's a new tool but when you play with it it feels like you're talking to REST API so how do you differentiate what's the mindset needed to differentiate between MCP and a rest API >> awesome great question um I get this a lot and even I used to think about this the same way when we started off right what is the difference between MCP and and APIs uh the critical thing is Right?

[00:02:15]Everything at the bottom is a functionality, right? It's a service.

[00:02:20]The way you expose a service to different kind of entities is where the differentiation comes. So when we talk about APIs, what is it we're talking about? We are talking about REST APIs, right?

[00:02:31]>> REST APIs has a specific structure. You have those resources, you have the different verbs that you use. So that is what are REST APIs. Essentially, even a REST API is a wrapper on top of a service. MCP is the way that the models can understand right so end of it don't rewrite your whole service or whole functionality as an MCP right MCP is a protocol that allows LLM's models to talk to these services in a very efficient way the difference here is this traditional APIs um it's you make a call you make a get call you make a post call you get back a response something is done on the server side lns are by nature very chatty right? It keeps talking back and forth especially nowadays with a long running agents as we call it. There's multiple conversations that goes back and forth that has to be done efficiently without the overhead of passing information back and forth again and again and so on. So while rest is a stateless protocol MCP started off and I'm saying started off MCP started off as a stateful protocol specifically designed for these LM interactions. MCP has taken off and is being is evolving a lot. I can talk for length about MCP but this is why MCP is different as a protocol right it serves a very specific purpose >> right uh now since you talked about wrapper in between and then it struck me the thing is when I started my journey with AI and then I was talking to one of my friend that I'm getting into AI now and he was like what you're doing in AI maybe you're writing your own model I was said no I'm going to work as a you know build AI solutions and he was like okay you're going to build another vapor so because that's how things started right so We got LMS and then people started building rappers around it. But I see now things are changing very fast with the help of agentic world. Now since I was working with ADK, I was able to build a multiple agents, you know, working them together, they getting data from one agent, passing it to this specialized agent and things are working amazingly. But when you look at the user perspective, they are still with UI. And when I imagine things, it's more about you know in future if I just want to book a flight I I will not go to make my trip and book a flight or maybe Google flights. I'll just say hey uh Gemini book a flight for me. How do you see this transition happening in the industry now?

[00:04:50]>> Okay. Uh great question right? I think it's important to understand two parts of it right. One is how do users want to interact with these systems? That is one part right. Second thing is what can you do? What kind of experiences can you deliver? If you use some of these newer technologies, agents to provide the same kind of an experience, then that is where it comes starts being thought of as a wrapper, right? But if you extend the functionality, you build on top of it, then it is a full-fledged application with its own value with its own business logic. So the way we are looking at at models as at agents is we don't want it to we we have gone past this time of you know chatbased one turn interfaces we want to achieve end goals right like you said I want to book a flight ticket right now it's booking a flight ticket is straightforward right again you're being very specific here right what if you say I have a business meeting let's say in Hyderabad right and I want to go there uh so in on a pleasant day. That is my goal. That is my intent, right? Um and this is something where now the agent can help you. It has to go look at a calendar, find your availability, find the correct flight timings, maybe look at some weather patterns where it's not raining so that you are able to set up your meetings at the right kind of a slot. Right? So we are looking at systems that is more goal oriented and this is the kind of way that we have to think about it while building these systems. Don't build transactional systems using AI. AI as an as a primitive or as an entity is can do a lot more than that. Right? So I think yeah as you said a mindset change has to happen where people need to look at it as goals as how do you how do you achieve what the user wants without the user having to tell you specific steps of what what's to be done.

[00:06:35]>> Perfect. Now since we're talking about agents uh in our company as well we build education content or educational products basically and we're trying to use agents wherever we need. Now at this point we were we are working on few products where we are using Gemini and then we are building agents with ADK but one of the thing which I got to know and in fact I was exploring was Gemma 4.

[00:06:58]>> Okay.

[00:06:59]>> And then I made a video about it. I loved it and then I was exploring different types of this models. We got Gemma 4 E4B and then we also have E2B and you can run A2B on mobile phone. And then I thought okay if you can run a local model which is a open weight models on your mobile phone I'm just thinking what are the things you can achieve with this thing and then I thought okay let me build the entire solution which one of the product which you have using local models and again that's an experiment because I'm not sure if local models can really achieve what frontier models can give. What are your thoughts because in the open source world at this point in in the traditional applications open source is doing very good. What about the LLM models?

[00:07:44]>> Yeah, absolutely. Great question. Two things, right? Don't be shy of using local models. Local models can do a lot of things that you may not have imagined to do. Don't expect local models to be at the same level as some of the hosted models, right? So, it is important to understand what it can and cannot do.

[00:08:04]Now, let's start with Gemma 4. Gemma 4 as a model uses the same architecture as Gemini 3, right? It's a it's so we use a process of distillation and GE gemini 3 some of the capabilities have been distilled into Gemma 4 as a model. So there are two things that is differentiating or that are part of a local model right uh or an open weight model. One is the capability of the model. What can it do? The second I would say is the intelligence of the model. What does a model know? Now local models like you said the E2B and E4B right? So these are effectively 2 billion or two or 4 billion parameters.

