Transcription
Hey guys, I was just able to compete in Google's second episode of the AI Agent Bakeoff, which is a five-hour challenge where we had to build a brand new Agentic banking application in just under 5 hours. And thanks to enough caffeine and the right strategies, I was able to win with my partner.
So, in today's video, I want to break down those exact strategies I used into five helpful tips that you can copy and steal for your own workflows to build better and faster agent workflows for your own applications. Also, if you haven't had a chance to watch the entire episode that Google just produced where we're all competing and building out the best agents, I definitely recommend you checking out that full video. I'll have a link to it down in the description below. But don't worry, I'll still give you guys a quick recap of all the important things as we break it down here in this video as well.
And because you guys are awesome, I'm going to be giving away the entire source code for the project I built during the bake-off completely for free. Just click the link down in the description below so you can get access to the front-end application, plus all the different agents that I built under the hood. But enough of that. Let's go ahead and dive into breaking down everything that went into building out this agent project so you can start to copy my tips and tricks for building real-world agents.
All right. So, what I want to do is paint the picture of exactly what we were getting ourselves into in this competition because in a matter of five hours, we had to go from a list of instructions, requirements, all the way to a full-blown working application that we could show off to the judges. So, I just want you guys to see how much we had to do in 5 hours because you're going to see, man, this is an insane amount of work and the only way I was able to get it done is by doing like AI-driven development. So, that's exactly what I'm going to be sharing with you guys in the five tips section, but I just want to paint a picture of what did we have to build, what did we build, and then the rest of the video, I'm going to show you how we got there.
So, let's break down what we were tasked with doing because this was a ton of fun, but stressful because we had so much to build in such short time. So, we had to create a multi-agent system that builds the future of personal finance. Now, what does that actually mean? Well, we needed to build out a beautiful UI where users could go in and engage with agents. We had to make sure that our agents we were building offered proactive guidance to help people accomplish their major life goals for budgeting, doing vacations. Outside of that, we needed to make sure that our agents were intelligent. So, not only did we need to include chatbots where people could actively talk to these agents, but we had to add in workflows that automated a lot of the work that people would normally go do when like planning a trip and stuff like that.
Outside of that, just like you would have to do in Top Chef where you're like instructed to go make a certain type of dish. Well, you have to use certain ingredients as well. So, inside of the bake-off, we had to use certain ingredients such as, or in our case, certain technologies such as Agent Development Kit. We had to use the Agent Agent Protocol and we had to use Gemini API. So this was a ton of fun.
But what I want to do is just break down, like, let's actually get really concrete and let me show you exactly what Google was basically asking us to build. So at the end of the day, here's what Google was saying. They're saying, "Hey, we already have an existing bank. It's boring. It's an old bank, but this bank does come with a backend and a front end, but most importantly, this bank comes with an agent. And as time goes on, agents become more powerful and more prevalent. Agent-to-agent protocol is going to become huge where you are going to have agents that are going to be talking to other agents inside of companies to where they do all the work and basically make things magic for us."
So what Google was asking us to do inside this competition is to build out everything that you see right here in 5 hours. Meaning we had to create a custom application and we had to, for each one of the main topics that we were asked to do, like working with financial statements, helping people plan their goals, you know, a bunch of like analyzing perks that might be available to the customers inside the bank. Like we had to build out agent workflows for each one of these, which is insane for 5 hours. Not only that, but we had to figure out how to connect our agents to the bank's agent and we were able to do that thanks to the A2A protocol.
So, as you can see, this was a ton of work, but let me actually just show you a quick overview of the application that was built. So, you can see, you know, this is what me and my partner built in 5 hours. So, super, super powerful and super, super proud of what we built. So, this is a full-blown application. So, you can see we have multiple tabs to help people achieve their different goals. So, we have in each one of these pages correlates to different agents and agent workflows. So, you can see we have spending. So, inside this, we have a spending agent workflow where we had an agent go off and analyze the person's income, expenses, recent activities. Outside of that, we also had a, you know, a other agent that was responsible for asking and answering questions on everything that you see on this page. So, as you can see, like this was a ton of work that had to get done in multiple hours. So, this is one part, this is another part, and this gets repeated on every single page. And this was the way I wanted to structure building out the application. Definitely recommend you checking out the full episode so you can see how other people tackled it.
