#65 He Replaced 100 Employees with AI Agents (Here's how) ft. Ryan Carson, 3x Founder & AI Builder

I think really the difference is leaning forward, right? So instead of leaning back and just saying, sure, whatever, it's leaning forward. It's getting to the point where, you know, everybody that I know that's really smart doesn't write code by hand. For me, it happens at 1130 p.m. every day. And then after that, it runs another, essentially like a Ralph loop. And it basically looks at key metrics of my app and then it picks the largest problem. So I use mostly skills now and not MCPs. I think we're moving to a reality where there's going to be a lot of solo founders. So I've built three startups. The last startup I built had over a hundred people, but it's just me. There's still no shortcuts, like just because you have agents now, it doesn't mean there's magic shortcut to making millions of dollars. You still have to do the work, right? It's just that the work is now higher level. Hey, everyone. Welcome to the Founders Foyer with me, Aishwarya. This foyer is full of conversations. The space where creators, founders, and builders look for all the support and concepts to grow their ideas into products. AI coding largely means pasting prompts into random chat boxes for most people, but a few of them are quietly rebuilding and prepping themselves for a futuristic software building, I would say. Ryan Carson has been living in this future that most AI builders are just now waking up to. After spending a decade helping over a million people learn to code with Treehouse and also serving formally as a builder in residence at AMP, he's simply obsessing over one simple question. What happens when a coding agent just doesn't auto-complete your ideas, but actually ships production code with you across your real repo night after night? In this episode, we'll dig into how his structured thinking and ship code while you sleep, and I think it's really while you sleep. Loop actually works in practice, why he thinks unstructured vibe coding is holding most people back, and how his new reportees, aka automated workflows and cron jobs, can actually reshape what it means to be a one-person company. Ryan, I'm so excited that you're here. Thank you so much for joining. Thank you. Excited to have the conversation. So looking forward to it. Absolutely. Ryan, you've moved from teaching humans to code, which you did formally with your company with Treehouse, to now teaching AI to code with humans. So how does it really change your perspective around what it means to even code at today's perspective? It's wild to be alive right now. I crossed the generations from being pre-internet to internet to pre-mobile to mobile, and now to AI, and it's just, I can't believe it all happened in my lifetime. And it really is astonishing. I think all of us who are using Opus 4 or 5 now to code are just blown away at how good it is. So it's pretty, it's an amazing time to be alive. I often can't believe it's real. I think Ilya said in his interview with Dvorakash, can you believe this is all real? Can you believe it actually works? And I think all of us are astonished, predicting the next token turns out to create intelligence So here we are. I know, here we totally are. And you're right. I mean, this has been such a wave across decades, right? I think every decade we had this aha moment, and then it was first with hardware, and then it was, oh, like SaaS and software, and then there was mobile, and then now it's just, I would say the decade has come shorter with just a couple of months, couple of years, and then a couple of weeks to days right now. So it's insane that with every builder, I keep telling them that, what do you mean even product launched? It's just something that's going out every second. So it just feels like it's a wild time. So for someone who's just learning to code today, let's take these high schoolers, let's take folks who are just entering the whole job scene in this space. Are you seeing something different from how it used to be? Of course, we all know that there's AI agents that's helping us code, but beyond that, has something been on your mind, especially because you've been on the forefront. What have you seen this change across? So is there like a shift in perspective you see, especially for the novice folks starting? Yes. I think the most important thing right now is to be agentic yourself as a human. And what I mean by that is to have agency, to be the type of person that says, I don't know how to do that, but I can figure it out. So now if you have that mentality, because you have an agent to support you, you can really build almost anything. So I would encourage people that are listening to the show, if you don't code right now, all I would do is pick a tool. And there's a number of them. So I was at AMP for a while. I love AMP. It's great. I'm not being paid to say this, but we offer a free plan. So just go to ampcode.com and sign up for the free plan, and then you get 10 bucks a day and free credit. So just do that, and start talking to an agent. You could obviously start with ChatGBT. You could start with Cloud Code. They're all kind of the same in the sense that there's a smart agent out there that wants to help you. So get the tool and use it. So that's like thing one. And we've been trained for a long time to feel like we need to ask the right question or we'll look dumb or maybe we won't get the right answer. But now that we have agents, you can just start. So people think, well, I don't know what app to build, or I don't know how to build an app, or I don't even know what that means. Well, just say that to the agent, and then you'll end up at a point where you are building stuff. There's no such thing such as a dumb question with agent. It's just going to help. No, they have infinite patience, and they're friendly and kind, unlike humans. So I would start there and just train yourself to be honest with the agent about if you don't get it. If it starts saying something that feels too advanced or confusing, just say that. And then you end up being able to learn almost anything. A practical example of this is that I started off with a computer science degree. I was coding by hand, and then I started companies and started