Agentic Development Stack: Harnesses, Models, and Automated Workflows

The tech and the toolset that I use to build my AI applications continuously changes. We're getting new model drops, we're getting new tools, and that's why in today's video, I'm going to give you an update on all the things I use. I'm talking harnesses, I'm talking models, I'm talking tools, I'm talking tech. I'm even going to tell you what diaper brand I use for my child. That might be too much information, but either way, we're going to get in all of that today. Everything that I use to build software, you're going to know, I'm going to share with you, and I'm going to explain how, why, and all that. Let's get into today's video. So, first things first, let's talk harnesses. Now, I'm going to keep it a buck 50 with you. I use all of them, right? I use T3 Code, I use Claude Code, I use Cursor, I use Open AI, I use Devin. I use the CLI's, I use the apps, the GUI's. Now, part of the reason why I do that is so I can be informed and when I create my content, give you the best recommendation. Part of that is also this is what I find fun and it's a little sad low key. But other than that, I'm just going to share with you how I use these. So T3 Code has been my daily driver. Now T3 Code is unique in a sense where it's not a harness. It's a harness's harness. Meaning, when I use Claude, when I use Claude on my T3 Code, it's using the actual Claude Code harness. When I'm using Cursor, it's using the actual Cursor harness. When I'm using open code, whatever harness I'm using, it's actually using that harness. So it's sort of like this meta harness, but the reason why T3 Code shines is it allows me to work on the different machines in my network. Right? So I have a Linux box, I have a Mac machine. By the way, stay tuned, make sure you're subscribed, because one of the videos that I'm working on is basically showing you guys how I'm building a mini data center in my house. Gonna share with you my network and all that stuff. But, yeah, as you can see in my network, we have a Linux machine, I have a Mac machine, and I have the MacBook Pro that I'm using right now to record. These are the machines I use to build applications. And with T3 Code, it makes it really easy to connect to different machines in my network and to disperse agents on different machines. And the GUI is pretty nice. Another thing I really like about T3 Code is the sidebar. Right? The sidebar, in my opinion, if you're working on multiple projects and you have multiple threads, I like how it almost looks like an inbox. Meaning, first and foremost, I know if work is not done, I can settle a thread, meaning it's done work, or it automatically settles when the PR's been merged, right? I can tell what's done, right? So, I don't need to click and check to see if the work's done, it tells me that it's done. So, I genuinely like the UX of T3 Code because it's actually being built by people who mass spam this the way you and I do. So, big ups for T3 Code. Now with T3 Code out the way, let's talk about the Chat GPT app or Codex. Now, I do use Codex. It's more so for computer use stuff, research, and all these type of things, or whenever I need to do anything 3D. For example, one of the videos that I have coming soon is this hardware startup that I'm building, you can see I have all the tech right here ordered, ready to build. I'm just waiting for a secondary table to come so I can build it with y'all. But, when I was originally exploring this startup, it was through Codex's computer use using Blender and then like going back and forth with Astra. So, if you have an open AI subscription, honestly, Codex is great, especially for these type of things where you're either building a hardware project or you need some computer use work type done. Codex is great. There's no reason for me to not use Codex when I'm not using Codex to build applications. It's usually, like, you could see on the sidebar, I have it for, you know, budgeting, right? I have it for improving my setup, like, I built out my room in Blender and I asked it, okay, how can we improve my setup and the lighting and all that stuff, right? So, Codex has that type of place in my heart. Do you want to know what's being used a lot? These expensive diapers by my baby. So, let's hear a quick word from today's sponsor, please. There's a few people in this world that I can talk about my baby, but also explore different designs for my web and mobile applications. And no, I'm not talking about my wife, but yes, I am talking about Grockbot. And honestly, I think it's better that I show you versus tell you. This is my Grockbot. I have six bots. There's three I want you to focus on: engineer, designer, and baby. The baby bot helps me by answering all the questions that I have with dealing with a newborn, first time dad here, Grockbot to the rescue. But with designer and engineer, let me just show you my workflow. What I'll do is let's say I find a website that I really, really like, I will send a screenshot to designer bot, and I will basically tell it, hey, I like this, create a similar design on Paper. Paper is basically