But once again, super excited for it to work and just to show you guys, you can see like what did I spend the most on this month? This is triggering off an agent that's asking questions over here about what was spent on. So, you can see the most expensive thing that we did last month was rent and you can see it right here. So, yeah, all around, like, it's a working application. You can try it out. You'll learn a ton of cool insights on how we set up the agents to actually handle working with A2A. So, yeah, there's a ton inside of here. I won't go super, super deep into actually how everything was built. I definitely recommend you guys hopping inside of here and just asking AI like, "Hey, please walk me through this." There's a ton of cool, really interesting places that you can use to jumpstart your own full-stack AI development journey.
But what I want to mostly focus on is how the heck were we able to build everything that you see over here in five hours from all the agents to the full-stack application to the agents having tools to do, like, this was a monster task and I want to show you how we did it in 5 hours and I'm going to break it down right now in five tips that you guys can copy.
So, the first tip I want to share with you guys is the concept of an AI reference project. Now, this one tip alone is the single most important tip that allowed me to go and build out this full-stack AI application in a matter of 5 hours instead of 5 days or more, truthfully. And the best way to fully understand this tip, and don't worry, I'm going to share examples of how I actually did it. But the main thing that you need to understand before diving into what is an AI reference project, you need to understand the new job promotion you just got thanks to AI agents. So, you are now an AI developer. You're no longer a software developer. You're an AI software developer. What that means is at this point, you should probably never be writing a single line of code again. I have not written a single line of code in probably close to 2 years at this point. What I have been doing though and what I recommend you guys to do and exactly what I did in this competition is I provided context to AI agents. What that means is I'm telling AI, "Here's what you should be doing. Here's why you should be doing it. And ideally, here is an example of how you should do it." And that's what really what I mean by AI reference. We are referencing existing working code and different projects so that we can just go exponentially faster.
So, let me break down a little bit more what the role means and then I'm going to break down exactly how to apply and explain what AI reference projects are. Okay. So, here's what I mean. At this point, our job is to provide context to AI agents. We're not coding anymore. If you're coding, you're going slow. So, what do we need to be doing as AI software developers? Well, our job is to know what exists. So, what does ADK do? What does A2A do? What is possible with these frameworks? What can I do with ADK? What can't I do with ADK? When it comes to A2A, what does it allow agents to do? And also, more importantly, we need to know where can I find working samples and examples of these different frameworks. So, is there an example project of ADK working properly? Is there a working example of ADK using A2A properly? Because if there is, there's no reason for me to actually, you know, type out and explain and manually type out, "All right, new function, go get this, you know, card from this other agent." No, what I should be doing is I should be passing that information over to AI and say, "Hey, AI, I know that this tool can do it. Here's an example of another project doing what I want, and I would like you to tweak it for our use case." That is what we should be doing in the world of being AI developers. And this is exactly what I did to crank out five days' worth of work in 5 hours.
So, let me give you a concrete example. In our case, as you saw, we have to build out an ADK application that has a full-stack application using ADK where ADK is using A2A to talk to the bank agent. So, my whole job going into this competition, knowing my new role as an AI developer, is I'm like, "Man, I have to get smart on what ADK can do. I have to get smart on finding example projects." So, when it is game time, I can say, "And just, hey agents, go to work. Here's exactly what I want you to build."
So, what does this actually mean inside of your projects? Like, what should you be doing? Well, what you should be doing is you should be preparing your projects to have a reference folder. This reference folder is combined with every type of project reference that you can find. So, you should be looking up and adding in existing applications. So, in my case, I've already built out A2A applications in the past. I have found a ton of ADK documents that explain exactly what you can and can't do. I have projects of ADK for building out tons of different projects. So, these are all different projects that Google has created to showcase the different powers of agents. Outside of that, there is other agents where there's projects that showcase how to use ADK plus A2A. So, my job at this point, heck, there's even an A2A samples repository that shows exactly how to use ADK plus A2A. So, my job is just to go search the internet and I'm going to use AI to do it. But I'm going to go search all example projects and basically prepare and stack the deck in my favor when I'm going to start a new project. So, whenever it's time to hit go, I already have a thousand repositories of projects that explain exactly what functionality we need to build.
