You are not using your AI Agent tools correctly

What's going on everyone? How are we doing today? How are we doing? I see people already started rolling in. So, that's always a good sign. Let me go ahead and make sure I ping everybody on what it is that we're doing. Um, I'm excited for this one.

So to give you an idea on what we're doing here, we've been doing uh for those that aren't familiar or do not know, my name is Danny Thompson. I'm the director of technology at this labs. And one thing that we've been pulled in a lot for lately has been um uh lunch and learns. And so today I will be doing a lunch and learn that we have done for several clients as well um and helping you number one be educated on maybe some um best practices and things like that. But honestly, what we're starting to realize and learn is a lot of um software developers that work at these enterprise level companies, they're not necessarily given a ton of resources or training to kind of understand what's required um in order to kind of utilize some of these um toolings out there. And so we kind of wanted to help bridge that gap a little bit and help educate the developer community as a whole on some of the tools that they could be utilizing.

Now, I know uh for some I know quite a few people have shared this out. So, if you feel inclined to do so, please do so. Share it out. You never know who could be benefiting from this. So, uh if you're seeing this on Twitter, maybe retweet it. If you're seeing this on LinkedIn, maybe reshare it. If you're seeing this on YouTube, maybe share internally with your teams because I can tell you right now that a lot of companies are bringing this in to kind of just help educate on this. And let me share this in um my Discord real fast. We are live AI lunch and learn on AI coding. tools. Sharing this here and then sharing this in my Discord and we will be good to roll. Um, if you are watching this, drop something in the chat. Let me know where you you're tuning in from, where we're hanging out at. Um, I am come hang out. All right. So, uh, I see more people rolling in already, so that's always a great sign as well. And I'm just checking here on LinkedIn, making sure this is working the way that we intend, and then we should be good. Okay, perfect. I I see it, Jeremy. First comment. Yes, sir. I see. I see you.

Okay, fantastic. So, um, several things to kind of go over today. So um obviously this is a lunch and learn that we often do um for um sometimes executives but sometimes like senior developers architects leads etc etc uh in order to leverage. Now this is definitely a lunch and learn that is going to be kind of really positioned and geared towards the enterprise level developer. So if you're not an enterprise level developer it's good to get that insight. Um if you are one you could probably leverage some of the talking points in this for internal communications to kind of share with others. I know yesterday I had several engineering managers DM me saying that they're sharing this with their teams. So hopefully y'all all get a lot of value out of this one in particular. So kudos to all y'all for hanging out. I see some comments joining in from the UK. Hey, thanks for coming out, hanging out with us. Really appreciate that. I know Jeremy's in Dallas, so that's awesome. Um and so let me go ahead and share my screen and kind of dive into what it is that we're going to be going over. I have um several coding editors opened, but I also have a um a slide deck that will just kind of go through and hopefully help people kind of learn something additional.

All right, so it's somewhat of an introduction to AI coding agents, but really it's it's taking it on another level. So, um, a lot of the things that we're going to be discussing today are definitely going to be on, I guess, somewhat of the higher level. But I guess if I had to ask you a very blunt question, if I were able to give you back 30% of your week, what would you be building instead? Think of one thing, I'll give you like 30 seconds to decide, but what would you be building if you had 30% of your week back at that? That's a lot of time. Um, Serena, hey, I finally see live. Hey, thanks for hanging out. Um, but genuine question, if I can give you 30% of your week, what would you build? And if you want to drop that in the chat, I'd love to see what it is that you would build if you got 30% of your week back for you.

And so, when you see some of the rhetoric online where it's like, oh, AI AI is going to take our jobs and and all these talking points like it's going to do. I personally have never seen it that way. Um, I've I've been a very vocal uh person against that rhetoric. I see it as a completely different thing that I don't think a lot of people talk about. And it's a massive productivity tool and productivity to the point where it's enhancing and enabling me to kind of go in other areas. So, one thing for example is I've never been excited to make a button, but I've always been excited to do things with that button. Meaning the design of it I'm kind of whatever about it's fine. I I do love I will admit I do love a good hover effect. I am a sucker for a good hover effect. I don't know what it is. I love hover effects, man. But uh outside of that, I would care more about what I'm attaching it to, the logic, the the processes and all that. That's kind of more so where my focus has always been. And so if I can use and leverage AI to knock out some of the monotonous tasks where it's like prime example and I I'll show a demo of this as well is like what if I am building a UI prime example I'm building a website right now for my Dallas software developers group meetup group right and as I'm building the UI elements I'm just speed coding right like I'm I'm manually writing out the strings but if you look in any production level codebase you would never commit hardwritten code and strings to a production website like it just wouldn't happen. You'd use labels and variables, maybe use a CMS, tons of options. But I could have AI then go through and just replace all of that code, those hardwritten strings for me and place it in a labels file and then I'm good for internationalization at that point, right?

