Supercharging Development with Kiro - Technical Track - Episode 06

Hey. Hi Ashish. Um, thank you for joining me here. I'm very curious to know about AIdriven software development. Uh, it's a buzzword everywhere and I think we are a little uh, you know, outdated in our ways of doing things in our so maybe you can you know guide us. — Sure. I I'll try my best. Again, it's a very moving target in a way because KO is developing so fast that I do not know everything. But I'll try my best to explain you these things. So let's go ahead and talk about that. So I put it like this. This is not a marketing but it is reality. Like if you can think of it, you can build it. That's how the ko operates. So you think what you need, you give some ideas. Ko enhances those ideas and then you can start building those solutions if needed. It could be simple apps, it could be complicated upgrade on existing ones or it could be just starting everything from scratch. So we'll talk about all those things as we move forward. Right now you may have a confusion like how this Ko fits in or where are all these components fit in because we talked about lots of things. We talked about MCPS, we talked about bedrock, we talked about agent core. So I want to give you a highle picture first that where ko and all the software fits in and then we can go into the details of Ko. Okay, if I say Amazon Ko, just assume it is KO. We do not use a word Amazon Ko, it is just considered KO because it's a quite independent product. Even if you are not using any of the AWS services or don't plan to, you can still utilize KO. It work completely independently. So that's why we just call it KO, not Amazon Ko, just KO. Okay, — got it. — Good. So probably you know that how we get started. So for getting started with AWS or sorry AI what you need that you need infrastructure obviously you need the place where your code would run whatever creation you have created would deploy and run so that's why you would be needing infrastructure and when we talk about infrastructure one thing you can follow is like you can have your own compute like AI compute which I mentioned on the right side where we have chips like AWS trrenium infia GPUs which can help you into getting started or building your own stack but in case you need to have like little more managed approach that time you can go for Amazon SageMaker AI which gives you all the resources like model building resources for training you can run MLOSS pipelines using that you can perform fine-tuning of existing model or you can perform deployment and also ensure governance so that is where the infrastructure layer is right and once your infrastructure is sorted out you could then think of okay now I have to start with my AI agent development what software and services are available is there for me. So let's talk about those concept that what we would be needing in this space. So bedrock obviously plays a foundational role here. In bedrock we already discussed about the different models availability, what capabilities it has. We also discussed in previous sessions about agent core and knowledge basis. So bedrock gives you a solid foundation to get started but it is not the only way because obviously everyone has different use cases and that's why we want to give you additional context in the sense that hey I could go ahead and have integration with Nova act for SDKs or agent or I can have open system software like strands agent I don't want to get into proprietary things so I could utilize S oss or open system software which are very flexible like strands agent so we can integrate with that and once you have this sorted you could get that started with your applications like you could use Ko could use Amazon Q as transform connect and there are a lot of options available into market now I put them into a stack that doesn't mean that to use Ko you need to have SDKs or bedrock or infrastructure no these application can be used independently so I can just started with get started with Amazon Q or I may just use transform Ko it purely depends on how you want to utilize but these are the various options which are available for you to get started. Does it make sense? — Got it. Yeah, makes sense. So like Ashisha said we are a little behind in this a IML uh trend. So we are still using EC2s to deploy our applications. So I was wondering I didn't see EC2 anywhere over here. So can I still use an EC2 to deploy my AI applications? So some of the powerful ones that AWS has to offer can I do that? — Absolutely you can do that. But if it comes to optimization then it would be preferring that we go ahead with these AI compute options we offer. So yes in a way it is EC2 which has compute power but when you are doing this AML training or inferences which are running probably having a GPU based instances or instances which are optimized for running AI workload will be more suitable. They will give you best price performance ratio. So yes, EC2 can be leveraged but you have better options where you can get better performance, lower cost, more stability, more functionality. So that's why EC2s are there but we recommend that you go ahead with the compute option we have provided here. That's what is the

— Got it. Makes sense. Right. — So I so the stack that you've shown me is great but you know um I would just like to understand what has the evolution been like we had those typical AI coding assistants which used to help us and some of our developers in our organization who are early adopters they have themselves used help of AI but we still haven't caught on. So I would like to know you know how far behind are we or what stage has these AI coding assistants reached. Can you help me with that? Yeah, why not? And don't feel like you are far behind like every customer's journey is different, right? Somebody may still be using cobalt that doesn't mean they are far behind, right? They that's that technology serve their purpose. So probably that's why they are using it. So it's not far behind. Maybe we can talk about like how this evolution happened and then we can move started. And one more thing which I want to mention here that with this AI stack, we also give you integration with MCP. We talked already about it. A2A options are there for protocol and obviously all of these can integrate with your data. It could be in S3 or it can be databases or graph databases or in on-prim environment we could have integration on that. So how AI is changing software? Let me talk about that. So in 2023 around we found a lot of autocomplete like a glorified intellisense built into the IDE. ID stand for integrated development environment uh interface where I go and write my code like visual studio or it could be anything for Python or anything Eclipse for Java something like that. So those ids have started giving a feature of autocomplete. So that helped developers in writing the code faster like so AI was providing general help to you and autocomplete allowed you to complete those sentences or those syntax much faster. Obviously that evolved as the technology evolved. So what we have