How a “Non‑Technical” Founder Built World‑Class Agent Tools — with Matt Van Horn

[00:00:00] I get the pleasure to be in build sessions with MSPs just like you, building real things with AI. And the number one friction point, the number one blocker is I can't connect to all the systems. In fact, my vendor makes an MCP server or a connector or whatever they call it, yet it's read-only. Another one tried it and it limited them to 100 records. How are you going to search for all the tickets when you can only get to 100?

[00:00:25] If you're tired of vendor connectors and MCP servers holding your business back, check out the MSP Skills repo. The link is below. It's got over 50 different MCP server skills and connectors that you can use to connect your entire stack to the agentic AI of your choice. Now, why do you care? The reason is you can now go through and say, find all the unused licenses that I'm paying for that I shouldn't be.

[00:00:55] Would you like it to go through and prep across all your different tools for that QBR? That thing that took you hours to prepare for for one client is now one sentence. So if you want to make sure you're getting the value out of AI, you've got to connect it to your real world business systems. And there is no better place than the MSP Skills repo. Check it out on the link below. And if you have questions or if you just want to join, learn more, come to the one of the build sessions. You'll see us putting it into action.

[00:01:26] Last year's Thanksgiving was the moment when it stopped being a toy and actually became tool. It's also allowed someone like me who, again, I've lived in tech my entire career. I've been in software as a suit my entire career, but I have not shipped anything of value. And the fact that I can come up with a crazy idea, sometimes it's good, sometimes it's bad.

[00:01:56] I have many unlunched projects that are not shippable, but the constraints are no longer there. Like I can actually ship decent software. Hey guys, Damien Stevens, founder and CEO of Servosity, host of MSP Mindset. Today, I am blessed to interview Matt Van Horn. Now he created a company that eventually became Lyft.

[00:02:22] He created the self-driving oven, which was acquired by Weber Grills. But he had spent his entire career in the sales or business development side and is a self-admitted non-technical person. So how has he gone on to create such impact with agents? He created something called Last 30 Days, which is a skill that is the agentic search engine for agents.

[00:02:49] He created something called CLI printing press that allows your agents to connect anything in the world. So he has, as a non-technical person, had a bigger impact than probably anybody I'm aware of. The printing press has made a huge impact in my life. What that does is it allows me to connect my agent to anything I can imagine. It could be something for personal use. It could be my CRM.

[00:03:16] It could be my ticket system, finance system. If you need to connect to it, this helps you do it. And in a best-in-class or GOAT manner. So if you're ready to go from I wonder what I do to being inspired by a non-technical person having a world-class impact, you're not going to want to miss out on my conversation with Matt today.

[00:03:42] So Matt, I tried CLI printing press, I think, fairly early on. Fell in love with it because I appreciate a good CLI wrapped in a good skill, all the optimizations that you've done, why those are better, and all kinds of things. And then it was only after pushing the granola CLI to your library, your printing press library, that I thought it didn't work.

[00:04:11] And then it was only after that that I realized, well, you're the same guy that wrote slash 30 days, last 30 days that I use also. So the irony I thought was great today is I ran slash, you know, 30 days on Matt Van Horn this morning. I haven't even done that lately. It was kind of funny to see Claude get so excited. Wait, this is the guy that made the skill that I'm using right now. And I feel like that's unusual. I've never seen that.

[00:04:40] Yeah, that's unusual for 4.8, I think, to do. So it was kind of funny to see how excited, you know, it was. And it was how I felt when I was like, I love the CLI printing press. But wait, I love this, you know, slash 30 day, last 30 days. So I'm digging into a lot. And I want to cover also some newer things. But I just want to kind of start with like, what drove you to create these? What was the problem that you were like, I've got to solve this?

[00:05:10] I was definitely not thinking about problems. I was definitely not thinking about businesses. Was thinking about, I want to make a thing. I wonder if I can make a thing. Yeah. And, and last 30 days is a particularly fun one. I felt like I was always behind on the latest prompting techniques. Last 30 days is a lot more than prompting techniques. But that was, that was all I wanted to do was, I feel like my prompts are outdated. This is in January.

[00:05:39] So this is a very long time ago in AI world. Yeah. It's like 10, 10 years and dog years. But in, in January, I wanted to be able to search Reddit and Twitter on X for the best prompting techniques for whatever new thing just came out. And so I had this idea. I was like, wait, my agent can use multiple keys.

[00:06:05] And a lot of companies don't, don't talk to each other and are not friends with each other. Like Grok has no access to Reddit. Right. Gemini has access to YouTube, but not X. ChatGPT has a relationship with Reddit, but not X. And there's no place you can go in the world. Can't go to Google.com and get all these things. There's no place you can go. So, and so I was like, what if people just brought their own keys or what if I just brought them?

[00:06:34] It wasn't even about people. And so, and it was the circumstance of, I live in Washington and Seattle area. And our ski season was really bad. Our snow was not good, but my kids still had to ski because we paid for all the ski racing in advance. They couldn't even get gates up. So ski racing didn't even happen, but they still skied. And I was just at the lodge one day with my laptop, like, and I was like, I don't really want to ski today, which is sad because I love skiing. Right.