Um, so some of the comments that came in, I'd ask my team to knock out more tech debt if they had 30% more time. Of course, um, I would work more on my personal projects. I love that. Uh, Chinaza says an AI platform. Totally totally see that as well. And so today I'll show you how coding agents uh essentially the three that we're going to be talking about today. Not someday, but by let's say the end of your next sprint, by the end you'll have three predictable ways to take hours back from maintenance and put them back into your product. Um, and why this works, it's honestly uh for a variety of reasons. Number one, it's removing monotony of course as as we've discussed earlier. And so the the agenda today, we're going to be covering agents. We'll have three demos. We'll briefly discuss a very big topic right now, MCP model context protocol. So if you're not familiar with that, um I'll share some findings about that. But I will say this right here, MCP, if you're not in the AI space, I highly recommend putting this on your radar as something that you're paying attention to because we at this dot, we're having more clients than ever before. And I'm talking like Fortune 500 level investing resources into this. And so this is something that if you're not familiar with, you should really be paying attention to. And if it's something that you're, you know, maybe not used to or something along those lines, maybe give us a call and maybe we can help you out with that. Um, but I digress and let's continue. So we have our agenda there and so let's kind of talk about this.

What would you ship if you can move two times faster? I've kind of already asked this in some variation. Um, but really the idea here is kind of going to uh Travis's point here, right? Um, I'd asked my team to knock out more tech debt if they had 30% more time. That right there is a huge thing. Like I know some companies I've worked at too that did this that every quarter they would have a sprint zero to try and take care of tech debt. But by doing a sprint zero, you're literally dedicating two weeks of employee resources to kind of deal with tech debt. That if you have some way that you can start solving it in real time without dedicating a full two weeks every quarter, that's thousands upon thousands upon thousands of dollars that the company by itself is saving. Number one, which I know the average developer doesn't really think about that and they they shouldn't have to, but what about you not having to now deal with code that is for the lack of a better term duct taped and WD40 together to kind of just make it work, right? And so, um, this right here will solve a lot of those pain points. And I know as a developer, I constantly dealt with them as well, um, everywhere that I've worked to be honest.

So, what are AI coding agents? Number one, in this context, for sure, it's an agent with a chat interface of some sort, right? And so, we're talking about GitHub Copilot, we're talking about Juni, and we're talking about Rue Code. And if you're not familiar with them, don't worry about it. Um, we're definitely going to go over a lot of things. Um, it's often integrated into your environment. So you're not copying and pasting it like you would like let's say chatgbt.com or claw.ai you're literally using it the thing that it is and it's operating on high level instructions for seconds to minutes to hours in some cases. Uh, often times um there are some tools out there that will iterate continuously until it feels like it's satisfied something. Then there are tools that reach a context window limit and so once that hits is pretty much the end of that. And so, um, that's kind of where we're looking at.