a Figma alternative, but really, really good with AI agents like Grockbot. Grockbot got to working, and after a little while, it got me my design ready to go. Now, here's the sick workflow. I haven't even left Grockbot yet. What I'm then going to do is I'm going to create a group chat between designer and engineer, my two bots, and I'm going to tell designer, "Hey designer, can you pass off the designs to engineer bot so we can start implementing the website? Give me a deployed version." So designer tells engineer, engineer gets to work, designer checks engineer's work, and then I finally get a URL with my site deployed. What's also cool is each Grockbot has its own computer. So I can tell it to download apps and sign in and log in, it can access all the information I want to give it. Now, finally, I just want to show you that Grockbot isn't just a toy. You can get real work done. In this case, I had six PRs from a developer that had to be removed. I told Grockbot to use Cursor cloud agents, and I got six PRs merged without leaving Grockbot, and I was using my phone. You can see that all my PRs got approved, some had merge conflicts, after the merge conflicts were fixed, I merged them all. It even waited for Gremtau to give me a five out of five. All of this engineering work, designer work, baby work all happened on Grockbot without leaving it. This is only possible with Grockbot or a wife who's an engineer, but let's stick with Grockbot. Make sure to check them out. The link is in the description. Let's get back to the video. Now, let's talk about Claude Code and the CLI. I almost never use the app. I'm sure they made some changes. I've almost never used the app. The last time I used it, it was terrible. Now, I will use the CLI, and the reason why I'll use the CLI is, maybe I'm SSH'd into a different machine, a new machine, and I need to set up some network config. Claude is great, especially Opus 5.5 is great. So whenever I'm using Claude, I am literally using the CLI. I'm never using the GUI. And you have Cursor and Devin, and for both Cursor and Devin, I am using the GUI. You can use their web apps, I usually use the web apps. And what I use both Cursor and Devin for is the cloud agents, right? If I need to spin up like tons of work on the cloud, like, I have my own cloud, but let's say I have already stuff running here, and I need it running in a cloud agent, that's what I'm using Cursor and Devin for. Cursor, I actually do use the CLI as well, but the cloud agent product is really impressive. And how my workflow for that actually works is I have it connected to my Linear via MCP. So if I want a cloud agent to work on something, all I have to do is literally tag either Cursor or Devin, and then it just works on it. So for, like, some of our apps that are live in production, people are using it, and a bug's reported, I'll get a cloud agent working on it ASAP before I even read the bug. So by the time I'm getting ready to understand what's going on, I already have a PR ready to merge. So all the different harnesses have different strengths. If you were to ask me, which one should I go with, I would go for the one that is the best in terms of cost, right? Where do you get the most subsidies? Now, it was Open AI for a minute because we were getting resets, but with the new Opus 55 model, it low key is entropic. Either way, I think you should be very disloyal and not have to, like, stick to one subscription. Keep switching around, you know what I mean? Get the best bang for your buck. Now, that was everything harnesses. Let's talk about models. Well, in terms of models, there's only two models to use when building applications. And for me, it is literally Opus 55. Opus 55 is by far one of the best models, not because it's this giant leap in intelligence over Astra or Fable, although it is better. It's just a well-rounded model. It listens to instructions, it's really good at long running tasks, like, it, it, it's just an all-around great model. Like, ever since Opus 55 came out, I haven't seen the need to use another model. Now, Astra is pretty good for really deep work, for research, for, maybe, understanding a code base deeply, but I'll be honest with you, I'm pretty happy with Opus 55. Now, when it comes to models, this will most likely change in the next two to three months, right? Because a new model will come, it'll be better, it'll be next level, and we're going to go flip through that. As of today, September 2026, Opus 55 and GBT 6 Astra are the best models. I'm using Opus 55 more. Number three, Tech Stack. Now, this used to be a conversation we used to have often back in the day when we were writing code by hand. What framework are you using? What tech stack are you using? Are you using React.js? Are you using Next? And we used to talk about this all the time. Now, AI came, and it changed everything. But I still think there are smart decisions to be made on the tech that you use to build your applications. I'm going to share with you step-by-step what I use. My framework of choice is SvelteKit. I use Svelte more than I use React. It's honestly between SvelteKit and sometimes