[00:07:04] I'm just going to go build that thing. I don't even know if it's going to work at all, but I literally banged out the first version of the last 30 days and a remotion launch video in the same day. I didn't really test it. I think I sent it to one person. I think they ran it once. They're like, cool. I was like, I don't, I don't know if anyone wants this or needs it, but I'm just YOLO.

[00:07:30] Just put it out there and it got a million views and started getting PRs. And then I remember I was like, I don't know what a PR is. I don't like people are submitting things. Like, do I just click accept agent? Should I accept this thing? Is it, you know, does it give North Korea access to all my search history? Nope. Okay, great. And then like, I accepted some PR and broke the whole thing like two days later. And then, so I was like, I'm not accepting any more PRs for a while.

[00:07:58] And then obviously I, I've learned what a PR is and an issue and have gone deep down this agentic path. But I, I just wanted something for myself and I put it out there and I accidentally built the world's first agentic search engine. And I accidentally built the most expensive search engine in the world and the slowest, but it is really, really good. Yeah. Yeah.

[00:08:24] And I think what's funny is I, I remember trying it going, I get it. But I remember kind of describing it to others, even though we're already using, you know, cloud code and other things. And they were like, just deep research, right? Like every model has some kind of deep research. And I'm like, no, like you said, they don't, they're not friends with all the communities. You don't have access to Reddit and X and, you know, all the other YouTube and whatnot. So I think that's, that's really cool.

[00:08:51] Um, I, uh, I, I didn't look, I'm sure cloud cancer this, but like the CLI printing press, I think does an unusually good job of finding prior repos or examples. Um, you know, I, uh, I'm, I'm hosting a weekly build session with, uh, with entrepreneurs building live and, um, normally they're building.

[00:09:18] But I actually two weeks ago said, look, I want to show off this printing press, uh, that I've been using. There's something to it. And so, you know, I was doing that really just answered some of the beginning questions because then it was ambitious. I was like, it's going to cook for like an hour now. So we'll, we'll go do other things. Uh, but, um, but yeah, that was, that was really cool. Um, and what was fun during that, you know, it was just a zoom meeting and there's, I don't know, 10 or 12 people on and they're giving input into the thing that we're building.

[00:09:48] Uh, cause I'm sharing my screen. And then I, we already just let it rip after answering just a few questions. And then they were like, Oh, there's this one guy that built this one thing that is, it's actually really pretty good for this use case. And of course it came back with that as one of the top ones that cited. Um, and so, uh, so anyway, somebody else described it as auto magical. Um, I think it's kind of, kind of has some whimsy to it. It does a good job. Is last 30 days part of that?

[00:10:17] I know I don't have to give it my scrape gators key. Um, but somehow you've kind of infused it with like, it finds the good stuff because if you go build something random and without learning from others, you're, you're just reinventing the wheel. So I'm curious how you, how you did that. Yeah. So, so printing press, again, it's one of those things. Like, obviously I've, I've again, 10 years and dog years since January.

[00:10:44] So printing presses, I don't think it's quite a month old, maybe almost a month old, but that was one of those things is like, I don't know. I, I kind of want this for myself. Actually, sorry, that that's not, that's not where it started. But soon I wanted it for myself. But where this one started was I was, obviously I'm, I'm very agent pilled and spent a lot of time thinking of, you know, I was a early, um, what was it called? Claude bot user. Yes. Or was open claw.

[00:11:12] Like I, the week that that started to go viral, Jason Calacanis had me on, uh, on his show as, as an expert because I had five of the most popular skills, um, for, for agents back in the, Claude bot days. And so I was thinking about agents a lot and it was Sunday morning, I think. And I just said, CE plan, compound engineering, CE plan.

[00:11:37] I want to invent a programming language from scratch designed for agents. Go, go. I love that. I love that. And, and then it's like, okay, let me dive in. There's something called Mog. It's, uh, it's gotten a little bit of traction, but agents don't know about preexisting languages. And I was like, okay, well, what about Peter Steinberger? Like what, what languages does he spend the most time in?

[00:12:05] And he's like, oh, well, he makes all these CLIs there and go. And I'm just like, Hmm. CLIs you say. And like, I, I'd previously posted like CLI versus MCP. Like I ran last 30 days on, on that and definitely was a big CLI nerd. And I was like, I wonder, again, this is a brainstorm with compound engineering and Claude about programming language turned into what became the printing press.

[00:12:35] And what if you could just make a CLI out of anything? And as I started to venture in and see what the magic was, it, it was actually a few different things. So one, again, shout out to Peter Steinberger, the, the, the claw father, the, the goat.

[00:12:54] And, and, and by looking whenever his agent struggled to do something well, he made his own CLI from scratch in go. And so I said, compound engineering, why are his CLIs so good? What's the TLDR? And the TLDR is that most companies, when they create a CLI or MCP server, they just kind of take their API and wrap it.