So, in our case, when it comes to building out this application, all I'm basically saying is like, "Hey, I would like you to look at project one where they connected their ADK agents to A2A. I would like you to look at project two where they connected ADK to the front end. And when it comes to project three, I would like you to look at how they built out multi-agent workflows where they had agent one work with agent two to then spit out a result." So, I'm hoping you guys can see at this point, the kicker here is we just need to be basically understanding what is possible and then just having like, you know, in our heads, just a bunch of like little memories of like, "I remember that did that. I remember that did that." So, whenever it comes time to coding, we can just pass all this into our projects, specifically inside of our reference folder, and then let our agents do the work. Because the agents, they're so smart. The second they see the code from here and the code from this other project, they'll go, "Oh yeah, I can copy that for your new project. That's going to be so easy. Thank you for providing the necessary context so I can implement exactly what you want." So, I'm hoping this is super helpful because this is the exact process I did to crank out weeks' worth of work in 5 days.
All right, so let's go ahead and hop over to tip number two because if you thought this one was super helpful, you're going to love tip number two. Also, if you're liking the idea of reference projects, you're absolutely going to love what we built for you inside of Shipkit.AI, which is a combination of pre-built ready projects, AI-driven courses, and everything else you need to launch AI applications in days instead of months. And just inside of Shipkit, like I said, you're going to have access to a course where you get to learn how to go from an idea. And AI is going to help you every step along the way to go from idea to a full-blown application to deploying your application. Like I said, we have every type of common AI project you would like to use. Everything from Agent Development Kit to RAG to Chat, everything. We have a full-blown pre-existing AI application that's working. You get the source code for it. You also get example walkthroughs of how I converted each one of these pre-built templates over into a custom project. And you get to see exact breakdowns of how everything works. So, Shipkit is the number one tool if you're looking to build out real-world AI projects using these new technologies and tips and tricks and see exactly how I build real-world applications in days instead of months. And if you're looking to get Shipkit, we're doing $50 off if you want to use code bakeoff. And if you have any questions, feel free, let me know. But I definitely recommend checking watching this video if you have any questions or shoot me an email. But yeah, thanks guys and let's get back to tip number two.
So, the second tip I want to show you guys is that you need to be talking to your computer and this absolutely broke everyone's brain at the competition because everyone else is sitting, you know, heads down, earphones in, and they're just typing as fast as they can. And I, on the other hand, am just talking to my computer saying like, "Hey, I think we should build this. Oh, and this needs to connect to this." And basically, I'm just talking. And the way I'm able to do that is using the tool called Whisper Flow. And Whisper Flow literally just listens to you. And they have a bunch of cool tools in here to help out if you're using tools like Cursor or Claude Code where they'll also look at your files and add them in. And at this point, why I love Whisper Flow so much is because I'm able to just go as fast as I can talk. So, if I can talk faster, it'll listen. If I talk slower, it'll listen. It doesn't matter. At this point, I wish I could talk and type at 122 words per minute, but my fingers just don't do it. And it's so much easier just to sit back in your chair, look up, and just describe what you want and talk and then have the AI build it for you. It's a funny thing at first, but the second you start doing it, you'll never go back to typing unless you have to, like, you know, call out something specific. But just when you're trying to solve the context problem, which is what we're trying to do right now, we're trying to put as much context in our heads into the computer as possible on how to solve the problem so the agent can go off and do the work. So, the more you can talk, the more you can explain what you're looking for, what problems you're running into, and what the ideal goal state is, the better these agents are going to do. And that's why I love Whisper Flow because it breaks down that barrier of preventing me from putting more context into the system. And if you want to get uh Whisper Flow, not sponsored by them or anything like that, which I was, but I have a link in the description below where you guys can get two months for free of Whisper Flow. Absolutely love it.
But what I want to show you is just Whisper Flow in action really fast. So, I can update the shortcuts on my computer to where whenever I now hit the option and spacebar, I get this to where I can just talk to my computer. It's listening in real time. You can see down here at the bottom of my screen, it has the little squigglies. And this is where I can just say like, "Hey, we're trying to solve this problem. Make sure to look up at setup.md file to see how we are trying to tackle this problem." And then, "Oh, by the way, this other reference project solves the exact thing." And then, as soon as I'm done talking and rambling, which is totally fine to do, like you can ramble, it's totally fine, the AI will just put that in as context and boom, it'll work on solving it and it's hooked up to Cursor so it knows whenever you call out a file, "Oh, I need to include that file." So, this thing is just insane.