hiring engineers, and then I got abstracted away from the code. And I wasn't really a developer, I was a CEO, founder. Since then, I've obviously gone back in and learned how to code again as an engineer. And now I feel like I'm a pretty comfortable, accomplished engineer now. I understand what's happening because I've asked the agent so many times, how does that work? I don't understand what that means. And now I understand what it means. Yeah, no, totally with you. Super practical tips. And I love the way that you phrased it as agentic, because I think the minute something feels like an agent, and it has AI in it, people just try to complex the whole situation, the whole scenario. But then the fact is that most often it's just our likability to start, and it's just not about AI in itself, but just our own hesitations that's just leaving us from getting started with the conversation. So yeah, absolutely with you on that. Amen. As we talk, Ryan, about starting these conversations with, let's say, chatGPT, your codecs or AMP and whatever agent that is available, most people claim that they do white coding, right? Which is, you know, just chatting with AI without some sort of a structure. They just go about asking things, which I think is very good, because as we just spoke, it helps clarify a few things. But is there something like a good session and a bad session with AI agents? Especially, you know, what does a bad session look like? And can you walk us through that? And also, what does a good structure look like? So I think there's a big difference now between vibe coding and agentic engineering. So I would say they're very different things. And so if you just want to have fun and just want to build and you don't care, vibe coding is great, right? So if you think about, like, I just want to make a game, I don't care, you know, how it's coded or how many people are going to use it, I just want to have some fun, just go with it, right? And that's vibe coding. But if you're building real stuff, you're really doing agentic engineering. And I think there's a big difference in that. So what is the difference between the two? I think, really, the difference is leaning forward, right? So instead of leaning back and just saying, sure, whatever, it's leaning forward. And as soon as the agent does something and you don't understand why it's doing it or what is happening, you start asking questions, right? You might say, well, why are you doing that? Or how does that work? Or you know, what happens if? And then what happens is you start going into agentic engineering, where you're really collaborating with the agent versus just letting it, you know, do whatever it wants. The thing to remember is as brilliant as Opus 4 or 5 is and the agentic harnesses like AMP or Cloud Code are, models don't have what we would call sort of like common sense. They're not able, like humans are, to have this sort of medium to long-term context around the project or what you're trying to do. And so they often make what looks like dumb mistakes. So a good example of that is, you know, Opus 4 or 5 doesn't spend a lot of time looking through your code base to make sure it understands all the code before it starts editing. It will usually just start writing code as soon as it thinks it understands. And that can cause all sorts of problems, right? So what you should do is like learn slowly, talk to the agent a lot, get more and more comfortable with your code base. And then you can start asking good questions and saying, well, why did you do that? And then often the agent will say, oh yeah, good point. I hadn't thought of that. So that's the big difference between vibe coding and agentic engineering. And everybody should be moving towards agentic engineering. Like that's where the jobs are. That's where the future is. You know, I think vibe coding is just kind of a fun thing to do in your free time. Yeah. No, it also sounds more like the same person can actually oscillate between the two and choose to be the vibe coder when the necessity is and, you know, doing something for fun, like you said, versus I think the alternate sounded more like being in command, basically like trying to just be out there and then trying to decide, which you made a very interesting point about the common sense part of it, because I've heard a lot of them say, hey, these models are either like genuinely too smart or too dumb. It's just like behaving like probably like a preschooler, or it's just behaving like a PhD student. And then there's like no in between. So yeah, definitely. I think personally from using some of these models across and also looking at the way that each of them performs for certain specific actions, like certain like, you know, Opus is really good for design versus a Gemini that's really good for something else. And there's this also like demarcations between these models. So from what you say, it's also looking like you need to be constantly talking to these models and sharing what your input is for it to give you like a very contextual output. So that's what a good structure would look like there. Yes. Yeah. Engage your brain. You know, don't outsource your thinking. That's the primary difference between agentic engineering and vibe coding. I think the other thing is people should really look into concepts like compound engineering, which Karen Clarkson from Every introduced. It's pretty simple. The idea, and it's actually a skill that you can install for your agent, or it's a plugin if you use cloud code. And it's pretty simple. Like you have this idea of at the end of a session, you basically say, you know, compound the learning. And what that does is it caused the agent to go back through the thread and look for things where it ran into gotchas or it made mistakes. And then it updates either agents.md if you're on, you know, AMP or codecs or anything except for cloud code or cloud.md if you're using cloud code. And then you start to have less of the frustrating common sense pitfalls where you're like, oh my gosh, why didn't you know that? Well, because you haven't written it down in your agents.md file. That's why. So I think once you use compound engineering, you start to have less of those common sense pitfalls and then you go on top of that, which