[00:13:24] But what Peter does is he says, I am a power user. I have an agent that has infinite compute, infinite tokens, and wants to use the crap out of this service and in the most aggressive, powerful way so that my agent is fast and efficient. And so he, so some of the ones he built, he built a disc crawl, which downloads everything

[00:13:54] that happens in discord. Like if you were just to like make an API call, that's like, okay, what was the last message that came in? And then the last message that came in, like, no. And so he puts an SQLite database where he sucks in everything and it's constantly refreshing. And so, okay, SQL database, that's smart. Thinking like a power user, that's smart. And so he built a few of these. He built the Google one, which is great. And I asked last 30 days, hey, Google came out with their official CLI. Like I'm trying to connect my Gmail right now to my agent. Should I use Peter's?

[00:14:24] Which I probably shouldn't because Google's official one came out three months later. That would be ridiculous if Google's was worse. And last 30 days was like, you should not use Google's under any circumstance. Peter's is much better. And I was like, what? What? That's crazy. Why? And so that just kind of got me in this headspace of what if everyone could be a little baby Peter Steinberger on their own and could print their own CLIs.

[00:14:50] And as I kept running into walls to figure out how this works, that's when a lot of the creative things and genius things that entered the printing press came to be. And I had worked on the printing press for a few days. And I had lunch with Trevin Chow, who was just leaving his job. We were both in the Seattle area. I met him. He sent me a DM because he was a last 30 days fan. I don't even know if he's a fan. I don't even know if he uses it, to be honest.

[00:15:20] But at one point he used it and he complained about a feature. And I kind of ignored his request. And then a few months later, Matt Mullenweg had the same feature request. And I built it for Matt. And then Trevin said, WTF? I literally wanted that a few months ago. And that's why I originally was like, I'm sorry I'm not Matt Mullenweg. I'm like, yeah, sorry. Anyway, so Trevin and I met on Twitter. And then we had a good lunch.

[00:15:48] And then I sent him and he was running compound engineering at the time, which I've already referenced. It's one of my favorite tools to build with. And I was like, what are you doing? He's like, I don't know. I need to figure out what I'm going to do next. I'm enjoying. He's like, I'm enjoying just building things right now. And I was like, I've got this crazy project. It's like Friday. I'm like, I'm trying to launch by Sunday. Do you want to like beat it up and give it a try? And he's like, sure.

[00:16:13] And then I feel like he did not sleep much for like the next two days and just started taking my, I'll call it mediocre printing press project. And he started to make it amazing. And I remember on like Sunday, he's like, can we at least launch like Tuesday or something? Like I have a family and it's close, but it's not quite there yet. I don't think we launched for another 30 days, by the way.

[00:16:40] Like we put in the time, put in the effort. Trevin made it amazing. But to your question, the four things that it does to make a CLI are really magical. So the one you referenced, which is it goes on GitHub and finds other open source projects that have done interesting things. And then it also captures the star count. Do people care? How old is this?

[00:17:08] Like, it's funny, Domino's Pizza, for example, there was like a Python script that someone had made years ago that still worked. That's funny. Where people could order pizza with this Python. It was like pre-agents. And so our program was like, all right, cool. Yeah, that's a smart method. And then what's funny is actually like the day before we launched, Domino's finally changed it. They weren't trying to get it. They just coincidentally changed it. So we had to reprint Domino's the day before launch. That's funny.

[00:17:36] Look at the community because the community gives you a signal for what people want to do. But it also can show some of the hacks that people have done so far. Like, does this Python script still work? Do these APIs, is this still the command to order Domino's Pizza? And we can test that during the dog fooding. So that's the first thing it does is look at the community. The second thing that it does is it looks for the official APIs, CLIs, MCPs that exist.

[00:18:03] And then the third thing that it does is it does a harsh sniff where it takes over your browser and starts sniffing for secret endpoints. And that's a very magical process. And what's cool is like you've seen it many times because you've printed kind of web pages and it's a fun experience. Two things that we're working on right now, which are kind of magical and kind of crazy.

[00:18:29] One, Trevin really wanted to connect his Bluetooth treadmill to a CLI. And my last company, we made a lot of Bluetooth accessories for grills. Our startup was acquired by Weber. And so we spent a lot of time with Bluetooth. And most stuff isn't very high stakes. And so you don't encrypt it.

[00:18:54] You just, the Bluetooth device just emits it because it's not like secret stuff like the treadmill Bluetooth is. And so anyway, you can harsh sniff Bluetooth signal, take it and control it. And so literally, Trevin built a Bluetooth sniffing for the printing press CLI and can now control his treadmill via CLI. And so I literally just saw PR come in for that 30 minutes ago, which is crazy.

[00:19:24] And the other thing that I was working on that I have not submitted a PR for is a mobile harsh sniff. So being able to install this weird VPN thing on your phone and then mirror it to your computer. But you can actually find the secret endpoints for different things. And so the Tesla CLI, which I just published, I used mobile harsh sniffing to help get some certain endpoints from that, which was really neat.

[00:19:51] And then the fourth thing that we do during the process is we put on our Peter Steinberger goat hat that says, what does a power user want to do? And if you actually, I bet 99% of people have not looked at this. You might be one of the few that has, maybe not. You can actually go in and look at some of the files that your printing press produced for your CLI. And there's a persona section in there. I forget what it's called exactly.