Now, if you want to do the poor man's version of Whisper Flow, totally understand. I have a cool thing I want to show you. If you're on Mac, what you can do, you can actually go over to settings and type in the word dictation. I wish I could make this bigger, but for some reason, they don't let this get bigger. But what you can also do is update the dictation shortcut once you turn it on to hit the right command key twice and you can actually talk to the computer. So, let me show you what this one looks like because this is also super helpful. So, you can say, "Hey, I'm just trying to show to YouTube right now how cool it is to use the dictation tool where you get to use your Mac and just talk to it completely for free." And this is automatically included right out the gate when you're using Mac OS. So, yeah, as you can see, this is an insanely cool tool. So, if you're not able to type at 122 words per minute, you definitely want to consider using tools like Whisper Flow or dictation so that you can get all the ideas out of your head and into the computer to help solve that context problem so that these agents can work their magic. So, yeah, that's tip number two.
So, let's go ahead and hop over to tip number three. So, the third tip I want to share with you guys is the concept of building out an agent workflow digital twin. Every time I have shown someone in my free school community or inside of Shipkit this concept when they're building out multi-agent workflows, the light bulb goes off. They go, "Thank you, Brandon. I will never build agents the same, and I'm going to be using this every time going forward." So, let me explain what the digital twin concept is and how it makes building agent workflows a thousand times easier. So, I think what's best is to let's cover the problem first. So, the problem is when we're building out multi-agent workflows, we're creating a bunch of different files. So, for example, to build out this workflow right here to where we have a root agent where our root agent has sub-agents and each sub-agent has, you know, different callbacks, different tools, they might access state differently. Well, as we begin to build more and more complex workflows, these grow in files. So, we are right now ending up with at least four different files where if we ever want to make a change to, let's say, our root agent and our root agent changes the way that it accesses state. So, instead of writing, you know, "Here's the goal," instead of writing the goal, it now writes, "Here's the task." Well, every one of our sub-agents that used to be looking for "goal" well, now it needs to be looking for "task." So, we accidentally just broke all of our sub-agents because as we're making a change, the AI like Cursor, for example, is going to go, "Yeah, I I'll happily update your root agent, but as a result, we accidentally break everything downstream." That's the issue.
Now, here is the solution. What we end up doing is creating a digital twin that is a direct replica of our actual workflow. So, you can see what we're basically doing is at a high level, we're just creating a markdown file that says, "All right, our root agent, our root agent has access to these different sub-agents. In addition to having access to those sub-agents, here's the tools that it's going to do, here's the overall goal and what responsibilities this agent does, here's the inputs of the agent, here's the outputs of the agent." And we're basically just replicating what's on the left and we're putting it into a markdown file. The reason why this is so helpful is because anytime we go to actually make a change, like, "Hey, your goal now is to not write the output to be like I said, 'goal,' it now should be 'task.'" Well, whenever we go to make the change now and we actually pass in this markdown file, what's going to happen is our agent's going to go, "Hey, I can see you're trying to make this change, but did you know you're accidentally breaking the three other agents that need this as a required input?" And you're going to go, "Oh shoot, thank you for telling me that. I will would like you to also go update these other agents to make sure that they reflect our change." Now, that's, I promise you, the second you start building digital twins of your agent workflows, you're going to be like, "Oh my gosh, this used to be such a pain in the butt where I'd fix one thing and it would break four other things." But now it's working like a charm.
So, and here's exactly what you need to do. You need to create, like I said, an agent digital twin. And all this agent digital twin needs to do is list out at a high level what each agent does. And every time you make a change, you're going to update this file. So, this file is going to include, like I said, the name of the agents, what they do, what tools, callbacks, and everything else. And then outside of that, you also just want to list like the high-level overview of what each, like, how do all the agents connect. And you want to just basically keep replicating everything that's done over here and actually just replicate it over in your digital twin. You can definitely create yourself. It's already included in Shipkit. But this one change right here, every person that has started to create digital twins inside of their own workflows, they never go back because the second you just, it works flawlessly and you get to move a thousand times faster and you can build more complex workflows because everything's stable. And when you make a change, you know where else that change is going to cause impacts downstream and you can fix it down there, too. So, promise you guys, digital twin is an absolute cheat code if you're building out agent workflows.