So, if we, um, kind of look through some of these, uh, items, here is like some very brief points that we can kind of go over. Uh, what are the differences between the three? So, if you're not familiar with, uh, Juni is the AI assistant that in uh, um, JetBrains created. So, if you use WebStorm or IntelliJ, uh, or any of the other tools that they have, Juni is something that they offer. Copilot obviously is something that VS Code and and and uh GitHub and Microsoft offer. Um, GitHub Copilot if you use um VS Code in any way, shape or form, it is integrated within the um experience. So that way it's it's pretty seamless and easy to utilize in that respect. Root code is one that maybe if you're a power user of AI, you probably are very familiar with it. Um, if you're not, uh, this might be the first time that you're being introduced to it. When it comes to root code in particular, it is a bring your own API key. So the way that these two work is it just it's basically set up. You have like a number of requests that you could do per month or something along those lines. Um, and it is like context windowed which we'll briefly go over but root code in particular is you bring your own API key. So you would go to OpenAI or Anthropic, Claude and you would get an API key from them and you would use it. So root code is free but with the API key you are basically putting it in there and then saying I will pay for every call that it makes. Now one real big thing that we'll go over in a minute is when it comes to some of the calls that some of the tools make out there you don't really know what it's costing. So like if you use the cursors of the world or or um maybe warp wind surf etc. you don't really know how some of the credits are being calculated. With root code you know exactly how much each call cost at the time that it's making that call. So, uh, definitely different models in that respect, but with root code, we have quite a bit more. Um, one thing in particular that a lot of developers for some reason don't even know about or think about when it comes to using their AI agent tools is rules. And with rules in particular, um, you are basically saying by setting these rules, the AI tool itself will follow said rules. Um, one prime example, I hate seeing, um, like the very very obvious comments that AI tools leave, right? So um let's say maybe for example yours like hey make this hero section for me. It'll leave a comment hero content. It'll leave a comment uh made this CSS here to style this. It's like we know that this is what you're doing. You shouldn't leave this very obvious comment. So you can create a rule where it basically says hey do not leave very obvious comments and give it some examples and it will never leave those comments again. Um, there's tons of rules that you can do. So uh giving you an example um again with root code here since we're using that one for today's stream, with root code you could actually put in additional things like the architecture and the cool thing with the rules is it will utilize that with each prompt going forward and so it's never going to be stagnant or um maybe forgetting to check them. It will check reference rules all the time. So you could have system architecture designs in there. You could have very detailed things around code styling and the way that your organization likes to do things. Maybe preferences on variable naming and it will reference it every single time it does a prompt. So you never have to write that context or information ever again. It'll be there once and replicated over and over and over again which is a big deal. Then there's custom modes. This probably won't make sense until I show it. There will be sub agents which again, it's only available in R code out of the three that we're showing today. MCP support now at AI Engineer World Conference in San Francisco that I was at Copilot unveiled the MCP support, which was a very very big deal, but with root code itself, um, there's MCP support as well. I love the fact that more companies are getting that. I think it's really really good um and then we have the environments right like uh so GitHub Copilot it's available in VS Code it's available IntelliJ and several other tools as well. Juni is a JetBrains specific product so for today we're going to be showing it in IntelliJ but You can get it in WebStorm or anything but it is specific to the JetBrains tooling. They believe that they have the right mixture of prompting in order to utilize it and leverage it in the best ways possible. Root code is available in VS Code and now it's also available in Cursor. So it's available in those two but you really can't utilize it anywhere else as of right now. This may change. It is an open source tool. Um, so if there are changes or things that you would love to see in Rukode, you can always commit it. Very very active community as well. So, I've actually made some friends in that community that I've discussed things with and learned from. So, highly recommend checking them out as well. Really cool people.

So, I'm going to go through like the first demo here, right? And so, um, here's an example of the code that we have with Copilot, right? And so, just basic cursor, I mean basic editor and some things that we're going to be looking at here. If you're not familiar with Copilot, let's kind of go over some of the basics real fast. And so, here you see your um your prompting window where you would type in whatever prompt it is. You can see the current file that we have here as well. And you can of course see uh the model. So the official position right now of um this.labs is there are many models out there. Some new ones that just popped out as of right now. Uh our recommendation especially if you're thinking about this for your company is leveraging and utilizing Claude. Claude on a 4 right now. We do believe this to be uh cost effective but also we do believe that it is um probably the best bang for your buck and it also delivers the highest level of value in regards to coding logic. Um I see all the chats. Sorry I haven't been paying attention. I apologize. Um what's up? What's up Zack? How you doing? Thanks for coming out. Uh Devon, hey y'all. AI is definitely now and in the future for sure. Uh you reposted the stream. Thank you. Thank you. Thank you. Appreciate that. Uh any good places to start? Uh we'll definitely talk about some resources after the fact. Um so but officially our position at this.labs is to leverage Claude on it 4. We do think that this is probably the better models as of right now especially in regards to coding logic. There there are other things like if you're maybe ideating over um ideas or chat or something along the list, this would be completely different. But in regards to LLM models and coding logic, this would probably be the best one.

Okay. So, let's say we have this um coding logic here and we have an error. And so, I'm like, I have all these unused imports. Well, I could just tell it, hey, remove the unused imports. But I'd say, well, I'm constantly getting into this area where when I build my code and try to compile it, I'm still getting it. So, instead of doing that and just giving it a test, let me ask for additional information. So in this example here, it's like what is the best or sorry, what is the preferred way to remove unused imports in a node repository and prevent them from ever being committed. This right here would give you the exact fixes that you need. Of course, you can add a TypeScript rule of ESLint, no unused imports. And that way whenever you compile, you'll get that error and that way you can kind of solve it before you try committing it. Um, which is a great way to like leverage it. Like you have the context there, you're getting some information based on your specific needs. I think it's a great way to kind of leverage it. And it kind of also shows one additional thing. It's that the tools aren't just there to simply um just do a task. There's so many levels to what they can deliver. And so again, like I've said before, I don't believe AI is taking our jobs. But now, if I were to go research this in another way and go through multiple blog articles and things like that, there's a lot of things that I could get wasted time on where this right here would expedite a lot of my debugging process for me. And it helps me narrow it down significantly faster than perhaps other areas as well. And there's a lot of things that I'll even give you some ideas on that I personally love to do with my agents that I don't ever see people talk about, but I think this will be like the start of like some very big ideas if you haven't gone down this path. Um, we have Sebastian. Danny, love the stock. Thank you so much. I appreciate that. Um, doing great, Danny. Just getting back into the swing of things. Love that. Love that. Love that. Uh, yeah, it would have taken a long time to figure that out for sure. Thanks for doing this. I appreciate you. Thank you so much for hanging out.