is compound product, right? Which is something I'm toying with, which is this idea that every night you run two cron jobs. The first cron job is a compound learning loop where it's a bash script and basically it fires up your agent and it says, okay, look at all the threads today and use your compound engineering skill and look for things that we need to put in agents.md. So it does that. For me, it happens at 1130 PM every day. And then after that, it runs another, essentially like a Ralph loop. And it basically looks at key metrics of my app and then it picks the largest problem that we have with the app and then it solves that problem and then it submits a PR and then I wake up to a PR in the morning. So it's like this idea of compound learning and then compound product, that's how you start to really turn yourself into an extremely productive person. Right. So compound learning is more to do with prepping your agents and the model stem cells. I really like the phrasing to it. In some sense, there's like this hack that I personally use as well on all of my projects, which is there are certain times that there's always like an instructional.md file or a, you know, like an agentic file that it refers to certain times when I'm just riffing with the models and then trying to make it understand really why I'm in command for, you know, executing certain things. I would ask questions in between the conversations for which I'll ask the model to keep summarizing what it understood from my conversation. I'll basically copy all of that, put it into a .txt file and I would say, Hey, like whatever we've spoken so far, all the steps that you've given me in terms of execution is in like random order in this, go read that through. And then like basically tell me if you're missing something or, you know, like if there's something that you would add to it. So, and honestly, that's been like the biggest unlock for me because it's always like come back to say, Oh, you know what, this is like perfect order. And then I say, there's this joy in like saying, let's go implement this because now I kind of know, Oh my God, like everything is just so on point. So I think it sounds so much like a compound engineering as much as, you know, something that I didn't give it a phrase for, but sweet. I would love for you to dig a little bit, yeah, thank you. I would love for you to dig a little bit deeper into the compound product concept because I think there was one thing I really wanted you to show these people as well in terms of the listeners. I would love for them to see how you're running this whole system of, you know, setting up an agent that goes and fetches a certain set of reports for you. And in the morning you're constantly like being updated. So do you mind walking us through how you've set up this whole thing? Yeah. So the first thing to do is not to make it more complex than it needs to be, right? So I think take a step back and forget about agents or AI and just think about, well, how would I improve this thing I'm working on, right? You know, whether it's a business or an app or whatever, you know, you think, well, what would I do every day if I could improve it? And most of the time, you know, the first thing you do in a business is you would have a VP of marketing or a VP of sales or a VP of engineering and you have a meeting with them in the morning and they would say, Hey Ryan, you know, I noticed that our free trials signups are down. You know, it looks like we're getting less traffic from this source. You know, I'm going to dig into it and I'll, I'll come back to you and, and you know, suggest something. And then they would do it and they would come back and you would see the results and you would just iterate. Right. And so identify what you would do to improve whatever it is you're working on. And then you think about actually, well, how could I have an agent be a part of that loop? And so what I do for Untangle, which is the startup that I'm building, is I have key, you know, statistics in the database, things like free trial signups, you know, landing page visits, et cetera, et cetera. And so I thought, Oh, what I'm going to do is I'm going to output a data to a markdown file every night. And then I'm going to have an agent look at that markdown and then make a call to Opus four or five and to decide what would be the one thing that I would do if I was a VP of marketing, you know, and I saw this data to improve signups. Right. And then it comes back with a suggestion. And then I would say, well, now I want the AI to implement that suggestion. Why wait until I wake up? Right. Like it might as well just try it. And then I wake up to a pull request instead of having to decide, you know, whether or not it should try it. Excuse me. So I think that's the loop. Right. And it sounds complex, like agents and cron jobs and loops and compounding. But the truth is, it's just take data, look at the data, decide what you could do to improve the data, do that, and then try it. Now, now, technically, what's happening is, is that we are executing AMP or cloud code on the command line in a bash script. Like, what does that mean? Again, you should just ask an agent because it will tell you but but the idea is, a bash script can execute commands on a command line, right? And so you can run an agent like cloud code or AMP by saying amp dash X, and then you put the prompt in quotes. And what that does is it fires up AMP. And it runs, you know, in the background. And then when it's done, it stops. So all you're really doing is, is creating a bash script that starts the agent with a prompt. And then, and then it does that every day at a certain time. So the question is, like, how do you get the right data into the prompt? Right? So that's where you pipe in a markdown file and say, you know, here's today's marketing stats, you know, and your job is to look at those, figure out what's right or wrong, make a decision. And then, and then it fires up another bash command to start AMP again, and say, okay, your job is to fix this problem, you know, go do it and then submit a PR at the end. So it's, it's just two calls to AMP. That's all that's really happening here. And it sounds complex and hard, but it's, it's not. That's all that's really