[00:20:18] And it literally makes like, you are a, like, let's go with the treadmill example. You are a power runner. You are training for a marathon and this CLI is designed for you and this treadmill. What features do you want? And it like role plays and then designs the features around that role play. I don't know if you've seen that. It's pretty wild. Well, the funny thing is in the build session was when I was showing others, I had built several, but it was the first time I actually noticed.

[00:20:46] And I wasn't looking at the files, but it was, I just was scrolling by when I was talking about other things. And, and so we looked at it and I was like, look at the, and you know, they were saying what they want to build. I hadn't, I was just showing the power of the, of the CLI printing press off. And I really didn't know, you know, I wasn't the power user. So it was really cool to see the, at least at that time, I think it was four personas that it pushed and they were like, no, that's good.

[00:21:16] Because even if you're the power user, you're going to miss the analytical user or the, uh, the backend user or the billing user or the, you know, whatever different persona, um, viewpoint. And so I noticed it, uh, you know, while I was kind of like, look at this cool thing, guys, let's build stuff with it. Um, and, uh, the other thing I also realized is when you say CLI to people that aren't agent pill, you know, it's really kind of like, why would I care?

[00:21:45] Um, right. Kind of, uh, uh, you get lost in the weeds. Um, so, uh, yeah, I, um, I think I've, the mobile one obviously doesn't exist. I haven't done that. I know I've done the wonderful web pages. Um, I, uh, I pushed it in like areas where it's like, it was not designed to do this. Are you sure you want to do this? Um, and cause you know, in Steinberger's way, you'd print a CLI for each, but, um, you know,

[00:22:15] I, uh, my daughter had some health issues. So I'm monitoring for like relative humidity and then radon and then particulate counts and, you know, all these things throughout the home and, you know, under the crawl space and, you know, all this stuff, but it's like across three different brands. So I was like, uh, it's all in the land, go figure it out, go find it. One might've had an API or something. The others didn't. It had to sniff them out, figure them out.

[00:22:43] Um, and it was, it also told me, it was like, we're not really designed. I'm like, no, I want one because you're going to pull it together because I can go into three different apps. Uh, you know, they're all work via, you know, iOS apps. This would have been, if I had the version you're talking about, this would have been easier, but it went, you know, across the land and is trying to figure out what exists and why. And it gave me what I wanted. Like it pulled the three together. Um, even the CLI. That makes it a goat. That makes it a goat. We invented, we invented the combo CLI. I'd never heard of that before.

[00:23:13] I said, we call them goats. Yes, that's right. The combo CLI really is. And it made that. And it's like three different vendors of IOT devices for monitoring, you know, home health type of things. And it's all put in there. Did you publish it? And I will. I haven't published that one yet. That sounds fun. You made a goat. Congratulations. Congratulations. It fought me on publishing just for feedback, because it was like, you got to submit the whole manuscript, but the whole manuscript included all the, uh, you know, um, every

[00:23:42] piece of the data and all the law, you know, some of the, so it should, it should be, it should be smart at stripping that. And just say like, you have to be like, I know you have to submit a full manuscript, so please do, but scrub everything first. So yeah. Yeah. Clyde at first was like, no manuscript. And then it was like, no submission for you. And I was like, okay, I'll come back and we'll, we'll, we'll do something in the middle. So it's cool to push it in different ways and see that, like you said, it is, it is the goat.

[00:24:11] Like it's pulled these things together and my agent uses it, but that was the first one I asked it to build a dashboard feature. And so you can run the fly with dashboard and it, and it shows like 12 rooms in the house, but it's pulling from different vendors and can kind of give you a picture of the trends across all the vendors, all the rooms. And that's just something you couldn't do. And even, even in the terminal, like that's more than I had before. And then of course the agent can do a lot more with it. So, uh, so yeah, it's, it's really cool.

[00:24:41] Um, and, uh, just impact, uh, kind of, I took your idea and I said, look, I shared it with the MSPs that are in this, in this build session. And some of them are kind of tinkering, but, um, I also realized a lot of the ones that I was going to print, they were kind of, uh, for this, for MSPs, you know, they may not submit them, but you, a lot of your audience may not really find any use. And so do it, please.

[00:25:08] And so, uh, the more needs, like we had a guy submit like a financial thing for Italy, like some sort of financial institution in Italy is like, great. Like, that's funny. Like, yes, submit, submit all the things like people will, will find value and especially your community will find value of being publicly available in the printing press. Yeah. Well, just for everybody listening, um, I think I've printed six or eight when I say manually

[00:25:35] using yours, but like where I was really doing that in interactively. And then last week, um, I, this community, they needed it, but they weren't building them fast enough. And I was impatient. And then that collided with, um, Anthropic just resetting the weekly limit. Great. So I had like 36 hours left to burn a week. And so I quickly said, all right, let's create a skill. I called it sea life factory. And I said, look, just like you, I'm using CMUX. You're going to orchestrate multiple pains.

[00:26:05] Here's how I want you to answer the questions when it comes back. And basically I said, I want you to keep five printing presses running concurrently, uh, 24 hours. What? And I managed to use 100% of my 20 X plan and printed overnight 20, um, CLIs. And so that was, I think you might have the record. I don't think Travis or I have done that. That's, that's, that's amazing.