[00:08:38]These are small models. They do not have the world knowledge that the Gemini, the larger 3.1 Pro or the 3.5 flash models have, right? It does not have it. So don't expect it to do that. But in terms of capabilities, what are capabilities?

[00:08:52]We want the model to have reasoning and thinking capabilities. We want the model to be able to do tool calling, right?

[00:09:00]Given my use case, which MCPS uh that I call, right, or which uh functions that I call, these are the capabilities of the model. Gemma 4 retains all the capabilities of the Gemini model but just not the world knowledge right so it's great now that we have these models that are small enough that we can run even a mobile device uh we can make it to a lot of things so what is the what think about what kind of use cases you're building out if you are building a use case where you're summarizing a piece of text you don't need world knowledge right you need understanding of texts a local model can do that really well a local model is better because it is it's a smaller model it running locally doesn't have to do network ops. It's much faster, right? So you don't always need to go to a cloud model. If you're able to access a local model, you can definitely use that, right? So these are the ways that you have to think about various use cases.

[00:09:51]Uh use cases where you want to use a model to do a rag application. Again, a local model can do this, right? Because if you have your PDFs and all your documents sitting locally, your open weight models that you're running locally can access that and give you the response, right? You don't need to go outside. But then yes very complicated use cases compute inensive use cases deep reasoning you may want to switch to a cloud model in those cases >> right maybe when I was learning this uh maybe I was learning in a good time when tools were not changing that fast uh I'm talking about two to three years back but now if you look at the launches especially today we got a lot of announcements and I was talking to few of the developers in the community hall and they were a bit scared for The reason they're working for a company and not a startup so they are working for enterprises maybe a bank and uh when when they were asking question you know how AI will take away their jobs because a lot of things are happening and I I just said upgrade yourself you know work on the AI products and they're like we can't do that because in our company there's a limitation you can't just uh introduce a new service and say I'm going to do do this there's a protocol which you have to follow one of the sgest which I gave is maybe you can work work on the side projects. So let's say if someone wants to get into this AI world where they want to be future ready maybe they go for the new interview or they maybe they looking for a new opportunity. So let's say if someone is just a developer now working as a traditional developer what should be the pathway how should they start and what should be the goal because a lot of people they want to start but they don't know the goal. So just keep learning keep learning keep learning never achieving there. So what should be the starting point path and the goal?

[00:11:35]>> Awesome. Um I think you gave a very good use case or a very good direction already, right? So working on some side projects and so on. What I like doing is so we have gone past this time where you know you learn spend 6 months learning and then implement the world is changing so fast. By the time you finish learning there's a new technology out there, right? You're going to only keep learning learning or catching up to tech. What I like doing by myself is solving my own problems. Okay, let's break this down. So, first thing, access, right? Please do use AI, right?

[00:12:10]Try to find use cases where you can use AI. Understand what AI can do and cannot do. That's really, really important.

[00:12:16]Right? So, yes, if your company, your organization doesn't give you access, try to get access on the personal side, right? When you go back home, you have enough time at home, right? 6 hours at home hopefully or more. All right? So you know you you get your night of course taking sleepers out of it. Um spend that time leveraging AI for your personal reasons right. I use AI for learning. Uh if I want to distill a knowledge a document I use AI to you know ask pointed questions. Okay this sounds interesting. Give me more information. Find me related documents.

[00:12:46]Right? So I use a whole bunch of these AI capabilities. First thing is this.

[00:12:49]Get your hands on AI. Right? It is non-negotiable. You have to use it. AI is here to stay. Right? So please accept that firstly right? If that is if that is a question in your mind. Uh so that is it coming to access right uh get access to it. Now the question of uh you know um upskilling or leveraging AI or how do you how do you how do you stay relevant have this loop where you learn not learn something learn about something try it out right so like I said come up with your own personal use cases like for example like I said my own use cases let's say uh learning languages I love learning languages okay um so initially what I did was I had these models where I would build an application that would teach me a few words uh right in a different language that I want to And then I would record me saying it.

[00:13:34]The model like Gemini is a multimodal model. It can understand various languages. It would tell me if I'm repeating it, if I'm doing it the right way, right? Um, so I built this while building it, I understood how to use a frameworks. I understood what the model can do. I didn't spend one month understanding the framework, right? I started using it and then learned as I went along. Now, Gemini has a 3.5 uh live model which does live translation, right? So how easy is this for me to learn a language if I use that API and directly interact by voice with the AI and it responds back to me in the language that I want to learn. Right?

[00:14:08]Again I don't want to spend my time learning that. I use it in a project that means something to me. So yes, please pick a personal project. Uh pick something that matters to you, right?