So with uh obviously with Copilot we were able to kind of ask that question and do that and kind of leverage it. And so now that we know that you can now see based on what we recommended we asked them to commit it and now it's made those changes for us. Now here's one big distinction with Copilot. Once you hit that keep button that code is yours. You've committed to keeping that code with you. It'll stay there unless you make the changes yourself manually. Or based on the changes that you could see, you could hit the undo button from there. Um, but that's Copilot in a nutshell. Really great utilization. Um, here, let's kind of go over R code for example.

So, R code, if you've never seen it before, it's also I'm using the VS Code integration here as well. I'm I like VS Code a lot. I have a lot of my code snippets and whatnot there. I'm just a big fan of it. And so, here we're like looking at like let's say a registration page. I actually created a mock um example of a course um platform just to kind of show this on the Java. I do I write a lot of Java code. So I'm using IntelliJ for Java in the next demo. Um, but for just regular TypeScript I often leverage um VS Code. So here's Rode. All right. And so obviously we have some things here. So we have our prompting area where we would have um we also have if you notice here this button. This button in particular is what's called the modes, right? And so right now we're currently on our coding mode, but you can see here that there are like five basic modes that Ru Code gives you. There's an architect mode. There's an ask mode where it will just exist to answer your questions. There's a debugging mode where it's like I specifically have an issue. I cannot figure it out and it will help diagnose and fix that problem. And there's an orchestrator mode where it will coordinate tasks across multiple different modes like architect, code, ask, etc. debugging, finding, and creating the best actionable way to solve your problem. And then, of course, there's this button. Now, if you're working at a company, I don't really recommend doing this. If you're working on a side project, you can experiment if you want to. Um, but you can hit the auto approve, and it would actually let you go in there and select the things that you would want to auto approve without having to audit it every single time. Now, as you can see here, like I mentioned, co the big distinction between root code and Copilot is you have to bring your own API key. So this obviously is a blurred out version of my API key from Anthropic and the model that I'm choosing. You can see here that the input price it's telling you what everything will cost. Um and of course you can use custom temperatures, you can kind of rate limit it, whatever it is that you wanted to do and kind of go from there.

So here I've given it the task. If we'll go back, we can see for a second I've given this task and you can see here I have placeholder values. I've got string text, email address, password. I've got all this text and I'm like, I built out this UI and now if I were to go back and replace all the text with variables and labels, it's just going to be a very monotonous task that I probably am not going to enjoy anyway. So, I was like, now what I could do is let me give Ruk the task of, hey, replace all the hard code, replace the strings, put it as labels, and go do it for me. And I'm going to go work on something on the back end while you're doing this. While it did it, it took what? A minute, two minutes, three minutes. Look at the placeholder value. It's now labels.register.form.holder. It did all this. It not only did it do it for the small section, it did it for the entirety of the file. That would have taken me couple hours maybe just going through an entire look how big this file is. It would take me time to go through this entire file and replace that one by one by one and put it in a label. Put it in the label form. Like there's so many steps there. It just did it for me and I could work on something in conjunction outside of a different maybe in a different section of the codebase while it handles that for me. So now it's allowing me to multitask in a way that I could never multitask before, right? But now I want to kind of bring your attention and you can kind of see here like the labels file. You see that there's a off and then there's a login area. So you see all those labels. You see the register one and all that was done for me. All this file was done for me without me having to do anything. So, if each line is a hard-coded string that we're replacing, look at how much time that would dedicate. It's just too much, right? And so, there's a couple things I want you to kind of notice here. Number one is unlike um Copilot, uh Copilot, again, great tool. Nothing wrong with it, right? And if your company is leveraging Microsoft tools, it's a great tool for that. But let's say I am doing these API calls. Well, you see each time it's doing a request, it's telling me how much this costs. So, this one of replacing the labels, it was 0.01. 0 one penny 1.6 pennies essentially. So almost two cents. This right here, it's allowing me to see that. But now, as you see this checkpoint, anytime it does a task, it creates a checkpoint every so often. If I don't like this code, I can roll back to the checkpoint in real time to then say, "Oh, I don't like this. This didn't really work out the way that I wanted." And actually, I'll pull up my editor here, and I will show you, um, what this potentially looks like. And so, uh, let's see here. Wrong project. Give me one second. Course platform. And I'll go to Root Code. And so, here you can see this is the exact same um task that I gave it. And let me zoom in so that way it's easier to see. This is the exact same task that I gave it um in the the Copilot uh demo. Right. Well, we're here. Um, in particular, you can see here I'm seeing what these calls cost, right? Well, let's say I go to up here, it did all this work, but I'm like, you know what? I don't really like what it did below this. I could then restore this checkpoint and I can restore the files to this point of what I like. And I can see here like the things that I was changing and then say, I like this. I don't like the changes after this. Let me roll it back to this point. I can then restore those files. Or if I want to, I can restore the files and give it the task that I originally gave it to see if it goes down a different path. Right? Big big distinction there. On top of that, I could use the different modes. For example, maybe the coding mode didn't work out. Maybe it's too big of a task. I can use orchestrator mode or boomerang tasks as they're commonly known and it will break it down into many subtasks and it will execute one by one instead of trying to just bite off this massive chunk in one go. And that right there like gives you a huge difference. You can kind of see here there's all this stuff that it's kind of knocked out for me. And that's why it's like a very very big deal. But there's other things that you can leverage this for, right? Like for example, what if you wanted to really think about if you think about it like let's say you see all these things here. What if I said, "Hey, we have all these hooks that we're leveraging and utilizing. Can you audit this?" Or the other idea that I really love to do is after I have maybe a PR done, right, and I'm like, "Hey, you know, um, I'm about to submit this for code review." I can let ask Rue code, right? Let me bring this back for a second. I can ask Rue code here and say, "Hey, can you pentest my code? Can you search for any obvious vulnerabilities that I've left open?" It will probably check all the input fields. It will check everything in my codebase. And based on that, it will then tell me, "Hey, this is a vulnerability that you need to address before you submit this." I can even have it check my PR. Hey, I'm about to submit this for code review. Can you please check these changes to see if I've used the most optimal path to solve this? What would you suggest differently? And instead of putting it in code mode, you can put in ask mode and ask it that question and say, "Hey, can you can you check this out for me? What did I do right? What did I do wrong?" And from there, that right there will give you like the most opportune time. And so then if you're if especially if you're in a company where you have a habit of you submit a PR, they send it back. You submit a PR, they send it back. This could be a way to have a once over before you send that PR through to make sure like it's in the best possible position before you give it to somebody else, right? And so tons of value just by utilizing that approach alone.