[00:26:34] The point is not that I'm the greatest. It's just that this, this stuff really does work. And, you know, there's one thing to say I've used it once or twice and there's another to say, okay, now this whole community, these are the common tools they need. And now it's like, we can access those. And I would have done them once, you know, one a week or whenever fitting it in between time and meetings and, and cloud limits.

[00:26:57] Uh, by the way, there is, there is, I don't know if you also pay for, for codex, but there is a codex mode built into the printing press in cloud. I noticed that. So you could just say, cause I, I've only run out of my codex limits once, but I run out of my cloud limits every week. And so you can just say like slash printing press.

[00:27:22] So you want to print granola for the first time, uh, slash printing press granola space codex. And it will build the ghost CLI in codex and send it away to codex. Yeah. That, that is really cool way to spread your tokens out. Um, and, uh, I had, I think already, um, in that particular case, you know, use a lot of my, uh, codex ones, but my cloud ones were a little more available. Okay. There you go. Yeah. That's amazing.

[00:27:51] Because it's like, what do you do this productive, not just burn tokens, but do something productive with it. Yep. So I have an idea for a, a goat CLI. I don't know if it's going to work, but I, we, um, we're, we have to move offices, my company that doesn't exist that we haven't announced yet. We're moving offices soon. And, and, uh, we don't have an office manager type person to like move the office.

[00:28:19] There's not that many of us, but like, it's still, it's still a move. I have to like call a mover, get like internet set up, like these sorts of things. And I was just like, is there an API for a CLI for humans to do work? Right. Like, can I just have cloud code managed this whole move and I do nothing? Like what dots would have to be connected to that?

[00:28:47] And so I've, I've started seeing if it's possible. And so the first thing I'm going to print is a task grab at CLI. Nice. So being able to enable my agent soup to nuts to book a physical human with good reviews that can accomplish real world tasks. Right. So very easy to find a mover, for example, on, on task rabbit and some with good reviews and an agent can do that.

[00:29:17] And so I was like, okay, so that would be like the first step. And then I started looking at a few of these agent platforms. There's one called magic. There's another one with duck in the name and I'm forgetting exactly what it is, the duck one, but they are APIs to humans that can make phone calls. Nice. And so I tested both of them and had them just call my cell phone and ask if my refrigerator was running.

[00:29:43] And, uh, and, but they, the, the magic one was pretty good. It was a woman in the Philippines. She admitted she was in the Philippines, but she was very nice. She said, English was great. And so I'm like, okay, what I'm trying to think of what are the building blocks that I can print CLIs to make the, um, I don't know what to call it, but I think like the, the, the, the human goat API, the CLI, the human goat CLI.

[00:30:10] Um, and then there's a startup it's called like rent a human dot AI that has gotten a little bit of buzz. I think they're a YC company. I went to the webpage and I, I just typed in where I live, Seattle area. And there were a lot of people, but they didn't have bios or reviews. So there were like a lot of software engineers on there. I'm like, I don't need a software engineer. Right. I need a movie office space. Uh, like I need a mover. Um, so I, I don't think I'm going to include that in the V1.

[00:30:39] Cause I didn't think there was, didn't look like there was enough vetting. It looked like anyone could sign up to be a human for rent, but anyway, I'm going to try and see if this were, I think it, it, it depends if TaskRabbit will let me steal their secret APIs. And it seems like they will. And if they will, then I think this is actually going to work, which will be fun. I want to, I'd love to hear to, so obviously CLI printing press is amazing. I love, love the story of why you built it.

[00:31:06] Um, I use last 30 days just to tell me what's going on with you, which is really cool. Uh, just this morning. And, um, I love to talk about like, you're just taking off with some of these posts. Um, I actually not only had Claude or my Hermes agent, but I read also the, uh, the one about the agentic hacks in June. Um, and, uh, there's some really good takeaways in there.

[00:31:35] Um, and, uh, one of them I think is controversial is like, basically don't read the plan. Like, it's not exactly how you phrased it, but tell me about that. Yeah, no, that's exactly how I phrased it. I think I said plans, plans are for agents. You silly human don't read the plan. Yeah.

[00:31:55] I, um, so I'm, I'm a big, big believer that LLMs are very lazy in terms of getting you like trying to make you happy with the least amount of work and to get your agent to not be lazy.

[00:32:21] You need to give it very specific instructions to do something thoroughly one by one. And my best method for doing that is I use compound engineering and I, I pretty much don't let my agent do anything unless it's reading a plan.md file that it's made for itself.

[00:32:45] And I, when I, when I wrote my original agent to get my best cloud codes hacks article back in the day in March, so I guess that's what, six years ago in, in AI years. Yeah. I, I, I talked about how, um, I, I use Zed to open all the plans and read all the .md files. And I still love, love Zed, but I don't have a use case for it. It's a good product, but it's not in my flows anymore because I stopped reading the plans.

[00:33:11] And in, in general, if you give a good prompt, you give a good CE plan prompt. And I'm just often just using my voice and talking things. And, um, my plans that CE and Claude pull together are generally pretty good. Like, good enough.