All right, let's go ahead and hop over to tip number four. So, the fourth tip that I want to share with you guys is that you need to train AI to work with your tech stack. This is one of the biggest unfair advantages that you could have to code a thousand times faster than everyone else because while everyone else is, you know, working on implementing a feature and their agents are adding mistake after mistake after mistake and they're just going in a circle saying, "Please AI, stop messing up. Just build the feature. You're already on the 10th feature and you're already done with the project and on to the next." And the way we're able to do that is once again by training AI to work on tech stacks. So, let me just give a quick noob versus pro because I'm hoping this is going to make sense when we click it in. But how noobs usually use these different tools like Cursor and Claude Code is they open up a new chat and they say, "Hey AI, please create me a new ADK agent that's going to make a tool call to get the weather." Now, they're going to hit enter. They're like, "Hell yeah, I have AI working for me." And then all of a sudden, it's going to come up with, honestly, I have no idea what because ADK was not released in 2024 when the training cut-off was for all these models. So, you're basically just having AI in no man's land trying to do something it doesn't know how to do. So, it's just going to produce something. Good luck. I doubt it'll work.
So, what we want to do is we want to train AI to implement and actually know how to work with our tech stack. And the way we do that is we create something called a task template. This is basically just an expert who knows everything about the tech stack that we're using. And this expert is going to generate tasks. So, like if we were thinking about this in the terms of people, we're basically having an expert. This expert is going to produce documents where these documents provide the exact step-by-step instructions on what code changes need to be made to actually get the desired functionality that we want inside of our codebase. So, this is the expert. The expert's going to produce the task document and the task document is what's going to be used to implement, "Oh, we need to change this file and we need to change this file." So, that's what we do.
Now, how the heck do you make an AI and train AI to be an expert at it? Well, it's actually not that hard and I'll show you an example um right now. So, this is what a task template looks like. It is nothing more than a list of instructions on how to use a specific technology. And ideally, what it's trying to do is take in the desired functionality that you're trying to build. It understands the tech stack in and out and all the mistakes to avoid. And then it's just going to write all of the functionality to say, "Hey, when we're trying to add the new feature that they're asking for." Cool. Yeah, I'll happily make that task and I'll do it well because I know exactly what to avoid and what mistakes to not make. And that way, whenever you go to make the code change, this code change is actually just going to be perfect.
So, here's what it looks like and I'll show you how to make one yourself in just a second. But basically, this document just contains all of the common features and functionalities that we'd want to do. We've basically standardized writing code. So, what you want to do is you want to say, "Hey, when you're working with ADK, you need to make sure that you can fully analyze the current project from beginning to end. So, it needs to understand the structure of an ADK project, which usually starts with a root agent and a root agent has sub-agents. In addition to that, it has tools. In addition to that, it has libraries. In addition to that, like there's a, you know, it basically just needs to know the common structure of an ADK application. From there, it needs to understand what is, what problems we're actually trying to solve. From there, it needs to understand some technical requirements about like what actually goes into building out ADK applications. So, it needs to know like, "Hey, here's what you should do when you're trying to add a callback. Here's what are all the types of callbacks there are. Here's what you should never do when trying to add callbacks." And, you know, this is just like a long, long document that describes all the best practices of working with a specific technology.
Now, they're not that hard to make. And that's what I want to show you right now so that you can basically train your own AI specialist so that when you're trying to work on a task, you can easily knock it out. So, here's what we have done. We're trying to create an ADK task template. And this is the genius on the tech stack. And the whole purpose of this genius is to create task documents. These task documents, once again, they just contain the instructions on how to update the code. Now, so you can see an example of this. I just made one for you guys. And let's look at this. So, here's an example of a task where a task just describes, "Hey, you're trying to, in our case, add the weather like we showed earlier. So, your goal is to do this. Here's the goal. Now, here is what you need to do. You need to understand the tool that you're working with, the models, the different files that exist. So, the current state and basically just keep going on from there on what changes need to be made, the current state of the application and, you know, it also is going to include all the instructions and it's automatically going to do a good job because it knows not to do A but it needs to do B."