All right. So what are we doing with Juni? Now if you remember this is IntelliJ tool. So Juni, this is a Java application, right? And remember what I said about like the unnecessary comments. I didn't have rules in Juni when I did this prompt. And you can see here where that just puts that right. I'm not a big fan of that. I I'm a very very big fan and um very um overexpressive and explaining variable names, function names, etc. My goal with this for the one of the demos that I did was to try and oneshot this entire application, front end and back end, as close to as humanly possible. Now um there was definitely some iterations but overall this was basically done with that goal in mind. So if I give Juni here you can see here this is where in IntelliJ you would see the extension it would pop up and I would ask it something simple right? Um, now I will say for the vast vast vast majority of things, I'm a very, very long prompt creator. To give you an idea, I like structured prompts. I like very good detail. Um, I'll take a lot of time making a prompt just to help me get to as close to the accurate response as humanly possible. And generally speaking, that will save me the most amount of time. So to give you an idea, if I have a task that's a Jira ticket and it's sectioned off for like three days, right? And I'm able to create a very long prompt and I knock out 80% of my task and I'm now adjusting the last 20% to fix it where it needs to be. That saved me two and a half days of time, give or take. That's a great use of time. So when it comes to small prompts like this, it's great for the demo that we're doing. I would personally probably never do my prompts like this. Um, but for the sake of the demo and to for you to understand the idea that we're discussing and the concepts this works great. So I would like to refactor our enrollment service Java file to follow best practices and ensure it is secure for our codebase. So kind of what I said previously like the whole um pentesting your code or finding vulnerabilities. This is a great way to kind of leverage that and do that. Um and so here kind of going over Juni as well. This is what it looks like. Here's where you would find it in IntelliJ. Here are the models that you can leverage. Now, one thing in particular with Juni, they just launched this literally last week, I believe it was when I gave this demo last. So, I added this in that moment. But generally speaking, when um Juni has this belief that uh or IntelliJ rather has this belief that they have the right infrastructure to solve their problems and so they use a different mixture of LLM models versus their own versus um the way that they wrap it in the infrastructure. Um, but generally speaking, they're letting you pick between the two as of right now. Notice after I give it the prompt, Juni's first course of action is always planning, right? So, in this, it created four major steps with several substeps, which is similar in some respect to root code's like um orchestrator mode, but root code's is like detailed in a different way. Um, but here's the plan overall. So, right now, it's getting the structure, right? So it's examining the file. It's checking related files. We're refactoring. So we're it's choosing the things that it's going to refactor. But as it examines, it may add steps here and there. And so it's working on it. So right now you can see that it's getting structure for everything. And in doing so, you now notice that a lot was added to it as it did the examinations because there was like, oh, there's a lot more that's involved here from our first glance. And so it started adding in more subtasks and substeps to ensure that um it got the right solution which I think is fantastic. And then it gives you the option at the end to build it or even test your code which I love to be honest with you. Um and and so it shows you the changes, it shows you the files, you can roll it back there. Uh but like Copilot once you accept it it's kind of yours that situation.