now that is our 2024 we came up with the idea of assistant means it's the general thing I'm not talking about AWS specific here. So generally hey I want to get a change done into this function. I want to validate this particular log file which I have. I want to troubleshoot that error. So it become more interactive where we could talk to this AI as assistant and they would be able to help us out and that has started the way for giving things like okay so we have now agent who can help us in completing the end toend task not just focused on one single task but whole life cycle of SDLC or a big project's maintenance can be automated through the use of agent. So we started with a basic autocomplete then we started using AI as assistant and now we are going into the world where full-fledged AI development is happening and agents are helping in shaping how we actually program the thing right — great I get the picture now where we've reached where we were and where we've reached so can you give me like an overview of what are the tangible benefits that I can get by following this AIdriven software development life cycle — makes sense yeah obviously it is always important to see that hey what why we are using this the why is always important right so let's talk — and not because everybody's following it I would really like to know what are the benefits — so let me try to zoom in a little bit because there was some question so the promise of agentic development is like what let's talk about that so idea is that we want to give autonomy to our developers right so agent can take decision on their own and complete the increasingly challenging task so we want our developer to focus on the more business side of thing rather than figuring out the syntax or troubleshooting and all. So we want to use agent so that we can get more autonomy for it and that helps us in performing two collaboration. So agentic development can then help developer and agent they work together they're not competing against each other like lot of time I hear that okay because of agentic development no developers are needed that's not true we do need to have agents also to facilitate the heavy lifting which is needed plus also have a option to that developer can also come up with their own creative ways of things. So we need autonomy, we need true collaboration and we also need what we also need to have a higher quality of the code. We don't want AI to generate junk of code and then my developer is spending time to just figure out hey what is right, what is wrong, what should be optimized. We don't want like a junk of code generated. We want to have a higher quality also apprec created by it. So that's how this is what agentic development we are looking forward to. But yeah, obviously it is what we want but it may not be that where we are right now. But this is what the promise of agentic development is that once you start following it, you would have autonomy. You would have true collaboration between a agent and a human and you would have higher quality code coming up which will be optimized for performance. It would not have security vulnerability and it would start working as expected. — Nice. So uh while I understand all these three points are really great benefits for software engineers and developers as such but how does it benefit our

business or how does it bring business value? Uh could you highlight about that too? — Yeah. So that is important because see technology is good if it is there but if it is not solving a business problem it is of no use right. So obviously businesses are also getting benefit. Let's talk in terms of faster time to market. Right? You may have idea let's say you wanted to build something called MVP like starting point for a product that may still take a lot of time but once you start using this kind of agentic development you could start creating your PC's much faster your MVP products can be really created faster so businesses can iterate faster find out what is working and move forward if it is not let's try another idea so faster time to market for business and also it helps in productivity gains so agent will help your developer in performing all the things like cost validation, maybe security vulnerability assessment, maybe performance fine-tuning. So that will obviously drill down the cost and it will also cut down the time they may have spent on the coding before. So for businesses it is faster time to market and productivity gain. These are the two prominent thing which may happen. — Great. So uh while we are you know there's a huge ecosystem of this AI aided uh assistance and ids. So when we are selecting and choosing them uh are there any sort of challenges that I should be aware of beforehand what I might encounter. — Yeah mean see we want technology to solve all our problem but again having technology doesn't mean that everything is solved automatedly. So we may still have some challenges to face right. So when you're talking about let's say AIdriven development what challenges can be there. So let's say you have started coding through your AI assistant and you started writing a code. Now it could be good for one person but what if you have to scale it to 200 developer team or there are 500 people working in different time zone on different projects or different features of the product then how you will scale that AI development. So that becomes a bottom line because we do not have a centralized way to get all the knowledge stored into AI and then utilize it. So that is one problem which are being faced by the customers right now. Also there is limited control on that. What I mean by this that let's say you have asked your program that or your AI assistant that hey write me a code for checkers tic tac toe that fine it will write a code and give it to you but you are not seeing that what decision point it has gone through what type of validation it has performed so we see AI as a black box where we send a input it generates a output and then we start figuring out whether it is output is aligned to my requirement or not do I need to fix it modify it do I to enhance it. So we do not have lot of control on how those outputs are generated for me. And obviously when we do not have control then code quality cannot be guaranteed because GI may have generated based on decision point A, B and C whereas for your business those points were not relevant. You wanted it to decide based on X Y and Z. So that may always happen when you are using this AI based tool. So these are some challenges which we want to address through Amazon through KO. We'll talk more on that but yeah these are the common challenges which any AI tool would face like scaling that and then limited control on what is being generated and ensuring that quality of the code is proper. Does it make sense? — Got it. Yeah. Makes sense. So I had a question. So you mentioned over here that um you know in using our existing uh AI tools that we have currently we can have limited control or we can you know do some really small task really well to create PC's but uh they may not