Now, here's what you can do to make your own task, basically your task genius, because this thing is the most important part to help you go a thousand times faster. So, what you need to do is you need to say and create a brand new markdown file and say, "Hey, you are a specialist at ADK. Go use context 7 to learn as much about this technology as you can and your whole goal is I'm going to give you an input of what problem I'm trying to solve. And your job is to create a task document that includes all the code changes." And then what you do is once you have V1 of this task template set up, you take it for a run. You take it for a spin and you say, "Hey, please go add the weather to this specific agent." And it's going to come up with a task that looks just like this to where it's going to go update the weather. And then you're going to apply that task and say, "Cool, I'm going to go update the agent now." And to start, it's going to fail. It's going to write garbage code because it's it's learning. And you're going to say, "Hey, you made this mistake. Never make this mistake again." Then you're going to say, "Hey, make sure you add some instructions in this task template to avoid this mistake. Here's what you did wrong. Here's what you should have done." Then you do it again. And now it's going to make a new task to add a new tool to an agent. And you're going to apply that task again to the code, and it's going to do better, but it's going to hit another error. And you're going to go, "Hey, you actually forgot to add context to the tool call. So, there was no way for us to save the state of the weather to global state." So, once again, you made a mistake. Please, ADK genius expert, never make this mistake again. Please update your instructions. And then you're going to keep going through this over and over and over again in a feedback loop until you end up with what I'm calling the task template that is responsible for creating the task until this thing can make tasks perfectly every time. So, you're going to go through this over and over and over again. And you want to do this for every technology stack. You can do this for ADK, Next.js, everything. And eventually, what's going to happen is when you go to implement a new code change, the task document is going to be perfect and golden and exactly implement the exact thing that you want. And it's going to write perfect code whenever it comes to actually implementing the change. And that is what's going to allow you to just move faster and build out real functional code way quicker than everyone else who's constantly like, "Please AI, please work." And they wonder why it doesn't work. And it's because, like, it was never trained on their tech stack in the first place. So, of course, it doesn't have context on how to solve the problem.
So, hopefully seeing this workflow in action is super helpful because this is exactly what I have done every time I go to work on a new project. It does take some time upfront to, you know, build out these types of task templates, but the second you put up the upfront work of maybe a few hours, you're going to move faster for the rest of time. So, I definitely recommend checking out, you know, building out your own ADK task templates. If you have questions, always feel free to drop into a free school community. I'd be happy to go deeper with this with you guys. So, you can make your own or we always have these in Shipkit so you can grab them there too. But yeah, thanks guys and we're going to go ahead and head over to test number five which is where we really put some gas on to cranking out code a thousand times faster. And I'm excited to show you this next one as well.
All right, so the fifth tip that I wanted to share with you guys is that you should be doing parallel AI development to move as fast as possible in building out your AI projects. Now, what does this actually mean? Well, it means inside of tools like Cursor, Claude Code, WindSurf, you should be opening up multiple tabs at the same time to work on different features at the same time. And what this means is, you know, we're going to kick off this job right here. While that guy's working, we're going to trigger off this one. While that one's working, then we're going to trigger off this one. So, we basically have multiple tasks running at the same time to start building out our application. And by implementing this, we are going to go so fast it feels like cheating.
Now, what does this like? How do you do this in practice? And why should we be doing this? Let's go why first. Well, if you were at a corporate company and let's imagine you are the boss. Well, if you had literally unlimited employees, it would be silly to say, "I'm only going to allow employee number one to make code changes. That's it. He's the only employee that's allowed to make code changes, and I'm going to wait for this employee or intern to get back to me before I do anything else." Like, it would be silly to build out an application literally just one small feature after another. Especially when most of the time when building out applications, there's some stuff that you could make changes to on the front end while at the same time you could also probably make some changes to the backend for another feature or a new bug fix. Or if there's multiple pages of your application, there's nothing stopping you from working on this page and this page at the same time. Or if you're doing agent development, there's nothing stopping you from working on agent one. And while that's taking 1 to 3 minutes to implement, there's nothing stopping you from working on agent two at the same time to where you are, like, the main concept I'm trying to get across here is like you are the boss and there's the best way to be working is to think about it this way to where you are saying, "Hey, intern one, you know, it is your job to go work on this task. I know you're going to need some time so go off and do it. While you're working on that, I'm going to call in another intern. And this other intern is going to work on the next feature. And I'm just going to continually do this over and over again for as many features as I can handle in parallel." Like, what's crazy is the bottleneck in this workflow is us. It's how many tasks we can, you know, keep up with in our heads at the same time.