[00:14:20]that you have a strong feeling for that you're attached to and that will give you the motivation and drive to actually implement something and you learn while doing that. So I I would recommend somebody to go in that way.

[00:14:32]>> Perfect. And that's what I'm doing now.

[00:14:35]So whatever AI I'm learning is not a structured approach. It's more about let's build something. What are the problems I facing? Just building it and then on in the process I'm learning okay this is something you can do extra. uh sometime when you talk to the Gemini and then you say this is my project I have built till this point I just say no roast me or grill the project absolutely >> and it gives lot of different perspective what is going wrong uh what are the things you are missing in the project which will make it more helpful and I love to do that so that's perfect >> see in fact what you said right the way you're asking the air to grill you or ask you those questions air doesn't hold any emotions or feelings right it's not like a human tutor where you say what will they think about me if I ask a question like this you know go ahead ask whatever you want to the AI it can be a dumb question it doesn't matter right the AI is there to help you and you are bringing the AI in a level of capability that you want you can use it to train yourself or you know upskill yourself >> right perfect lot of lot of insights Brashant and uh I will not let you go before we have a rapid fire >> oh sure why not >> yeah so I have few questions prepared let's see if that hits right so name one Google cloud service or a AI tool which you have used yesterday while building something or learning something.

[00:15:54]>> I use anti-gravity on a daily basis. So if there's one tool that I've used, it's definitely anti-gravity.

[00:16:01]>> Okay. Which one? The ID1 or CLI?

[00:16:03]>> CLI. I mean the So it's it's like this, right? I mean I IDE or the UI makes it a lot more easier, but CLI gives me that, you know, the traditional developer command line kind of a feel. So my preference is a CLI but essentially both are equally good.

[00:16:18]>> So CLI feels like you you are in power.

[00:16:20]>> Exactly. Yeah. Yeah.

[00:16:22]>> Perfect. Next one. So let's say at this point language doesn't matter anyway. So you can you are working with maybe Java, Python, JavaScript but let's say if you want to start a new project today irrespective of a programming language and if you ask AI to build a project for you irrespect of the language what will be your guess the AI will serve the language. So this is something I've always been telling people. Stop thinking about language, right? Let the AI do whatever it feels like. And since this is interesting, right? Different AIs may have different language preferences. I want the AI to build my functionality, right? So I will tell the AI give it my use case and tell the AI build an application that solves my use case. I am not concerned with the language. That said, most AIs today end up using React for the front end and Python is the preferred language for the back end. So that would be the go to language for the AIS that I have seen from experience but it's not something I try to concern myself with.

[00:17:16]>> Perfect. In fact what happened is uh I was working on one of the project where I was recording I was building a recording application which will record my face and the screen everything and I started with anti-gravity as a UI and then I moved that project to IDA and then I realized I was thinking it would be a Python project. It was a rush project.

[00:17:36]>> Oh wow. So it understood the use case what exactly is needed since we have to work on the hardware level now let's use rust and I loved it >> awesome >> I realized that after working for 2 days because I was full into UI I was not checking the codebase I wanted to do that without checking the code base but I open the code base I realized it is rust >> got it >> that was fun >> so in fact one very interesting thing that I'll I'll plug in here is um so at Google we have created a whole bunch of skills so you can find these skills it's a github repository all the skills are available there there is one skill called the well architected framework.

[00:18:08]>> Well architected framework.

[00:18:09]>> Yes. Right. So these are the modern architectural best practices and guidelines. Now if you use that skill whenever you give any problem statement what the model will do is look up that knowledge base decide which framework language technology is a best suited for your use case and implement in that use case. Right? So this is relevant information. Tomorrow today it's Rust maybe tomorrow there's a Rust++. Right?

[00:18:32]And as long as architecture framework is updated, it will be able to leverage that and then you don't you're not relying on the model's knowledge but you have the you know a resource that tells the model how to do things.

[00:18:42]>> Right. Perfect. Now with that one last question. What's your take? Maybe you have let's say 20 seconds to answer this. What's your take on hardness engineering?

[00:18:52]>> We have a lot of things right? We have context engineering, hardness engineering, loop engineering, a lot of things, right? Essentially all of this is how do you ensure that the right amount of information is given to the model. How do you let the model take its own time to implement your use case and use use code? So harness engineering is really really important. What I'm saying is most of the frameworks and technologies will eventually incorporate it. So it's important for you to understand how how the systems work right building out uh having these longunning agents and running in a loop.

[00:19:18]>> Right. Perfect. Okay. So on that note, thank you so much uh Prashant for giving your time. I know we are the it's almost end of the day but uh you have shown a lot of energy. I thought we'll be chilling here and uh but then topic is interesting and I love your energy.

[00:19:34]>> No, thank you so much. I think it's it's definitely the end of the day but then it's good to end it on this kind of a note. So great conversation. Thank you so much.

[00:19:41]>> Thank you so much once again for coming on the Disco channel. Yeah, >> thank you.

[00:19:44]>> Thank you and keep watching. Let's wait for the next video. Bye-bye.