happening. Yeah. And I think you also put it in the right way, like, when you have the right data, and when you know what to do with the data, then I guess your steps are also simplified. And to what it's, you know, as a fairness, I think most operators and founders today have this data. It's just the fact that we don't know how to operate with the data, or we actually don't know what to do with the data. I think that's the problem, because at the end of the day, everybody understands that, you know, go to any founder or go to anybody who's building an app and ask them, hey, like, how many signups do you have? And then there's always like a DB, there's always like 10 things that they can pull off and show you there's this clarity, there's mix funnel, there's so many, you know, products that they technically set up as soon as, you know, an app or a certain product is into picture. It's just that there's always this confusion about how do I make one talk to the other? And I think that's where the whole, yeah. Yeah, no, what do you do? What do you do with the data? What do you do with the data? Exactly. It turns out, like, probably the data is already just in your Postgres database, like, just start very simply, you know, instead of piping the data out, you know, to a mixed panel, just keep the data in your database, then pull it out, give it to the agent and have the agent decide what to do with it. We're in this kind of world now where it's actually very powerful just to have the data yourself and then, you know, output the data to a markdown file or you could even, you know, install a Postgres MCP and have the agent, you know, query the database. There's obviously some security risks with that, but we're in a whole new world where, like you said, that you don't need to have analysis paralysis anymore because you can have the agent take the first cut at, you know, wow, there's all this data, like, what are two things we should think about, right? Yep. Absolutely. And like you rightly pointed out as well, you ask the agent what to do. I mean, at the end of the day, if someone's just listening to this part of the episode or someone's just seeing your tweet, all it takes is just put that tweet to the agent and then ask it, hey, I want to do this with my data and I basically have this data in Postgres, or Superbase, or whatever it is. And tell me the exact steps that I need to do to set it up. I'm sure the agent is going to tell the exact steps in terms of step one, step two, step three. So it's almost like a... I think the whole... I really resonated with you when you said people get scared with AI. It's just that it's too huge, it's too big, it's just too hard to sort of like put into that agentic mindset. But I think it just comes down to being willing to ask agent and then like try to say, you know what, I really don't know how to do this. Can you just walk me through this? Yeah, or can you do it? I mean, so there's so many times where, you know, AMP will say something like, well, here's how you do it. And I'll go, you do it. It's like, okay, you know, I mean, and I think it's basically a learned behavior. What people need to realize is as soon as they say, I don't know, then ask the agent, right? So for instance, like say that they're listening to this and they think, oh, cool. So I need data to give to my agent every day. Well, I don't know what data to give. Well, great, ask the agent that, you know, say, hey, agent, you know, given what you know about our app, what are the key stats? Like what are the five stats that really, really matter? Okay, are those in our database? Are we collecting those? Okay, we're not. We'll go do it, right? Okay, you did it. Now, I want you to build a script that pulls out that data every day and puts it in a markdown file. So it's an amazing time to be alive because anytime you used to say, I don't know, now you just say, how do I, and then... It's like, how will you, to the agent? How do you? Do it, do it, okay. No, totally. As I always say, there's so much joy in just trying to say, go implement this or, you know, go do this for me, especially after the test. Oh, it's magic. It is magic. It's so good. Every waking day, I just derive so much joy in putting that line and sitting back and watching it execute stuff. I know, it used to be so much work. Oh my gosh. Like, I think that's why everybody I know that remembers the old way, you know, that build products by, you know, reading up. I have like an O'Reilly book behind me because like, it used to be hard. Like you used to have to get the book, open it up and look for how to do this. And then copy and paste the code and then try it. It didn't work. And then you go to Stack Overflow and kind of get an answer. Someone would make you feel bad for asking, you know, that's all gone. Like, and now there's this joy in building again because all the crap has been taken out and all the fun is back in, right? So, yeah. No, I am totally with you on this. And as someone who's experiencing it, I just wish so many people who are listening to this also get started so that they can partake in that joy and actually experience this for themselves. So, so cool. In fact, I have like a follow-up question to that because we're talking about data and we're talking about how these agents, technically just two types of agents, two types of calls working on the data and giving you all the output that you want. People also talk about MCPs, right? Because there's also like MCPs being the right way to connect data. And then there was all this fad around, oh, like you can connect different MCPs to your different agents and then watch it combine data. So, I just want to, I think, differentiate between, at least from your working experiences, where MCPs can actually be helpful, especially from the data context that we're talking about, as opposed to, let's say, these kinds of agents directly working on your markdown files or your data, wherever it is. Like, is there any sort of a difference in perspective or output that you see? Yes. So, I use mostly skills now and not MCPs. So, a good example, well, let me back up. So, in order to use agents well, they need to have a good feedback loop. And what that means is the agent needs to be able to understand if what it is