Now, when it comes to the this whole entire idea of MCP, model context protocol, this is like a a big big big deal right now. Um, number one, just out of curiosity, like can you drop it in the chat, have y'all heard of MCP? And for those maybe that have heard of it or maybe you haven't, have you actually done anything with MCP? I'm very, very curious. My team has well documented code practices from the actual codebase to the diff summary, verbiage, test plan, etc. I've been using Claude to do initial pass through code reviews as I review in real time. I think that's fantastic. Um, very, very good. Uh, okay. I'm seeing some comments coming in. Um, have heard of MCP, haven't used it, etc. Sure. Um, so for those that haven't heard of it, haven't used MCP yet, heard of it but never done anything with it, totally get it. I I definitely think the excitement around MCP, a lot of developers, some that you are household names like Kent C. Dots, if you noticed, as of right now, he's really, really big on MCP. Why is that? It's not just because it's the new thing, but it's the new thing that shows the most promise for organizations adopting it. Um, to give you an idea, MCP, the example that is often given is it's the USBC of programming right now. And essentially the idea was prior to MCP being standardized, it's a standard way for companies to kind of leverage and utilize um, uh, uh, prompting essentially. So to give you an idea here, whoops, I went backwards one. So this right here is a USBC. Will it zoom in or filter itself based on this? Let's see. Sometimes it works, sometimes it doesn't, but essentially same charger for your phone and all the good stuff, right? And the idea with MCP is it's created a standard that all companies can utilize and all LLM tools are utilizing. And the idea is you are taking a unique API, you are then wrapping it in this MCP and you're giving it additional context and by giving it that additional context because it now goes to the standard process, you can then plug that into any LLM or any tool anywhere and get output. So you don't have to come up with unique solutions every single time in order to get the LLM the additional context. Uh, I think this right here by itself is monumental. It's a game changer for scale. Right now teams are hacking together context and MCP standardizes it via docs APIs tickets security policies, all attached to an LLM call. So the result ends up becoming and having like context aware answers that your company data is utilizing using being used for, not the internet. So to give you an idea, I even did this um uh as a talk previously where I created a Spanish translation application, but the idea was what if you have teams that are international, right? And with those teams, you have teams in, let's say, South America, and as they're documenting their docs in their local dialect, they use some verbiage that um maybe it's very common to use within their region, but translate a word for what it wouldn't necessarily make sense. So, one example of this was um let's say you have a team in Chile and they use like uh the term or verbiage around like empanada. If you translate that directly, it could mean like a a pastry that has meat inside of it. But some may use it as to kind of use the expression of like an an American equivalent. If you break this, you will be in a pickle, right? It doesn't necessarily mean you're going to be in an actual pickle, but giving that to an LLM model, it may translate that directly. And without you having that context or understanding it, it may be very very confusing in order to get that. So instead of the company itself leveraging that MCP model um and being confused by it, what they could do is uh I mean uh the LLM to translate it and being confused by it. They could prompt the LLM with their own documentation and the LLM the LLM will exclude any external internet sources for additional context and it will only refer to the documents that you give it. This right here is the game changer. So if you're for example if you were to say hey I'm trying to solve this problem and you were to prompt Claude or Anthropic um with that problem it may search Stack Overflow the internet blog articles etc. But now if I only give it context around my codebase and then give it documentation architecture etc. and I say I have this problem, it is going to be more accurate and be able to pinpoint what that problem is and reference the docs that I give it instead of anything on the internet because anything written on the internet could be anything, right? It doesn't have anything specific to do with my organization. This is the way that you kind of beat around that. And so some examples would be conversational search and filtering. So maybe you're using natural language queries over structured data and catalogs. Like that's a big deal. Um AI powered dashboards. So, Copilot on Microsoft 365 to kind of generate contextual insights across documents. So, it's not just saying, hey, you use this for X% well, what does that what does that mean? And so, um it could be a great way for the internal organization to kind of understand more. Um context aware assistance, this is a big one. So, like Notion AI, for example, it recalls workspace context to help you write, edit, or answer follow-ups or um one that I'm seeing common right now is in Gmail, right? Like Google has put Gemini pretty much in everything. And so now as Gemini is writing email replies, it has a lot more context, especially if it's a thread of emails to write a more profound response than something that's very generic. Um, Jira, for example, um, create agent tasks by pulling tickets through MCP. That right there would be a massive one. An example that I heard recently was um imagine you have somebody in your organization that works in HR and they are um buying uh lunch for like everybody on the team. But let's say people on the team have allergies like shellfish allergy, nut allergy, etc. You would have all that in documentation that the uh HR could kind of reference and as they're placing this online order, it could tell Claude, hey, I need to place this online order for 100 employees. Their likes and dislikes are in um this document. Reference it. And it would be able to then tell, well, Zach here is an employee, but he has a nut allergy. So, make sure that whatever dish I order him abstains from nuts. But if I was just giving a generic likes and dislikes, it may not even see that, and the prompt would just go off of what it would condition from from basic internet information.