really be great at creating end toend enterprise uh AI applications. So I was just wondering I can do all of this uh using VIP coding right why do I need like a very sophisticated autonomous agentic AI IDE to do software planning and development if VIP coding can help with that why do I need such a — yeah it can help how much that is questionable I'm not saying wipe coding is bad or I would say that wipe coding is the word which has taken the whole year like the most prominent word of the year if I Say it would be W coding. Everyone is talking about it. It is good. It has given a lot of people ways to start developing or start creating solution which they never thought about. But W coding has its own limitation and where it fits in is different. So W coding concept for people who are not aware that you code first and think later and adjust as you go. So like some of us may want to plan ahead and then start with the travel. Let's say I am proper planner. When I have to travel I have to plan. Whereas some people may be hey let's pick up a bag. we'll move on and we'll figure out on the way what needs to be done. So V coding is similar to that. What happens in V coding that your developer writes a prompt AI generates a code. Now that code may be optimized may not be optimized. Developer again gives a prompt again generation happen. So we keep on prompting we keep on optimizing and at the end we get the required

thing. Right? This is what happens into a w coding format. But idea here is that let's code first we'll think it later and then we will keep on adjusting as we go. This is what a w coding observing is. — Okay. So this looks like more of a trial and error kind of mechanism. So uh do should I really use VIP coding you think or looks like I should not be considering that at all. So how do I decide when should I VIP code when should I not? — That's a great question. So let me give an analogy so that you can I can explain you in a different way that how we can leverage VIP coding if required. Right? So we all have some jobs in our house to do like last week I had to get my garage painted and for that I would need a handyman who could come and fix those garage thing like paint garage then I just stay for that and done but if I'm building a house that is a different story. So wipe coding is like which works fast. It relies on experience and intuition but it may not be doing a lot of forward thinking for it like okay this paint is not mold proof. Should I use a moldproof paint into a garage door or not? Or maybe should I use some wall filling first and then do the painting. So wipe coding is like hey I would improvise and avoid and dot thing as I keep on going on that. So it's like a handyman which is good for quick fixes very flexible. that you can ask him to do different type of work and great for small problems also but the risk is they are not always scalable. I would not trust a handyman to build a house for me completely or it may fix something but it may not take like I had experience like I asked the handyman to drill something into wall it didn't monitor properly and there was electric line behind which got drilled because of that. So those type of things can happen but they have their own use cases ways of doing things and that may fit in some cases but it may not fit into a huge enterprise case environment type of thing. Does it make sense? Got it. — Yeah. Yeah, it definitely. So, can you give me some examples when I should go for vibe or not to vibe? Yeah. — Yeah. To vibe, right? To be or not to be kind of a thing. So, let's say you're looking for a quick UI additions like you have already existing application. Hey, I want to add a confirm button. I want to do a simple fetcher enhancement. This is a bug. This is the error. I want to fix it. I want to refactor my code structure into a single file. So those time w code is really good for but if you are doing some complex multifile feature and architectural changes which can fold or which can unfold multiple files or which can touch multiple files or feature where you require need to have proper requirement and then you go ahead with design document that time your w coding may not work best in that cases or in the case like think of if your system is going to affect multiple system boundary then wipe coding may not be the best thing for it like I give you another example so in my garage I have to get my electricity sorted and in UK the electricity systems are not like hey you put a plug and then you get started. I have to go through council to get a permission to block those roads and then go for a permission from the health from the water department and all. So obviously a handyman is not able to go ahead and get these things done for me. What I would need an agency who knows what type of permissions are required and then that can help me out in that cases. So wipe code is good but it may not be able to solve all the problems you are looking for if it is an enterprise use case or if it is having a complex structure to handle in that case. — I get it. Yeah. So now looking at the bigger picture like I understand coding you can do quick fixes like you know some small implementation and changes and now looking at the bigger picture uh how is AIdriven software development different from the traditional software development that we were doing? So we've been using that till now. So is it very different? Uh how does Ko come into the picture? Uh so I just wanted to know how they are different from each other. Where do the different components fall in? — Yeah. So if you look about wipe coding, we do not talk about anything requirement or design. We just say okay give a prompt let's get started. But if I look at traditional SDLC's, traditional SDLC's work in this format. So we plan we first get a requirement then we do a design documentation then implementation happen then proper testing and QA phases are there deployment happens and then maintenance happens on that right and you may follow waterfall model or you may follow agile or maybe any other methodology but the concept is same that you start with your requirement documented cemented and then you move forward from that. So that is not going to change when we talk about koddriven development or what we refer to something called specdriven development. We have a new approach to follow. So we'll talk about that. So ko introduces a new approach for building software with AI agent. But we still follow those standards of having requirement finalized having design finalized and then start building rather than just jumping and say hey I want to build something that is not going to work. So mostly we should be able to