So, like you'll see me in the bake-off to where I'm saying like, you'll see this throughout the episodes where I'm like, "Hey, go make this change on the front end." "Oh, okay. Agent two, it's doing this wrong. It should be doing this." "Okay, cool. For some reason, something's going wrong when it comes to our login page before we can go to the application, work on this." And I'll literally do this in parallel. I'll have like five or six tabs working at the same time. And then as soon as I'm done, you know, looking at the task, because that's the kicker. The intern does the work and reports back to us the code change and we quickly sign off on that code change and then kick it off to the next thing or we say, "Go make this change." So, we're constantly in in a new paradigm as an AI developer. We're providing context saying, "Go work on this," and reviewing. Those are our two jobs: context and review. And, you know, to review, obviously, you kind of need to know a little bit about like the tech stack and what's good code, what's what's not good code. And you need to understand like, is this the actual desired functionality and were any mistakes made? But as long as you can just quickly review these code changes, you can work on many tasks in parallel. And our brains, in the amount of different tasks we can handle, we become the bottleneck. But what's crazy though is we're working on four tasks at the same time. So, like truthfully, me, I do this all the time now. I feel like I'm entering the matrix when I'm doing this, but I'm getting a week's worth of work done in a day by following this because I'm doing five things at the same time. So, this is, it's just a mind-blowing tip and I cannot recommend you guys enough to do this.
And so, basically, when you kind of start stacking everything that we've talked about today on top of each other, you guys are going to be unstoppable. So, you know, when you guys are using Whisper Flow to actually talk to your computer, you're going to be typing faster. So, you know, in this case, you're going to be typing at, you know, 122 words per minute. Also, you're going to be using in your own case, you're going to be using task templates, which are specialists at implementing code. So, these task templates are going to make sure that you write quality code to implement the desired feature. So, you're going to be inputting stuff faster to the computer. When the agents are working, they're going to be doing better work. In addition to that, you guys are also going to be working in parallel at the same time on multiple different tasks. So, you're like multiplying your efforts as well. And these just all stack on top of each other one at a time. And when you are working on a task, you also have a ton of AI reference projects that basically call out exactly the desired functionality that you wanted in the first place. So, between the task templates and the reference projects, my god, you guys are going to be moving so fast. So, this is one of the ways to where I was able to, in the bake-off, was one of the only people that was able to actually build and complete everything in the 5-hour allocation and have a working project to show at the end of it. And the only way I was able to get it done is to follow these exact strategies that I've been sharing with you guys today.
So, I'm hoping you've had the light bulb moment and you're like, "Man, I want to implement these so bad." And I think the best way to get started, like I said, I have the free school community. I would hop over. We have weekly free coaching calls every Tuesday at 6:00 p.m. Eastern time. Would love to help you guys. If you have any questions on any of this, walk you through exactly how you can use some of this inside of your own applications and your own workflows. Super happy to help. And we have a ton of awesome other AI developers on the call. So, I think you'll really like to get to meet a lot of developers who are on the same journey because we're all learning and getting better at the same time. But yeah, that is the five tips that I think you guys should definitely be implementing to help you build a thousand times faster when trying to crank out real-world AI projects.
So, I hoped you guys enjoyed seeing all the behind-the-scenes tools, tips, and tricks on how I was able to crank out the projects for the ADK agent bake-off and was able to win the whole thing. I had an absolute amazing time. Shout out to Google for hosting such an amazing event. And quick reminders, you can download all the source code for the completed project that I did at the bake-off completely for free. Just click the link down in the description below. You can also, I have a ton of ADK related content right here on this channel. Everything from an ADK master class, do some deep dives into voice agents, RAG agents, so many other tips and tricks on ADK. So, I definitely recommend checking out those on the channel. And I also recommend you checking out the school community I have for you guys where you get to hop on weekly coaching calls with me and a bunch of other AI developers on everyone's on their own journey. We all work together to help make sure everyone's making progress. And if you want to jumpstart your AI development journey, I cannot recommend enough for you guys to check out Shipkit where you're going to get access to all sorts of pre-built AI projects so that you can implement a lot of the different features we talked about and tips in today's video. And if you have any questions on it, feel free to always shoot me an email at brandon@shipkit.ai. But enough of that. Let's go ahead and I recommend you checking out whichever video is popping up on the screen right now. And I cannot wait to see you guys in the next one. See you.