And so, at this point, I basically implore you to have five experiments to try. So for your next code review, if you've never done this before, maybe ask the agent to perform a review alongside yours. Or when you start working in a new area of the codebase, ask the agent to explain it. And when it's doing so, find out more. Uh if you work with TDD, I'm a big fan of TDD, try the agent as a pair to perform one side. Um when starting your next task, for example, create a plan and use that agent and just see what happens. Once you have a plan, ask the agent to implement it. And at this point, I would honestly ask if there are any questions, any comments, anything top of mind that maybe um someone wants to know additional information about but that may not necessarily have it. Um and we can kind of go there and uh I do know that it needs a lot of thought around security. Yeah, for sure. Especially if you're working in an organization, you definitely need secure practices around that. Um there there's there's no beating that one. Uh and our official stance at this dot is um we often like to work within the guardrails of the security team's suggestions. So don't try to counteract it, move around it or beat it. Often times they've given you practices for specific reasons. You should probably follow that um instead of trying to find a way around it because uh and this is kind of the whole reason that we're even doing these AI lunch and learns. It's to educate, right? Because um if a company has a stance of only giving you the most basic AI tools or no AI tooling at all, you have shadow AI in your company without even realizing it. Um one big proponent of that is I can tell you right now they're leveraging AI and your stuff is getting put in there. Um so I would highly highly recommend giving them something that they can utilize and leverage. Otherwise it would be very foolish to assume that your engineers are not touching it in some way, shape or form. It's being it's being utilized. I mean, I I believe uh Copilot just announced recently that uh what is it? 18 million active users and companies uh are leveraging it and utilizing it. That's the ones that have access. How many do you think are using on their personal accounts like you need to think about that? Um so I'm a big fan of and proponent of avoiding that shadow AI issue. Give them the tools that they need. Um put security guard rails around and make sure that your stuff is not going in there. And if you're an engineer that is leveraging these tools, please whether it's even on company specific ones, be smart and remove PII from any of the prompts that you're using. Don't put in company specific information. Don't put in patient information. Don't put in customer information. Remove all that. Mock it all. Like, put a John Doe, John Smith kind of situation in there. Don't put actual customer info that would be bad for it to go out there. Um, people don't realize it, but you, as much as you feel like your LLM model is yours, um, those records are kept. It can be subpoenaed at any time. It could be exposed at any time. I wouldn't trust it in that level. Yeah, I was just talking about this literally at a work meeting this morning. How many subscriptions are you personally paying for outside of work for AI tooling? Um, for this dot, we are very very very heavily invested in AI. And so we have a lot companywise, personally wise, I have a lot. So much so that I've actually literally canceled Hulu recently just because I was like, I got to justify a new subscription and I just can't justify it. Uh, David Martinez, dude, yeah, I'm a very senior designer and I'm using these tools for content strategy and even for coding prototypes. Absolutely. Uh and I think that's one

of the best ways to do it. Uh, one prime example, like I think the speed to getting the right answers is a huge leverage here, right? So if I'm having a meeting with, let's say, strategy team, product team, whatever it may be, and they're like, "Hey, we want to go down this initiative." Instead of us just ideating and scribbling and LA, let me just leverage Vzero or something like that. And by leveraging it like that, I can now build up ideas that we can kind of go off of. And I think that right there is is and you can see like, hey, this is not actually the way that we were thinking about it. You just saved me so much engineering time by clearing that up because now I don't have to go code this thing and fix it. Like you can specifically tell me, no, I don't like this feature. No, this is the feature that I was talking about instead. That's the perfect way to prototype with a live visual representation that you make in a couple hours. Like Vzero, I'm a big fan of Vzero. I think it's one of the best ways to prototype as far as I'm concerned, but there's so many other tools out there too.