sort these things first and then we should go ahead and start building our applications. — Right. If you want a long-term maintainable software, we should go right from the beginning. Make sense? Yeah. — Exactly. That's the key. So yeah, again think of like what is your long-term vision? Are you looking for just a quick bug fix or you are looking to build the next Amazon. com? Depending on that you would use what resources are useful for you and what you should do. So if I give you a little bit more idea on that how we should get started with this new approach here it says that let's start with something called spec. Spec is specification. So we get a specification and we'll see how it works out and how it looks into UI concept. But specification is basically what is your starting point and that specification get recorded into a file which is called your requirement MD file. So this requirement MD file is generated. All your specification gets into it and you don't have to manually create the specification. You give a prompt to a Ko assistant and Ko would be able to generate those requirement for M. Now you do not okay go ahead and say okay like let's consider a situation. Let's say you have to build a software. So you do not go into a room and say hey these are the four bullet point my software should solve and then I walk out and then developer in six months will deliver the solution. No, you have to document your requirement. You have to design your document. You have to create proper designs on that. So same thing happens here. If you are happy with the requirement generated go ahead. If not make those changes run it through those again then design will be created with the design. If you are happy then you move forward. Then you go ahead with the implementation phase. This is how your based approach would look like. You start with specification you create requirement. You then design and then you start building your task based on the design you have created. Obviously there is feedback mechanism like okay this specific task has to be changed or this requirement is need to be verified or validated or maybe changed that we can always integrate in the system shouldn't be a problem at all — okay so what exactly do you mean by specification is it like guidelines instructions which are there in these I see that these are markdown files it's got MD so are these like instructions or guidelines what sort of things are there in the specs Can you give me a little — depends on what you are doing. So let's say I'm fixing a bug in existing system. So my requirement that time would be different, right? It would be based on what the code base already exist, what errors are there, what standards are in place. If I'm building something new, my requirement MD may look differently. But idea of speci development is that we want to follow traditional HDLC practice to ship the product. So wipe coding ship the prototype that is good for prototyping purposes. But when you're looking for a whole system to be developed probably you will go with these fail. So V coding is not being eliminated here. It is still there. But idea is that we are using it in a more structured way. So when we talk about specdriven development, it is not just producing the code but it is also about okay testing and QA needs to be done. Okay, I want to now deploy this code into AWS or any other provider and then I may have to start with the maintenance on top of it also. So that is how this whole cycle fits into the specdriven development not just the code generation part which why coding mostly focuses on. Does it make sense? — Yes, makes sense. So probably you could give me some examples or illustrations to uh you know give me an idea how does this uh the decision making part happen in this whole cycle. — Yep. So let let's talk about what currently happens. So let's say you have a prompt right? Okay. I want to build an application. a fitness application. I give this prompt to my AI agent or AI software whatever it is. So for that it has to go through different decisions, right? It has to figure out what would be would it be a three- tier app? web application? Would it be a mobile application? Will I be using Java for front end? Will I use NodeJS? What would be my database type? Where it would be hosted? Would I need serverless? Would I go with the server based approach? So what you see here that lot of decisions are happening and it may be possible that the AI system is not able to decide in the same manner you want it to decide because it has its own intuition, its own knowledge and its own way of thinking. So what comes as a output could be ambiguous and you spend a lot of time in correcting those thing which it has generated or fixing the error which has happened because of that. So that is a problem with the current approach. What specdriven approach says that okay let me not directly jump into decision making let me go ahead with a precise specification of what I actually need what my requirements are and once I have those requirement and I've given the requirement and design like okay I want to use Java NoSQL databases I want to deploy it into AWS as a serverless function so all these will help AI in deciding properly and that's why your output would become really precise so you may will start with the ambiguous prompt. You don't need to figure out everything on your own. But before you start writing your code, you

would go ahead and take those specification, get it precise, verify it, update it, and then you can keep moving forward to get the precise output you're looking for. So basically whatever the black box you are looking for where decision being was being made, now you have some control on how those decisions are being made and in what context that decision are being made. That is where this specdriven development really helps. as it makes sense. — Got it. It's so we are basically providing it more context and focused requirements rather than giving it just like hey build a sudoku game. uh we can probably give it more context what kind how what kind of grid do we want et you could still start with saying to AI that hey build me a sudoku game then it will figure out okay do you need a sudoku of 9 by 9 6x 6 12 x 12 right it will ask what is the difficulty level you are looking for those type of stuff or maybe do you need to have it autocomplete feature right so it will come up with those ideas you fill up those ideas you select some of them and say I need this I do not need this update it and then the development actually happens so rather than just starting code we take a step back we properly think what we need and then once we have finalized what we need we can then start it doesn't mean that I cannot change my requirement again it may happen right in product requirement changes in software development requirement changes so we can still go ahead with that but idea is that we do not directly jump to code — makes sense so in the previous process also we could have reached our goal but it would have taken longer because I think there would have been lot of back and forth interaction with the agent with clarifying here we've given it precise requirements so it knows the path. — Yep. — Get it? — That makes sense. — Now, now let me give you an analogy so that you can get it a concept better. So let's say we talked about handyman. Handyman walks in and says hey yes I can fix it and he will be able to fix it. And if it happens that okay he tried one way and he tried to open something and