[00:33:40] Like, and so like, okay, I put together the plan. And then if it's something that I think they might mess up, I'll be like, what's the TLDR? So I don't actually load the plan.md. Like, what's the TLDR? And I'm just like, all right, go for it. And then I immediately go to another tab and I kick off another CE plan on a different project. And then I'm like, okay, go. And then they come back and I'm like, okay, CE work. Like, I, I'm sure it's fine. And then like, sometimes it's building for 30 minutes.

[00:34:09] And I'm like, wait, what, what did you do? Like, that doesn't make any sense. And like, maybe I could have prevented that, but like, it's much better for me to move on to the next task in the next window. And the next thing that I'm doing then to dwell on nailing this perfect plan. Like it's generally good enough. And if one of the features got messed up, like can fix it in the next CE plan that I don't read and that's okay. Yeah. It, I have a software company.

[00:34:39] And so when I told my engineers, like, I'm not reading the code. In fact, you know, I was, I, uh, you know, I don't know if this is two months ago or something. I got a new Mac, got the new M5, got all this stuff, you know, cause the old wasn't keeping up, you know how that goes. And I was just like, I don't need an IDE. And they're like, how can you operate? I'm like, why would I need an IDE? Like, I'm not going to read the code.

[00:35:05] Um, and, uh, and so, but at least before I was kind of thinking through the plan a little bit more, trying to correct it. Um, so I think it's even more interesting to say, don't even read the plan. No, don't get, get rid of them. Get, get better prompts at the beginning. It means you didn't give a good a prompt if, uh, if your agent didn't understand you. Yeah. What's funny. Cause I see people and I'm doing things and other peers and I see you and others doing things. It's like, you know, you're just, you're just, you're creating things and you're learning

[00:35:35] from that and you can look at the output and see, did you get what you wanted? And then like, I've got folks on the other end of the spectrum and they're like, you know, it's hard to imagine not reading the code, let alone the plan. But the, the prototype or the CLI or the app or the, whatever it is, you know, can be reviewed, right. Right. Boring through a long markdown document may not be the best use of time. No, not at all.

[00:36:02] And it's funny cause you move the constraint, you know, code generation used to be hard and the thing you didn't trust. And then now it's like, that's easy. And then the constraint became reading plans all day long, which is not that fun. Um, so I love that. But I also think this is something I laughed out loud about because my team would, would again, think this is crazy, which I think is good is like, just change your shell. So the CLI doesn't open.

[00:36:32] It's like that, uh, cloud code does. Um, and, uh, and it's just like, yeah, because that's what you do with your, uh, your CLI. Right. Very, very randomly. Very rarely. Do you run some other command? Totally. It's funny. I, uh, I'm, I live in cloud code and every once in a while, Trevin will need me to like authenticate something on GitHub. That's like an advanced security thing to give him access to something on like the printing

[00:37:01] press. And he's like, okay, uh, CD into this folder and then run this command. And like, obviously I know how to, I can do LS and CD. That's like the extent of my terminology. But I, I literally desperately try to never leave cloud code. Yeah. And so I have cloud code run all the commands. And if I ever have to slash exit, I'm like mad that I have to do that. Like just, just my agent should be able to do it. Like why? Like, right. Just go. Yeah.

[00:37:31] Yeah. Yeah. Um, on that note, I want to talk about, because I use last 30 days and still use it. And then it came across printing press and I was like, this is, this is what I really needed. And then only then I was, did I connect it? Oh, you're, you're, you're that same person. Um, but then, uh, I think the other thing is like, you look at all the output and all the contributions you're making to all the other open source projects.

[00:38:00] Of course it'd be easy to say, oh, you've been a senior developer or, you know, something like that, but help us understand that's, it's not been your background, right? I'm not, I would love to actually go back to that, that example I gave you of the first week of last 30 days when people submitted PRs and I didn't know what to do. Like to where I am now is, is quite, is quite funny.

[00:38:30] So I, I think that there, there was, so I've been back when it was called vibe coding. Like I, I've just been in an AI early adopter. I, I, I, it was kind of like a, a funny joke in, in my company when someone would join the company, like it was a big ha ha joke that I would teach the vibe coding class because as the only non-engineer in the company, right.

[00:38:59] Non-engineer in the company that I, and I remember in the beginning I would, you know, try and just make like a hello world type thing in cursor and it would like kind of work. And then a few months later I would like try and make a, it's like years ago, it feels like an actual years, not, not months ago. And then I remember I was like, okay, I'm going to try and connect a weather API key to

[00:39:27] like display today's weather on a webpage. And, and then like, it wouldn't work half the time. Like that was like the extent of my agentic vibe coding thing. A year ago. And then in, on November 24th last year, right around Thanksgiving, that's when Opus 4.5 came

[00:39:53] out and a few weeks later, the next major codex release came out and that's, I call it like the BCAC works for Cloud and Codex. So yeah, she's one, but last year's Thanksgiving was the moment when it stopped being a toy and actually became a tool. And that's true for the best software engineers in the world. They're using these as the tools.