I'll use Next.js. Both of these work well with Vercel, which is where I deploy my applications, but the reason why I use Svelte is a, AI is really good with Svelte and SvelteKit, but b, I just find that any app that I built with Svelte by default without even me telling the model to optimize for performance, by default the app is just fast. It's snappier, it's quick, it's just a well-performing application and because Svelte in my opinion is a superior framework. So when I'm building web applications, I am using Svelte. Now the back end of this, I'm going to get flack for this because I work at this company, I use Convex. And the reason why I use Convex is AI is great with Convex. It's fully type safe, everything is real time, right? I don't need to manage web sockets, it's fast, it's snappy, it works well with Svelte, there's a library that works perfectly well with Svelte. But that's not the only reason why I like Convex. Convex has this thing called Convex components. And what Convex components are, you can think of them as building blocks, or third-party integrations, that are all code, that can simply integrate to your Convex app. A simple one is, this is the one I built, the Stripe component. I can literally give this link to my agent if I'm building my app with Convex, and say "Set up Stripe using the Stripe component," and it will. And let's say, I need Sandbox, it'll set that up as well. Let's say, I need Agent mail, Posthog. Let's say I need browser base. Whatever component they have, like, WorkOS is my favorite auth provider. It's the big leagues auth. I literally just give it the component, and it sets it up and it works right off the box. Like I have not had to read code when it comes to setting up Convex components because AI just works with Convex, right? So, Convex is a pretty big part of my stack, right? So, Convex is a pretty big part of my stack. A new addition, well, hasn't been new, but I just haven't talked about it, is whenever there is a service that I'm building, I am always and forever going to be using effect. I always wanted to get into effect, but the syntax and like the learning curve felt very verbose, but then AI came along and made picking it up very, very easy, writing it very, very easy. So, anytime I have a service, actually, a great example is Bezelo, is basically my app is one giant MCP server. The MCP server is all effect. It's entirely effect code and the website is Svelte, right? So, for any type of back end, external service, like, whatever it is that I'm building, that's back end related, it is always going to be using effect. The big reason why is effect is magnificent at handling errors and that gives me and the AI agent an edge in making sure we build apps that are not heavy with errors, right? For example, let's say you build something with JavaScript, there's so many places where things can go wrong, right? And this is why we use TypeScript, right? Type safety, types are necessary. It basically is the language in which we define a function and we know if code is being written the proper way. But the problem is this hasn't applied to errors in TypeScript land. Effect makes handling errors easy. I don't build an application that does not use effect. Now, let's say I'm building a service and not using Convex for the back end, there's two databases that I use, really, there's one, and it's PlanetScale. PlanetScale is the big dog Postgres provider. You can also use PlanetScale through CloudFlare's Hyperdrive, I think that's what it's called. But yeah, anytime I need a Postgres database, I am using PlanetScale, works very well with Effect. Now, when it comes to deployment, there's two places I deploy, Vercel and CloudFlare. Now, what's interesting is if I'm ever using CloudFlare, I'm probably using the entire stack, meaning I'm using Workers, I'm using R2 for storage, like, I'm the R2 storage bucket, I'm using Hyperdrive, I'm using Durable Objects, I'm using the entire CloudFlare stack. But, if I'm not using CloudFlare, then the app is going to be deployed in Vercel. Now, an app that I have deployed on CloudFlare is Goshen Email. I'm going to open source this soon, but this is basically an email provider for your agents, right? You were using this for Bezelo and Pluto, and what's cool about Goshen Email is it's fully running on CloudFlare. It's using CloudFlare Email, it's using Workers, it's using Hyperdrive, it's using R2, like, the entire app is running on CloudFlare. CloudFlare is an amazing infrastructure product, so is Vercel. Depending on what I'm building, if I'm using anything CloudFlare related, I'll probably stick to CloudFlare, if not, I'm going to be using Vercel. Svelte, Effect, CloudFlare, Vercel, PlanetScale, and I forgot Convex components. Basically, all these things, I'm not using all of them all the time, but I'm using at least one of these to build my app. This is the stack that I love to use. Now, tools, tools. What tools am I using? I'm just going to go through each one one by one. First and foremost, you need a code reviewer. I'm using Greptail. I love Greptail. Greptail is confidence score and their GrepLoop a