let's say into a plumbing it didn't open then it says okay let me start another approach let me cut it down and then fix it again. So idea here is that it may be good for a small problem. It work fast relies on experience and intuition. It may able to provide you quick fixes but it is mostly on the intuition based learning we have. But when we talk about specdriven development for bigger project think of like a team of architect working together. So they will be able to perform a predictable outcome for you. So strength of this particular team is like okay they have worked on these they have figured out all these ways how these all parts fit in where the beam would be where the plumbing will go where the electricity will flow that all has been figured out. So this is very scalable system no surprises later like sometime it happens that a handyman asked me to buy something and later we were able to figure out that this may not fit into the situation which we have. So I'm not saying handymans are bad or w coding is bad but idea is what you want it to solve that becomes most important. So obviously when you are going with specdriven development you will be having a little slower start because you have to figure out those requirement and requirement upfront thinking requires upfront thinking you have to figure out lot of thing and then you get started but yeah this is what analogy I would say you hire a team of architect for a bigger task for smaller task probably why coding can really work well for you. — Great. Yeah, it really gives me a good picture of what probably KO would uh help me achieve. — Excellent. So, so continuing on that. So, specdriven follows a well- definfined path. So, if I go into a little more detail on that, so we architect create detailed blueprints, structural, electrical, plumbing and plans are finalized, approvals are done before breaking the ground and that's how we are able to get the things done. Same applies to coding. You go from requirement to design and then implementation phase. There are clear interfaces, contract and architecture defined, predictable and scalable system comes out of. So everything fits together. No surprises at the last. That is what a specdriven development is all talking about. Does it make sense? — And Ko fits into this whole situation really well. — Makes sense. — Good. So let's talk about the topic now which is KO. But I wanted to give you extension of understanding like where we need it. The tool is good if we understand its purpose and if we are able to leverage it properly otherwise it is of no use. So Kino is a really nice tool a really recent product from AWS and it is able to solve a lots of problem. I use it on regular basis or lots of other stuff also apart from programming. So that is a really helpful tool for me also and for everyone I guess it will be. So KO is what we call it is agentic AI development environment from prototype to production. So you could go ahead and use it for prototyping and later when you decide that hey it has to be moved into a production environment that time also I can leverage KO should not be a problem. Right. — Right. So how can I get started like how do I install it? Where do I download it from? Can you give me an overview? — Yeah very simple. So very simple is like go to ko. dev and it is free to download. You could go ahead and download it. It is available for Windows or Mac OS and

download for Linux also. There is a KO CLA also available. So you could just go ahead and download and get started for faster. Like if it is a free account you are using, you still get 50 credits to use and if you use builder id then probably there are more 500 credits or something comes out of it. So it is easy to get started and once you download and install it and installation is pretty simple and straightforward. It's not like very complicated thing. Once you have it, you would see a application or IDE in your environment which you can get started with. Let me bring it to the inh to the screen. This is how it looks like. So if you look here, this is I am on to KO and through this I'm able to perform lots and lots of thing. We'll talk about some of these feature but basically it is a fork of VS code. So it looks interface looks like VS code. I'm using a light theme here. Yeah. And then here are the things which we talk about ko and we have them ko power. So this is where the all the benefits come when we are talking about all these things right. So this is how — before you proceed I have one question. So ko obviously uses uh the ML models right uh for the agent to work internally. Now when I uh download Keo and install it does it actually download and install even a um a model on my local computer? Uh no it's not it doesn't work in that way because if you do that would be too much for a local computer to handle because then it will not give you the flexibility we are looking for. So what KO is giving you K hero is downloading the interface or ID for you but at the back end it is making calls to the Amazon bedrock services and in those bedrock services these are the model which are optimized to run with Ko. So Ko can use any of these model depending on your use case. So if I have to do a very complicated calculations probably I will go with crowd opus but if I have very simple things to do maybe I could go with minax the sonnet or any other thing. So that is how I should be able to leverage these different models. So when I go ahead and launch ko let me show this thing to you give me a minute I'll bring it up and then we can get started from there. Where is my ko software? Yeah here. So see this one here. Here it says auto. So this auto is selecting the model automatically for me and if I want to change it obviously I could go ahead and use different model depending on the model you use your credit would be calculated differently. The same task if I do with automodel maybe I will be charged one credit is a unit which is a work done by AI. Let's say I give it a simple prompt maybe it will use one credit to fulfill that prompt. So based on that I would be able to select any of these models. Some of the cheaper as you could see like minimax some of them may be higher cost depending on your requirement. — I think the slide is not seen. Did you want to show the slide? — Uh slide is not visible. You can see the kilos. ID we are looking at Ko right now. — So here so if you see 1. 3 1. 425. 15 these are the different models which you can leverage in — so this is what I wanted to ask you Ashish I can see those uh I could not really make out what that meant because I understand bedrock pricing but how does the ko pricing work you spoke something about credit so is it like I have to subscribe to Ko and I get n number of credits and I can only use up to those credits. How does that work? — So, so Kir's pricing is made up of two things here. One is subscription. So, which you pay like a fixed fees for subscription what you have right and in their subscription you get some credits to you and on top of it there is a usage based pricing also. So, you got a subscription in that let's say you got 50 credits if you go beyond it then you pay on the per credit basis what we have utilized on that. So, subscription plus usage these are the two combined prices which happens for K. Okay. So what are those things which I'm seeing like you know X1. 