Um, even if you were just to leverage AI in one way, shape, or form, just to get it out there. Um, do you do these sessions every week? It's my first time joining. Uh, on my personal stuff, I do. I try to always do some kind of educational thing in one way, shape, or form. Um, I need to do them more consistently for sure. Uh, but um, on this channel we have a lot of educational content. Um, if you go, so if you're watching this on YouTube, you're probably watching it on the This Media channel. Um, this media is the media arm of This.Labs. And so we put out monthly workshops on here completely for free. If you go through and look through what's called JS drops, there are some massive JS drops on this channel in particular. Um, some really, really good ones if I'm being honest. There's some for, um, matter of fact, let me go through and and tell you, there's like one of my favorite ones. It's actually by Debbie O'Brien. It's an introduction to Playwright from end-to-end testing. Uh, phenomenal one. Or there's another one that we did. It's an introduction to SvelteKit. It's like a five-hour long workshop. Or I'm getting started with PWAs and Vite. There is an advanced master, um, uh, mastering advanced Nuxt.js. There's so much JavaScript related workshop content out on this channel. If you haven't leveraged it in the past, subscribe right now. Check it out. I think you'll love it. Uh, I think one that we did recently that was pretty good was, um, create a website from scratch using Nuxt Studio and Nuxt Content and, uh, Nuxt UI, I believe it was. Um, but there's a lot out there. So there's a lot of React content. There's a lot more coming out. We try to do them at least once a month. Oh, there's a, I remember there's an advanced TypeScript JS drop from, um, um, um, why is my brain going blank? What's his name? Dude's phenomenal. Josh Goldberg. Josh Goldberg, dude. I love that dude. Um, he came and he did a JS drop and, um, phenomenal JS drop. And so if anybody's ever interested, uh, in attending those, we do them once a month. If you subscribe, you definitely be notified by them. And if you're not on our, the this newsletter, we often reference that as well. So that's like a really good place. Also, if you are a woman in tech, we have our women in tech meetup that we do once a month for mentoring where, um, successful women in the field of tech kind of mentor others to kind of help build them out. Um, but there's also, and hey, if you're not coming, you probably should. Hopefully, we'll see you at the Commit Your Code Conference. If you haven't heard about the Commit Your Code Conference, it is a tech conference that I do every year now, I guess. Um, last year was the first year, this is the second. Uh, we have 125 speakers this year. Um, it is completely for free online. So, if you don't live anywhere in the Dallas area, you can join for free online, but if you with the option to donate to charity if you want to. Um, but we do the conference for $49 a ticket, uh, instead of $700. Uh, and we donate 100% of all the ticket sales to charity. This is a huge part of that as well. They're going to be there. Um, I'm going to obviously be there since I'm the one hosting it. Um, but we have a massive, massive conference there. Um, so if we could see you in Dallas, I'd love to see you. It's a two-day conference. Tons of external events, tons of networking. I think we have four external venues outside of just the conference itself. Um, more things are going to be going on. So, tons and tons of ways for people to hang out and kind of check out.

Such a good talk. Thank you. I appreciate that. 100% goal. This plus passive videos. Yeah, for sure. Um, thoughts on balancing when to use the tooling and when to leave it alone. At Meta, we obviously have extremely complex systems. I'm still trying to find the balance. Oh, yeah. I mean, for me, I'm on record on our podcast, by the way. Um, this has two podcasts that are phenomenal. If you haven't checked them out, I highly recommend it. We have the Modern Web podcast. We have the Leadership Exchange. Modern Web is like a top 30 in the world for JavaScript. Um, and then I have a personal podcast as well, but I, I'm on record on pretty much all three of those podcasts saying I'm anti too many levels of abstraction. I think once you get too far away, you lose the meaning, you lose the value in that abstraction. It's not great. And so for me, um, once I'm too far away and I feel like the the abstraction level is too far to where I don't know what's going on, I don't like where that is. I want to still have oversight, checks, and balances, and measures, and control over whatever that output is. And so whenever we get into that territory of it's a magic black box that's just doing something and I don't know what it's doing, I don't feel comfortable there. I want to still have oversight over what's happening. I want to be able to prompt accurately and control the output. And so, um, whenever we get to those levels, that's kind of where I pull it back. Um, but when it comes to like systems that are very, very complex in it of itself, when you look at something as a broad picture, that will always be complex. So I go more granular and I try to go, uh, keep it very, very tight. Um, so instead of just looking at an entire microservice, maybe I look at one subset of it to see what's happening in one region before moving to the next. That's kind of the way that I would maintain those guardrails kind of going and navigating that.

All right, y'all. This has been fun. If you enjoyed this, please share this out. Retweet it. Re-share it. Repost it. Share the link with somebody. You never know who could benefit from it. Obviously, it's free, right? So, you can't go wrong with that. Um, and hopefully, I'll see you all around. And if there's other topics that you would love to see a lunch and learn on, maybe, uh, drop them in the comments, especially if you're on YouTube, check out the This Media channel if you're not, but drop them in the comments and maybe we'll do one on that for you and maybe we'll try and do more of this for the public. All right, y'all. It's been real. It's been fun. And I will see you on the next one. Goodbye, everybody.