doing is correct or not. So, you have to always think about, well, how can I empower the agent to do that instead of me copying and pasting? As soon as you're copying and pasting, you're doing it wrong, right? So, as soon as you do that, you think, okay, I need to find a skill or find an MCP that allows the agent to solve its own problem. So, an example is, I used to have the Playwright MCP installed. And Playwright, for anybody listening that doesn't know, is basically an amazing framework that allows, it used to write code to control the browser, but now the agent can write code to control the browser. But actually, I've taken a step back and there's a new install and skill called Agent Browser. So, it's Agent-Browser. It was created by the Vercel team. It's open source, it's free, it's great. So, if you're listening to this, just ask your agent to go find Agent-Browser by the Vercel team and install it and it will do it. And what it is, is basically, it teaches the agent how to fire up Chrome on a debug port and then use Playwright to control it, right? And it's amazing. Like, and so, what I don't use the Playwright MCP, I just use the Agent-Browser skill, which is using Playwright under the hood. So, now, if that doesn't make any sense, all I'm really saying is, skills are really where it's at right now. And there's a website called, I think it's skills, is it skills.h? Yes, skills.sh. Okay. It's this amazing sort of list of good skills and you'll probably see the Agent-Browser at the top of that. So, that's primarily what I'm doing. I have one MCP for Sentry, which I use occasionally. So, Sentry is a tool to monitor your logs and look for errors and it's pretty great. So, what you can say to your agent is, hey, do we have any Sentry errors? And it goes and use the MCP to look and then it can sort of get a bunch of info on that. Right. I don't use that as much. Right. Okay. But yeah, MCPs, I'm using them less. It's mostly about skills right now. So, the Compound Engineering skill, the Agent-Browser skill, the Ralph skill. These are a couple of things that I use a lot. Nice. So, yeah, I think it's also safe to say that as there's more advancement in the way these models perform, there's also going to be like constant upgradation in terms of like, which gets better to work with the model. So, I guess like four or five months ago, or at least safe to say like about six, seven months ago, I think MCPs were the format to get the data from different tools. But I so side with you around the skills because it feels like these agents are working better with skills. It's almost like the feeding information that it needs for getting to do certain actions. So, yeah. So, slowly, I think the transformation, as you rightly pointed out, is coming off from MCPs to the skills. And almost, let's say, there are still some cases where MCPs can barely get you the data maybe. Like, you don't have to do the old school integration or the old school like, oh, I've got to add this with this. And then, you know, like get into the loop of getting stuck. But I guess that's where MCPs can really be helpful to get you off through the stuff and then get you the data. And I guess skills can then work on top of that data. Right. And it's interesting, though, because so I host everything on Vercel. I love it. I'm not being paid to say anything. Yeah. So I'm not in anyone's payroll. So everything I'm saying is just because I like it. And I love Vercel, as you do. And so they have an amazing CLI. Like the CLI tool, the Vercel tool is extremely powerful. And they just shipped an update. And it knows about every single API on Vercel. And so if you add in a skill, which helps the agent understand how to use the CLI tool, you basically get all the benefits of MCP without all the extra context bloat. Interesting. And so because agents are really good at running bash commands, like they're really good at using CLI tools. And so I try to default to using a CLI tool with a skill instead of an MCP because it just seems faster. Yeah. And it's just amazing. Yeah. So I would say if you're on Vercel, update the latest Vercel CLI tool. Make sure you have a skill installed and you'll love it. It's great. Yeah. I guess the end of this podcast, everybody's going to go look up for the agent to say, hey, how do I create my skill? If they've not already, but then really wish for more folks to sort of do that. And I think I'm recently a convert or rather I would say I shift between the ID chart boxes and terminals. So it's almost like I'm in the middle ground right now, not completely a convert to CLI yet. But I do see a lot of benefits, especially working with instructional files, whether it's like MD files or whether it's some of these skills stored in the MD files. More to be easier through the terminal than through some of these agentic chatbot interfaces. So yeah, I do see the point there. Cool. So I just want to close the loop on compound product part of it. So technically compound product is what we spoke over the last couple of minutes in terms of understanding what's really something you want to go improve on in your product. And then trying to compound engineer through these agents and do one task well one day at a time when you sleep. While you sleep. That's right. Yeah, it's pretty amazing. And then in the end, you'll probably end up running three of these loops at night. I think that's where we're kind of going as people managing these fleets of agents and figuring out how to parallelize as much as possible. We'll kind of see how it goes. Right, absolutely. Can't wait for, I don't want to say the future because that's already happening. So I just can't wait for people to get started with that setup. So as we're talking about the AI coding agents, I have a question for you around evaluating the output of an AI agent, right? Because I've often heard from people and we ourselves spoke about the context and sometimes the agent giving you like a bad answer or not being able to get through the memory of the last few chats. What's, I don't know whether to call it as a framework or what's the certain rubric that you go by personally from your experiences to evaluate where coding agents are today? And where do you see or how do you see teams using it better? Like when