3 what does that actually mean? — Consider like let's say you ask a question to Ko let's say I want to write a code for a tic-tac-toe game right let's say that prompt consumed one credit so it was ko auto it consumed one credit if you ask the same task to be done by cloud opus it would have consumed 2. 2 2 credits on that. If the same task was done by minimax, it would consume 0. 15 credit on that. But it may not generate the best output you are looking for. So credits are basically a way to calculate like what is the baseline we are utilizing cost associated on that. So basically one task you can think of it is using credits on that and that credit is basically the work done by AI and higher model would charge you more cheaper less. Very simple. — Got it. So the multiplication factor is more because it's consuming more tokens. I believe — whatever you consume us more both things are there. — Got it. Makes sense. Great. — So maybe you could give me a you know a demonstration of some of the UI and

features of KO. — Yep. Why not? Let's get started with that. So very simple interface not a very complicated one. So if I'm looking at this right now this is like a phase panes you are looking into. This is my traditional directory structure which I'm seeing here. And obviously you can go ahead and create new files here. Here are your ways to customize the layout for you. Maybe I don't want to use secondary bar. I want to sync into left or right or panel where it should be there. Then I can go ahead and select different options from here to get started. Here is my chat option. We'll talk about that in a minute. There I can go ahead and say okay I want to go ahead and start with VE or spec. We'll discuss that. Here is your command prompt also. So any command I run here should be visible so that you don't have to keep on moving from a command interface to a text file or to a session or a prompt everything is built into one single page and here are your integration of source control if you can see so here is a get changes if required when I did initial commit and all obviously there is option for running and debugging your code and then some extension which we can add if the need be those are there and here is what the ko thing is so if you remember I talked about that we go ahead and start with something called first starting with our requirement and then we go ahead and then design perform our uh task. So that is how this whole structure is designed for and there are multiple options here like spec agent hooks trading and scales MCP server. So we'll talk about it if the time permits. If you are not able to cover everything don't worry like in the workshop you would be able to go through all these topics which I'm explaining to you here. Right? So this is a very easy to navigate interface. Obviously you can use different theme, different font sizes and all but it is a desktop application which I installed and from here I could keep on moving forward and keep on performing the task which I'm looking for and to get started here are my option like I go into chat and say okay do I want to go ahead with coding or I want to go with specdriven development I could go ahead and put those prompt on that. So we'll talk about it if the time permits and we'll start building some applications based on that. — Makes sense. So just a small overview of if I have to build a feature a small feature. So what are the type of specs that we have and uh yeah what some of the capabilities also if you can tell me start by building a sudoku game very simple one I'm taking as an example here. So let me prompt get this prompt and this is a link which I have kept into the slides also. Obviously everyone gets the slide afterwards. So you should be able to get started with that. But let me bring this up and we'll get started that how easy it is to get started with Ko. I'll open it into a browser and we'll pick up from there some information and we'll get started. Right. Let me go ahead and pick up. So this is a workshop which people can run in their own environment if they want to. You can download it and then start with this hands-on lab. We have a different lab. So this is what we are trying to build. We want to build a Sudoku game. So we first need to get workshop. We need a builder ID. Okay. I forgot to mention it. How it integrates with AWS is through builder ID. So you need to register for builder ID. You could use your Gmail addresses also. Should not be a problem. And then you can start building your Sudoku game. So I go here, I say open a project and I go into specification and say hey I want to perform some automation for my sudoku game. So I could start by putting some prompt and that will be a starting point for me. So first thing I can do is here I open a project and get started. Now I have already done few of the things. So that's why here I'm just saying that please add few things to a string file. This is what we could do stating file. But let me go ahead here into ko again and ask that please go ahead and build a sudoku game for me. Let me put a prompt here. So I'm going into spec and I'm saying please build a sudoku game or should have web interface and option to auto solve. Right? This is what my requirements are when I go ahead and starting it. Right? So what KO is now doing it is creating something called steering document. Including steering document. If you look here, there are lot of steering documents I have. What is steering document? It is like a GPS. So when you are going from point A to point B, you could drive your own. But if you have a GPS, it will help you out in getting proper navigation. So let's say in this case, I want to follow specifications like okay, Java best practices. I want to follow a specific design I'm looking for. So all this information I can add into my steering document which helps Ko in guiding through the decision point it is making or to enhance the code it is creating. So now it has started with this process and it has included the steering document. I can see there is already sudoku spec some existing code in the workspace. It was able to find it and now I can go ahead and look sudoku game is already fully built. you got this you can open this or you can go