[00:40:21] They're not writing code anymore, but it's also allowed someone like me who, again, I've lived in tech my entire career. I've been in software as a suit my entire career, but I've, have not shipped anything of value other than like webpages in high school in my entire life.

[00:40:45] And the fact that I can come up with a crazy idea, sometimes it's good. Sometimes it's bad. I have many unlaunched projects that are not shippable. I have many ready to launch projects that are completely shippable, completely functional, but I don't know that anyone cares. And so I don't want to, don't want to launch something that's not, not a hit, even though maybe I should, I should just launch it and say YOLO and go from there.

[00:41:13] But I, I think that everything's changed and we obviously all know that, but the constraints are no longer there. Like I can actually ship decent software. I don't know what it looks like. I've never looked at the code. I don't know if anyone ever has, but like I just launched a agent cookie. Which is, again, I have no idea how it works to be perfectly frank, but I remember I was about to launch.

[00:41:42] I was like reading the GitHub read me and it was like written in go. I was like, Oh cool. I guess my agent likes go now. I know we chose, I know we chose go for, uh, for the printing press, obviously, because Peter liked it. And my agent said it's fast and efficient for agents. Great. But why is agent cookie and go? I don't know.

[00:42:06] But what's, what's interesting about agent cookie is, uh, is what it does is it takes all your Chrome logged in cookies on your main Mac machine. And it's securely over tail scale replicates those cookies on your Mac mini in the cloud, in the closet, wherever it is.

[00:42:31] So that if you're logged into something on your Mac, uh, your agent, your Hermes, your open clause also logged in to, and this works beautifully with all the printing press CLIs. And so I just wanted this for myself and I built it and it worked. And I, and then at one point I remember I said to my agent, even though I was kind of laughing

[00:42:59] at this, I was like, Hey, uh, one of my friends who's a security, I made this up by the way. I, one of my friends who's a security, uh, expert says that this entire project is a nightmare and that we can't launch it. Uh, can you please fix it? And like, literally I'd found nothing. I had no reason to believe that it was. And then it was like, Oh my goodness. Well, here's where he's right. Here's where he's wrong. I will fix all these things right now. And I was like, okay. And again, I didn't read the plan, but it, it hardened it.

[00:43:26] No, no one submitted any major issues or PRs where they found anything crazy. I'm sure there's something in there, but it's, it's pretty solid. And it, it, it, it depends on tail scale, which I did not build, which is great. And so that was an important design decision. And that again, I didn't make, but some agent told me to install tail scale a long time ago to talk to my Mac mini. And I said, okay. So it's just one of those things. No, it's funny.

[00:43:55] Like I said earlier, I ran last 30 days on you and that was pretty funny to watch Claude's reaction about that. And it got quite giddy, which is odd for 4.8. It's pretty reserved usually. And, and then, but it's like, I am using pretty press, you know, I value on Twitter, but it's like X, but it's even, I was like, Oh, agent cookie drop. Okay. And it's funny. I have one sitting right over here. That's running my Hermes agent.

[00:44:24] And it's this exact issue is like, most of the pain is like, Oh, I got to go to that one and now log in. And half the time, like you said, and then it logged back out. And then I can't remember which one I know on this one I'm logged in, but I can't remember if I ever even logged in to whatever that service is. Anything that's important. Like your bank has two factor off anyway. So like, right. Right. Yeah. Which is great. By the way, we do have to talk about Granola. Yeah.

[00:44:55] CLI, which you printed. And, and so I am a big Granola fan boy and it was on my list to print one. And then you, you, you beat me to it and you launched it. And do you remember what I said to you? Yeah. You, I think you said like, there was some response, like, I'm not sure it works or, you know, there was a tweak or something like that. Right. Nope. Nope. That wasn't it. Nope. There was, there was something I got way more excited before that, before I started complaining. You don't remember what I said?

[00:45:24] Oh, the, the X post. Well, I, what I remember is like, I did something. Here's the funny part to me. I had done a couple and then I was, it prompts you like, do you want to submit this to the library? And the others were like, let me run it. Let me make sure it works. And, but it worked so reliably that one. I didn't even look at it and I hit submit like maybe. And then to the library. And I just didn't think anything of it. And I was happy to contribute it. And then like, I just went, was with my family and did things like sleeping.

[00:45:53] And it was like the next day, you know, you, you were talking about how you're such a granola fan. And, and, and I forget exactly what you said, but then I remember like, you know, all kinds of people were picking up and I'm like, wow, this is awesome. So I remember exactly what I said, which you were not remembering, which is funny. And so I, I, I've, uh, over the years I've become a lot more reserved and less fun than I used to be in general. Like kind of second guessing myself too much.

[00:46:20] And I'm trying to be less like that and be more like the, the old me. And so I remember I was in the office and I was like, I was talking to someone on our team and I was like, I think I'm going to tell Damien that I love him. Right now in the comments for making this granola CLI. Is that weird? No, that's what I want to say. Like, how is this going to like negatively affect me? Like it's not, I need to be more, more out there, more fun. I'm going to tell Damien Stevens that I love him. And so it's hilarious.