skill, which is part of my skill's repo, if you have not checked out, make sure to check out. Is Greptail's GrepLoop skill, right? And this basically based on Greptail's feedback, loops the agent to get a 5 out of 5 score. Basically, I want to just show you an example. Let me open this pull request, and you can see here, a PR is made, Greptail gives a confidence score. The GrepLoop skill basically makes the agent iterate on Greptail's feedback until we get a 5 out of 5 score. This is just simply excellent. I love the feedback Greptail gives me and the coding agent that I'm using. It's one of my favorite tools. Another tool that I've been using is Orgo. When I, like, for Pluto, all every Pluto instance has its own computer. The computer we're using is Orgo. Now, just to make very clear, I am an investor in Orgo, but the reason why I invested is I use their product so much, I hit up the team and I'm like, "Can I write a scout check?" and they accepted. So, take that with a grain of assault, but I do use this in Pluto and whenever I want to give my agent a separate computer, I'm using Orgo. Another tool that I've been using a lot lately is Paper. Paper is basically a Figma alternative that is great with agents and I've been using this to design my portfolio site. I've been doing this to design Pluto. Pluto got a complete redesign and I will be honest with you, it wouldn't look this clean if it weren't for something like Paper. Paper basically takes me, someone who's not a designer, and makes me feel like when and allows me to build and explore different design styles all using AI by the way. They have a great MCP where you can connect your agent with Paper and it can literally create designs on the canvas. Another tool design related is Mob and Mob is basically collection of all the most beautiful sites that exist, and it's presented to you in one URL, they also have an MCP server you can connect your agent to. This is my favorite place to collect inspiration and build beautiful looking apps. Not as beautiful as you, but I'm trying. I use Whisper as my voice to text app, pretty self-explanatory, great to have especially when you want to yap a lot to your agent. I rather yell it than type it unless I'm trying to be specific, then I'll type it, but Whisper flow has been great to me. Now this is maybe getting a little too much if you're solo, but I manage a team, so we use Linear for tracking issues and work that has to be done and the status of that work. There's a lot more boards, this is the only stuff that I could show you. And what's cool about Linear is again, they have an MCP, you can connect your agents, especially Cloud agents, and I can fire off Cursor or Devin by just selecting them here and assigning them a ticket. It's splendid workflow. In finally, I'm going to share with you my skills. Again, I've done a whole video on this before, but for those of you just watching, my skills are open source, my entire software factory, my workflow is here. Basically, I'm going to explain to you in very quick detail how it works. Step 1, when I prompt an agent that has access to my entire skills repo and agents.nd file, when I prompt an agent to build a feature, it first isolates, creates a separate work tree off of main so my agents can work in parallel, so I can work on 15 different features, the agents won't be colliding against each other. Then it's going to start writing code and it's going to use a skill called code structure which structures the code in a way that I like for it to be written. And then after that what's going to happen, it's going to prove that the solution it presented actually works by recording a video of the solution if it's UI facing. If it's not, what it's then going to do is it's going to run some tests to prove that said functionality has been done. But let's say your machine doesn't have the ability to record a video, it will take a before and after screenshot and place it in the PR. What's once the PR has been made, Greptail, which is my code review of choice, will run in the PR and it will give a confidence score. If the confidence score is not a 5 out of 5, the GrepLoop skill is going to run and this is basically a loop that's going to take the Greptail feedback, feed it back to my coding agent and saying address it and fix it, and it won't stop until it gets a 5 out of 5. By the end of it you get a PR with a confidence score of a 5 out of 5, ready for it to be reviewed by you and merged. Now you know how I use the different harnesses, you know my favorite models, you know my tech stack of choice, you know the tools that I use and I've even given you the skills that I use to build software, there should be no reason why you're not in the kitchen cooking it up. So make sure to take advantage of everything I've told you. The links to my skills are in the description. For the other tools, you're just going to have to Google them. I'm not about to copy paste all the links. You've been awesome, my name is Ross, thank you so much for watching this video, make sure to like, comment, subscribe, with the notification bell, let me know what else you like to see, peace.