ahead and start want me to add anything to it like more puzzle or number pad for mobile or something so I could go ahead and do that so what it did when I have started this project it had gone into specification so in this spec this is my sudoku web app and this is my requirement which has been finalized on that see I was talking to you about the three phases requirement then design then task so these are my requirements so based on the prompt I have given it has figured out okay what I need a minimal browser based sudoku game built as a single web application which player will fill in a 9 by9 sudoku grid. What is app? What is board? What is cell? This is giving you all the glossery terms also. So there is no confusion between developers when they are looking into it. My interpretation of a puzzle could be different than yours. Right? So that's why it has come up with all these solution. Then it started creating these user stories on that. So obviously first thing we would need is display a sudoku board. So remember I just provided a prompt. We said I want to build a sudoku game and based on that prompt it is able to figure out what is all required and it has created my requirement user stories and all and with user stories we always create something called acceptance criteria. So this will be able to accept that this feature is valid or not once it fulfills these things. So second requirement player input as a player I want to enter digits into empty cell so that I can attempt to solve the puzzle. So this is my requirement. Now all the validation required. So when a player enters a digit now I don't have to write a code which will only accept digit into it because sudoku is only done by number. We do not need characters in that ABCD type of thing. So it will perform all those validation for me. We'll create those test cases. We'll run those test cases to ensure that we are following those standards we have defined for it. So everything starts with requirement. So here is my requirement file. Based on that once I'm happy with the requirement I can go ahead into design phase. Design phase is where you talk more on the technical aspect of your application. So what I need a single page game delivered in a static HTML file if I want to change it. This is what my file would look like. Sudoku index html. What my key goals are I don't need an installation. It should have minimal footprint and it should be have wind detection. It should have backtracking. It should have conflict detection. All those type of stuff. So index html is being created for me. Sudoku. js will be there. Well the all the game logic will be it is all done by ko for me based on the prompt I provided. So once I'm sure with requirement I will then move on to the next phase and we'll say okay I want to design something and once I am into design phase once I approve it I should be able to move on to the task list. So see this based on the design it has created these simple tasks for me and then I can start executing these tasks. So first it says how to implement a wind detection. So it will be done by non-zero value. What is my requirement? What would be my task associated with it? See this one it is also performing validation of those requirements. So it's not just writing code but ensuring that the code is of high quality and it is able to serve the requirement which you have specified on that. Obviously there would be checkpointing required. Then backtracking solver required. So all of this has been created by KO for me directly. Great. So just to summarize Ashish. So I think each stage is feeding into the others. So those steering files you starting point requirements right then we go to design. So this requirement is where we are able to get all this requirement sorted. Once I'm done with that, I will go ahead and say okay, now I want to go to design frame and in the workshop which you will be shown by more later, you would be able to go through all these things, right? And then I will create task and then I could go ahead and say okay, I'm happy with this. Let's go ahead and perform the task running. So I already done that to save time. So that's what it happened like it was able to give me this sudoku web app and it has created all these file like design my markdown client requirements and then task. md and any output which is coming up that can also be stored here like there is a sudoku and in that I have my index html file which has been created for me right so this is how it is easy to get started shouldn't be a problem at all in getting these particular things started and then move let me show what the output came out from that give me a minute I'll bring that something up here uh ko introduction sudoku so this was the output generated once I completed all this task it was done yesterday so this is my index html and within that it has created this sudoku game for me and obviously based on the design document I could control okay I asked it for a 9 by9 I could go ahead and say no I need something else I need a bigger sudoku I need higher difficulty those type of And here I could go ahead and add numbers and then move forward from there. So those is how a simple gam has been created for me just by a prompt. So it may look white coding because I use a simple example but I could go ahead and

create more complicated things like okay let's build a fitness application which will help me to manage my fitness activities and determine if I'm actually doing the exercises and how many calories I burned in the process or not. So these kind of prompt which you start with you are still talking in natural language but to get started it create those requirement file and then keeps on adding to those requirement or refiling those requirements with design talk and at the end it is able to create your required code or required application which you wanted it to make sense. — So I just had one thing so for example after you created that pseudo co app and I want to enhance it say I want the user to be able to choose the different levels. So if I vibe code at that point and say hey now make changes so that uh the first screen should allow the user to choose between beginner uh intermediate so it will go and make changes in all of them right requirements — exactly so it is able to see the context and the relationship between all the file let's say now I say I want authentication mechanism for this right so it has to look into where all authentication mechanism will fit in or I want to have those codes there into databases. So it will figure out in what files we have to make changes. It has those contexts and what it will do like if you are talking about an existing codebase it is able to store that information into something what we call sharing file. So in steering file it is able to capture all the information I had like this was my sudoku information. So all this information has been stored what my design would look like what would be my technical specification would be structure product Java. This is all because I have done multiple project. But when you start small, you would have only your specific files into it. And if you look here in this window, it says you can open index directly. And here it says, want me to add anything to it like more puzzle, difficulty selection, a timer, or number pad for mobile? Okay, let's see. Yes, let's add a 3inut timer. Right. So I'll go ahead and put this prompt on that. It will figure out that where this prompt needs to be implemented, what design change would be required for it, how it would affect the existing functionality, what code changes are required for that and it will give you those specific things for me. So here it is asking