[00:46:49] You don't remember this. And so it's, it's on GitHub. And I said, I just said, I love you. So excited to dig into the granola CLI. So it's so funny. I remember now that you say that. And then I was like, wow, he's really, he really likes his submissions. Like I didn't think as much of it, you know? Nope. No one else has gotten an I love you for a CLI submission. You can check your, send your agent to all my comments. I have not said that to anyone else. You, you are special Damien.

[00:47:17] And thank you for not canceling me for, for saying I love you. No, that, that was awesome. I remember like seeing, I remember just seeing like kind of waking up to like all kinds of people retweeting it because printing press was taking off and then the granola is popular. And then like the other thing that's powerful though, is I just answered a few questions. I had put a lot of effort because I was early, like you power user into granola. And I was tired of chasing every time they changed their database.

[00:47:44] And every time they did this, you had to, it broke the API or, you know, cause you're, I had some like, you know, they didn't coded up some Python script and I kept having to chase whatever the latest update changed. It broke. And so like this, you know, I put effort into it saying like, I've learned this and I've learned that. And, you know, think about this and test with this and test in all these real world scenarios that I've seen it break because I, you know, I am a, I am a power user.

[00:48:11] Like I think I average like at least 20 to 30 hours of meetings a week. And so a lot of, a lot of granola transcriptions. And, and so, so no, it's, it's, it's hilarious. I just remember reading through all of it and like, well, all I did was say sure to the library. And, you know, I've got you saying you love me. I got the printing press taking off. I was like, wow, this is kind of like, I just pressed okay. Instead of no thanks. I'm a little overwhelmed.

[00:48:42] So, so it was a little surreal. It was also kind of what inspired me to, to say, look, I'm gonna share this with others. Like we've got this group. I want to share this with them. Let's dig into it because I feel like most people, they struggle to connect this to the real world. Right. And that's how, that's how you do this. So, so really, really cool story. I want to make sure we're a good story to your time. What should we look forward to? What should we follow?

[00:49:13] What's the name of the secret company? What can you reveal to tell us what's coming? Yeah. Nothing to reveal about the secret company right now, but a lot of all these things will, will make sense. I think once, once we're out there and it's, yeah, it's just a really fun time to build.

[00:49:33] And you can just have a crazy idea in the morning and have a working prototype or convince yourself it is a bad idea to build a programming language for agents from scratch. What's funny is Chris Tate from Vercel, who's one of the, definitely a top 10 best agentic engineers in the world. He, he launched Vercel's agent programming language about two weeks ago, which was amazing.

[00:50:01] And I submitted a few PRs and it's pretty cool. I tried to build a really cool demo using the language and my agent pretty much said my ideas were not possible yet. So it needs a few more things, but it's, it's, again, they're doing it and it's amazing. And it's, it's really cool. And I'm, I'm a huge Vercel fan boy, but you can just do things right now.

[00:50:24] And I, um, like I will give an example of a project that I just had my agent build in a day and it was one of the most thorough, like I wasn't giving it any feedback. And it was like, I didn't say slash goal, but it was kind of operating in that, that sense. And I said, um, SQLite was made 21 years ago. Can you please read, redesign it from scratch designed for agents? Wow.

[00:50:54] I don't know if this is a good idea or not, but, but it spent about eight hours of tokens and it built a whole thing. And then I was like, okay, well, how, how are your benchmarks? Like, uh, and then here's what I said. I said, um, I didn't even mention it to Trevin at the time. I was like, okay, um, I'm going to go replace SQLite on everything in the printing press. What's Trevin going to say?

[00:51:23] And it was like, our thing sucks and is slow. You should not do that under any circumstance. So the project died right there. Even if it did come up with some novel new database for agent features that maybe it did. I have no idea. I never read the plan, um, but it was, uh, yeah, that was a one day dead, dead project as, as an example of just how, how my brain works. That's awesome.

[00:51:50] I love the experimentation and I do a lot of experimenting, but, uh, it's never been easier to share it now. And I mean that in, I've published a few things based on printing press to your library and publish a lot more. And so there's, whether it's your tool last 30 days research, other tools, it's never been easier. So that's, what's, I don't know, that's, what's exciting to me. It's like, you can try things and you can share them. And that, that is signal.

[00:52:17] Yeah, you should, you should post that, that story. And obviously you'll have this, this content, but you should post a tweet about that, that story of you having extra cloud credits and printing a ton of CLL CLIs while you slept setting the world record for most CLIs printed in agent history. I'll, I'll tweet that.

[00:52:40] I know I did actually explicitly say like, uh, don't hit submit on the end, at the end, because if I submit 20 and in a few hours, they might, uh, you know, get, um, might kick in and just block us or something. I don't, I don't know, but you know, we'll accept them. There's, there's a lot of Trevin set up a lot of automation where it'll yell at you. If your CLI is not very good and to fix it. It's off. Well, what's the best way for folks that have gotten value out of this to follow you? Connect with you.

[00:53:10] I'm just, I'm Van Horn on, on X. That's the best place. Awesome. Well, I appreciate, um, the time to do this. I'd love to do more together, Matt. Yeah, this was fun. Yeah. Thanks for being on. Thank you for reaching out. Yeah. Appreciate you. Of course. This was great.