do you judge that something is working well, something is not working well? And how do you get these agents apart from skills? Is there anything else that you can get these agents to give you like the kind of right output? Yeah. I mean, I would say, you know, the leading agents in my mind are AMP and Cloud Code. And I think a lot of people are using codecs. I don't use codecs. So I can't speak to it. Oh, okay. I'm a codecs fan, but yeah. Oh, cool. I keep wanting to try it because I hear GPT 5.2 is amazing. And you know, so I kind of. It's really great with your databases and something like anything that you have from a complex data perspective, it's really good. Cool. Yeah. Apart from that. Yeah. Yeah. I'm hearing that. So it's like, I'm tempted to try it. But I think most of these harnesses are pretty good, right? I think if you're going to choose Cloud Code or you're already using Cloud Code, great. If you're going to use AMP or you're using AMP, great. Like these are good tools. I don't think you need to be worried that, you know, either any of them are going to be bad. I think the most important part is pick one and get good with it, right? So if you can imagine, you know, this is like you are a runner, right? And you're picking your shoes, right? You know, pick your shoes and stick with your shoes and get to know your shoes and use them more, right? And then it's a very important tool. So I think the more time you spend with your agent and really get good at it, the better. So I use AMP. I love AMP. Like I'm very good at using AMP. It's got a lot of functionality that I use heavily. I only use it as a TUI or on the CLI because it's a power tool. And so I think folks, you know, just pick one. I think there's this like FOMO going on where it's like, oh, maybe I should be using Cloud Code or maybe I should be using Codex or maybe I should be using AMP. Just pick one and get good at it. You know, Anthropic's not going away. AMP's not going away. OpenAI is not going away. Like these things are going to be around. So just pick one and get good at it. And also, it's getting to the point where, you know, everybody that I know that's really smart doesn't write code by hand. And so I think if you're in there, you know, trying to do tab completion or editing the files, you're doing the wrong thing. Like you should be letting the agent write the code. And then if it's super important code, review it, understand it, know it. But you shouldn't be micromanaging the agent. I think if you're doing that, you're doing it wrong. Right. I think there's a subtle difference then between the whole commanding the agent to do it your way versus also micromanaging. I guess now the micromanaging is no longer about because there was a time when, you know, folks said asking an agent to do what you want was sort of the micromanaging. But I think we're far past that stage because what looks like micromanaging now is technically doing those tab complete or, you know, like sort of taking that engagement with AI agent in the wrong way versus if you really go very specific about the way that you want and give it all the context, I guess that's no longer micromanagement. That's more like the engineered way to get to the output that you want. Yes. And I think what we're finding is with good agent harnesses, you know, like a cloud code or like an AMP, you don't need to tag files really anymore. So you just have to do this, right? Like, okay, what's the right context to give the model? Most of the time you don't have to do that unless you want to, you know, shortcut maybe the first step where it's grepping, you know, a bunch. But most of the time you don't even need to do that. And then it does a very good job of finding the right file and reading the right lines. So I think that is something you don't have to micromanage anymore. That's very true. Yeah. I also have a feeling that some of these providers, like you mentioned, like whether it's Anthropic or whether it's OpenAI or AMP for that, I've not used AMP yet. It just makes me so curious to go get started. I'm giving you FOMO. It's bad. This is the problem. We're all going, maybe your tool, maybe your tool, maybe I should do it. And it's like, just use your tool. Yes, yes, yes, totally. I use a combination of Augment plus CodeX. So I'm on VS Code, just use both Augment and CodeX as the extensions. But, but you're right. Like, see, in some sense, I also feel like this is so much contexture at the same time, a contrary advice, because there's also like way too many things happening when you open X, when you open LinkedIn, there's just a ton of courses, a ton of new things that you get to know every day. So as someone who is in the forefront of AI and wants to learn everything, I think it's just not enough. It just feels like, you know, I have to know this. I have to know this. Like, what's CloudBot? Why did it change its name? And then there's way too many things. Why is it MoltBot now? What's happening? Exactly. What's happening? I know. Should I buy a Mac mini? Should I not? What's all this? I know, it's stressful. You just got to take a step back and, and take a deep breath and, and dig into what you're already doing and get better at it. Right. And yeah, there's a couple of things that I think people need to be aware of. Like, if you're not generally aware of this idea of compound engineering or compound, compounding, like, then you do need to get on that. Right. But your current tool can do it. Right. So don't switch tools, you know, or buy a Mac mini to do it. It's more like, what are you doing right now with your agent? And how can you compound that? Right. Like this morning, I was thinking, I just had the idea this morning for this idea that every day I should run a cron job to do compound engineering on the threads I ran that day. So I just ran down to AMP really quick and said, okay, I have this idea. Like, I'd like you to read all the threads in the last 24 hours. I'd like you to use compound engineering to look for what we learned. And I want to make sure that you updated all the right agents.md files, you know, and then ship that in probably like 30 minutes of back and forth with AMP. And now that cron job runs on my Mac. Right. Very nice. And I didn't know how to do that. No Mac mini. I