me see this it is not blindly going and forward doing all those thing. It is asking me question that okay are you building a feature or are you building fixing a bugs obviously it is a building a feature. So it is waiting for my input. I go ahead and say okay this is my requirement and go ahead and build a feature for me. So it will go ahead and make changes into all the files which has created before. It has context and based on the context it will go ahead and provide this. So see this what do you want to start with requirement or design? Obviously I this is my requirement. So I will go ahead with requirement. I submit the answer. So I'm building a feature. I provided my requirement. It will now come up with the changes required into design. Once I approve it, it will go ahead and perform those code changes if the needed. So including cering document, it is reading all the file which I have and based on that it will come up with what is required to be made changes and where it has to make changes. So here it is changing. So config. o exception my requirements are being created or updated as we see the same file which I already created is getting updated because of the prompt which I have provided. So it gives you the wipe coding also but it is not like completely going out of proportion. It is always keeping things on track based on the requirement you have and the design you are doing and before you move forward it will ask you to point out that it is interesting to note that it is also showing all the steering docs. So I guess I can see that you've there is a steering dock for Java best practices. There are coding standards. So it will never go wild, right? It will always stick to those. — Yeah. Exactly. Because you may let's say you're fixing a bug and in your existing code you already have standards to follow. You already have a way of writing your code, way of commenting, way of documenting. So you could put this information into your steering file and steering file as I said is a GPS which will keep you on track. You don't navigate here and there. You always see where you have to go and you have somebody even if you go here and there GPS will tell you that okay you are you have taken a wrong path go to this dane and you will be on the right path again. So it will take care of all those integration required because of this. — Correct. And I think then I don't even have to add those specifications in the prompt itself because it has awareness of those steering files. So it prevents the context overload also. It always knows. — So what happens let's say if you go to model and provide all those Java best practices again as an input then it will be too much big context for a model to understand. So what we do here instead of bugging that context window we provide that information into spec file and sorry into this your — steering file and that based on that knowledge the context is adjusted when requests are being going to the machine learning model behind the scene

— great yeah totally makes sense — it has come up with this so except creation of design MD now if I go into this design domd it would be able to tell me that what all changes it has done on this cases Right here it is my requirement. I will go ahead and see this my product MD visual design MD and I should be able to go ahead and make changes in that if so it is a little intuitive but it has like lots of pain so people get confused. So don't rush take your time trying to understand what it is trying to do whether it is working or not. So see this one it is working on config. 0 requirements. mmd and designd. So here it is showing me what it has done, what files we have and other information like the console port information all of these are there. So, so we do not have time to cover everything but what I'll do I'll give you all the details like if a spec is in spack we have user story in design we have architectural design in task we have implementation steps and then I have given you the analogy for steering document what is inside a steering document why steering matters and I use the analogy like we use GPS to move forward from one place to another place same way we have used steering document so we guide hero how a solution should be built based on the requirement we have given to it and then we have MCP integration also. So if you go ahead and check into the KO console I have opened there is integration with MCP server also. So it is not just doing things for you but it is getting things done through other agents also if you want to get started. So here I have integrated few MCP servers like fetching something from AWS documentation playright I may use a docu design app here like AWS diagram builder as MCP. So whatever I created it can be created into AWS diagram. So lot of integrations are possible which can be done. So we'll give you all these slide so that you could go ahead and get lots of stuff and then it has integration with third parties also. So let's say you want to follow best practices of postman or stripe payment integration or data dog observability. So it has integration through ko power. So the other icon which was shown into the console is of ko powers. So here the last one here. So this is ko powers which you can also install and create your own custom powers to enhance its functionality if the need right. So we have time constraint partner otherwise I would have loved to walk you through all these feature but yeah we'll give you slides to this so that you could go ahead and get more things started also we have key takeaways here like start with real problem not the technology is solving a problem but you should focus on what the problem is actually use ko for prototyping validate ideas quickly spec based development scale beyond developer so multiple teamwork members can work on that and you can insert K in your SDLC enables collaboration across technical boundaries. I have also included lots of workshops here you can try on your own and then I've included links for some reinvent videos which can be really useful to learn further on that. So we obviously had could not cover everything we wanted to because of time constraint but I hope it has given a lot of information for people to get started and start utilizing things. — Yeah, this definitely was insightful and I'm looking forward to my journey to AI based development. Thank you so much. It's interesting and thing is it is AI like start with something even if it is a small thing start with it and keep on building on top of it. So don't hesitate download ko there is no cost to download there are some free tokens on that and whatever I did like I would be able to see how many tokens I consume. So just the last thing before I give it to the uh behavioral track series. So keto free bonus I had 500 for to be consumed in 28 days for all the things I have done from last yesterday I have just consumed 90. 9. So that will obviously help you to be on track and what is your current status and how much you consumed how much is remaining. So there's no problem in starting these things. All right good.