don't. I was like this close to buying. I was like, oh, no. I saw it. I saw it. And I have my very own close friends have Mac mini. So I was like, literally pinging them to say, hey, is it any better? I've always been tempted, but now it's temptation to the core. Should I do it? Yeah. So yeah, it's crazy. I mean, my Mac just sits here on my desk at night. Like I can just use caffeinate to keep it on. Like it's not doing anything else. Like it's basically a Mac mini. I don't need to buy one. Yeah, no, this is super grounded piece of information, I would say, because there's always again, I'm going to I'm going to touch this topic back once again to ask you. But I think on the whole, it's really, really actionable to focus on what you're already working on, especially the tool of your choice. And I think it's just a matter of, hey, if you if you've not gotten started, then yes, of course, you have to decide where do you really want to go towards and then pick and learn about the tool. But I guess for people who are hearing this and who already have some sort of a process through what they're doing with AI, it's just doubling down on that and then finding more avenues, like you mentioned, to, you know, keep doing that in a better way. So, yeah, that makes a lot of sense. I mean, it's almost like there's still no shortcuts, like just because you have agents now doesn't mean there's magic, you know, shortcut to making millions of dollars. You still have to do the work, right? It's just that the work is now higher level. And what I mean by that is there's no secret trick that someone said on X that you're missing, which is keeping you from being a successful agentic engineer. It's just not real. Like what you need to do is get your hands on the keyboard, right? And and engage your brain and then, you know, and start talking to your agent. That's the most important thing, like hands on keyboards, you know, or if you're using WhisperFlow, you know, start talking to your agent, right? Yep. Yep. Oh, I'm a huge fan of Whisper as well. So, you know, and what you just said, you know, basically is going to be like a nervous system calming down for most of them listening because constantly like, you know, agitated with knowing stuff. But just as a sideline to this, let's say, you know, I am sticking to a tool. I am getting better with what I do, finding these avenues to grow. What would you still think as the most sanest way to keep in tabs with what's happening? Like, for example, like, how do I get to know something like skills.sh? How do I get to know that? Hey, like, Vercel has launched this and I would love to try it because I am already proficient with CLI. So is there like some sort of a sane approach to be on top and still like sort of choose to be away? Um, gosh, it's what I do is is mostly through X. I think if you follow a couple good accounts, you know, I think Kieran Clarkson, who is that the I guess they call him the project owner, the manager of Cora at every he's super switched on. I would follow him. Oh, yeah, I'd follow. I think it's Vercel. I think there's I can't remember if it's Vercel engineering or that. Just find a couple key X accounts, I think, and then pay attention to that. There's a couple of great podcasts. Clairvaux has a great show as well. You've got a great show. There's a couple of things like I think if you just pick a couple, you'll probably get the highlights. But just avoid the trap of thinking that, you know, you're you're you're just one, you know, trick away from unlocking everything. It's not really. Yeah, yeah. So basically, like filter out and, you know, learn where to tune in and learn where to step back and sort of work in your own. Yeah, absolutely. Makes sense. Yeah. And one question that, you know, I'd love to pick your brain on is where there's there's no dearth for existing founders, whether they're tech or non-tech. I think there's nothing called as a non-tech anymore. It's just that everybody is in tech space. So I shouldn't be calling the non-tech founders technically. But everybody wants to 10X their output. And that's what we're talking about. Is there like a. Is there like a one person or, um, you know, just one person being able to set up an entire company or entire apps of sorts and what's your, um, I don't know, not some sort of like, um, Conviction around it, I would say, because we're seeing like all of these agents being the replacement of teammates in every team. So is there like a certain future that you see, uh, which is one person behind all of the wheels and you are a good example yourself for doing that? Yeah. I think we're moving to, um, a reality where there's going to be a lot of solo founders that start companies and then scale them to, you know, five, 10, 15 people max, um, and, and, you know, so I've built three startups that, you know, the last startup I built had over a hundred people, um, you know, raised a bunch of money for it. Um, and I'm essentially building a similar, uh, app and complexity tackling a, you know, a similar difficult problem, but it's just me. Um, and you know, I used to have a CTO and a VP of engine, lots of engineering, lots of engineers and product managers and, you know, DevOps and, and marketers. And, you know, I, I'm kind of doing all of it now, not as well as, you know, as, you know, some of those experts that have been doing it for a decade more, but well enough, right. So I think that's where we're going. Um, you know, I'm currently raising, you know, a small seed round, but I almost don't even need to do that. Like, and there's a lot, I think a lot of founders are going to raise a little bit of money and never raise again. So it's a really exciting time to be a founder. Like it really is like your ability to multiply your output with very little capital has never been better. But it also means there's a lot of people building stuff, right. So, um, but Hey, that's just the world we live in. So, um, that's kind of where we're going. Yeah. And I think it's no longer about like one person or a solo person, because, uh, throughout this podcast, we were talking about how these multiplex of agents can actually be your teammates can actually be, uh, different avenues that you can get information from. So almost feels like as you spin up these agents and have them, you know, help you through time, it's just going to be like, yeah, I'm going to, you know, have people to work with me of sorts. Yeah, you have, I mean, managing fleets of agents now is, is your job. Um, and it's fun. It's like having employees really. I mean, you give them feedback and, uh, but they are always working and they're always happy to take the feedback. So, and like humans, I mean, you know, I love humans, but man, working with agents is pretty great. Yeah. Oh gosh. Yeah, totally. Totally. And, um, I, I kind of joke this, um, to my friends every day. I kind of say, you know, we're working with AI teammates is actually more fun than humans seems to me. So just doesn't, I know it could be true, but so I have this t-shirt that says humans are overrated. I prefer chat GPT and, um, uh, I was like, you know what? I want to like strike that and say a couple of things, not just chat GPT had like more, um, suggestions to that t-shirt. So have like multiple versions of the t-shirt be printed with different agents, I would, I would, I would buy that t-shirt. I'll, I'll, uh, link that to you. It's like so much fun wearing that up and the entire, so I work out of a small, um, you know, office space. And then everybody looks at me when I wear the t-shirt. Like, you know, is she crazy? Is she like really? Okay. Oh gosh. Um, so much, so much fun with AI. What a wild time we're all in so much to ship. Awesome. Um, right. Is there like, um, oh yeah, I really don't want to forget this part. Right. Uh, because I do know that we spoke about the whole system thinking, and I really don't want to leave, uh, you without, um, you having to walk through your systematic approach of working with these agents. I know you had like this three file system or like three step system, which you only love for you to like talk a little bit around that. I'd link that, um, you know, in the show notes, but yeah, I would love for people to have some sort of, um, understanding around it because we spoke a lot around how do you, you know, set up these agents and how do you start talking to them? So I think this is also very contextual to help them do that. You bet. So, um, I initially came up with this three part system where you, uh, plan, and then you take that plan, uh, which is called often a product requirement doc or a PRD, and then you create tasks to complete, um, uh, the implementation of that PRD, and then the third step is to manage the agent, you know, to complete each step. Um, but honestly it's outdated now. Um, and I'm basically, I mean, and that repo has, I think over 7,000 stars now. Um, it's a good place to start. Like, so I would encourage people to start there. Like you basically have a PRD, uh, plan, um, and the agent helps you write it. And then you, and then you have a task, uh, implementation and agent helps you create the tasks. And that's, that's fine. I think it's a good place to start, but, but I think now the best practice is using something like compounding engineering, which is very similar, right? So you have planning, um, you have executing, you have compounding, um, uh, you know, and that's kind of built in now. Um, so I think that's what I would recommend people do is, is just, yeah, I would just, you know, ask your agent to go find the compounding engineering blog post and then implement that system. Uh, and it's basically either a skill or a, uh, a plug in a fused cloud code. Um, yeah, that's, that's the main thing I'm doing. And then I'm adding on top of that, this idea of compounding product, this idea that it, you know, you should be creating a loop, um, where your agent is actually shipping stuff while you sleep so that you can wake up and, and, and move faster. Um, and I'm experimenting with that. I, you know, I think we, I think, uh, I've probably created with AMP, you know, three or four PRS using compounding product, it's a new idea and I'm testing it, um, found some things that work and don't work and, um, but that's my current thinking about the systems that I use, uh, to ship code. Right. No. Um, uh, the reason that I pointed out the three file system was also the fact that, um, a lot of that, um, kind of like gets a little bit tweaked or replaced with the, um, skills that we spoke about, because now it's almost like not just cloud code and not just Opus or Sonnet. I think, um, every other agent is also being able to recognize, um, a dot MB file and recognize what a skill training would really look like for it to get context. So the PRDs, I guess, at least would be largely replaced by, um, the skill and what it should do and what it should not. But I think the tasks part still remains to be, um, the whole, uh, planning out and then like, you know, execution of this agent, like go get this done, raise a PR. I think that part still remains to be the whole, um, you know, in, in, in, uh, uniformity with the tasks part from what you said, but crazy, um, just three months or four months, I guess. And then there's already like so much change. Yeah, it does change fast. And I'm glad you said while you sleep, because, um, it almost makes me feel happy that when I'm awake, I can still go command that agent. And it's not that, you know, it kind of does work when I'm awake. So that's almost like thinking what the hell would I do then? So it's almost like, you know, double the work. So, um, very, very cool. Uh, Ryan, thank you so much for joining me on the show. It was, um, fun for me as well as so much of personal learning. And, um, it's always exciting. You know, I said at the start of the show, it's always exciting to see people's workflows, what, um, really does magic for them. And when folks listen to it, there's almost like a feeling that they get in terms of, oh, like there's someone who's figured it out and I'm also going to like do the same thing. So I think your examples are like so spot on and very much actionable for folks to get started. So, um, really appreciate you for being here and sharing all of those, um, notes, links, apps, and I'm going to put all of that in the show notes, but, um, yeah, thank you so much. It's been a blast. Thanks for having me on and look forward to chatting to you and your guests in the future. Take care. Absolutely. Take care. See you.