Loading summary
Transcript125 lines
- [00:00]
Matt Van Horn
I certainly don't read any code. I don't even know how to read any code.
- [00:04]
Interviewer (possibly Peter Yang)
I thought you were actually an engineer.
- [00:05]
Matt Van Horn
I am not an engineer. I don't know what you call me biggest hack. Like if you do one thing is my favorite tool for building anything is compound engineering. Somehow I'm able to ship things of value, which is crazy and weird and still blowing my mind. What if anyone could print their own CLI about anything? It found all the secret APIs by sniffing and it worked.
- [00:31]
Interviewer (possibly Peter Yang)
Hey everyone. My guest today is Matt, who is arguably one of the best and most prolific AI builders that I know. And today he's going to walk us through his complete agentic AI. Well, I guess agentic engineering system that went viral on X. And I think, you know, from what I've seen from Matt, I think this interview is going to blow your mind. So welcome Matt.
- [00:48]
Matt Van Horn
Thank you. Excited to be here.
- [00:49]
Interviewer (possibly Peter Yang)
Let's talk about your. Like you posted this article, right, that had like I think 20 tips or something about building with AI agents. Do you want to share an overview of it?
- [00:57]
Matt Van Horn
Sure, yeah, I'll share the screen you made.
- [00:59]
Interviewer (possibly Peter Yang)
All right.
- [01:00]
Matt Van Horn
I didn't make the slide, you did. This is Matt's engineering system. I think you should present this. Peter, you made this?
- [01:05]
Interviewer (possibly Peter Yang)
No, no, I made this based on your article. So I mean there's a lot of tips on here, a lot of tips on here. But like what are some high level themes? You know, like maybe one high level theme, like you told me, is like you don't actually read. You don't actually read the plans. What are some high level themes?
- [01:20]
Matt Van Horn
Sure. So I think the biggest hack, like if you do one, one thing is the, my, my favorite tool for building anything is, is compound engineering. And I'm. It does. Compound engineering does a lot of things. Every has done an incredible job. Trevin's done an incredible job. Kieran and but the, the killer feature is CE plan and CE work. And what that does is if I ever have an idea, like I can give an example of an idea I had this morning. But if I ever have an idea, I always just start with ceplan. Here's my idea. I'm usually using voice. And what it's going to do is it's going to write a plan MD file and that's coming up with the execution plans for the agent. And the reason this matters versus just talking straight to Claude or to Codex is agents have a propensity to be lazy. Like they want to make you as happy as possible as fast as possible with as few tokens spent as possible. And by writing down a step by step systematic plan, it sets itself up so it's not lazy. And so for anything, I always say, CE plan, CE work, CE plan, CE work. And back in the day, and by back in the day, I mean like two months ago, I used to always read the plans. My last. I've only written two of these agent engineering articles. And the last one said, oh, we'll open the plan in your editor. Don't look at the code, obviously, but you should still look at the plan. And in the last two months, I found there's no reason to look at the plan and that I can just ask questions to the agent. So I'll give an example of something I'm working on this morning. So I launched a product, I think this week. Yeah, I guess this week on Monday, called Agent Cookie, which allows your. So if you have a main Mac laptop that you're using and a Mac Mini that you use for either Hermes or openclaw, the cookies are out of sync. And so you're logged out all the time. And so I build something over tailscale that makes your cookies perfectly in sync. And so this morning I was using another product of mine that I built with Trevin called the Printing Press. And often it's creating clis out of anything. And so I'm trying to create a TaskRabbit CLI right now. I can get into that in a second. Why, that's interesting. It's a human goat cli. But that's a side thing. Often on my own machine, my agent is logged out. And so I said this morning, CE Plan. Take a look at the Printing Press and take a look at Agent Cookie, two unrelated projects of mine. And figure out what Printing Press uses for har sniffing when creating a new CLI or agent skill and how Agent Cookie currently works and figure out how all my agents locally. So that's Vercel's agent browser. That's browser use. That's, I think playwright how they are always logged in to everything. And then it's like, okay, I made a plan and then I didn't load the plan because you should never look at the plan because you got to move fast. You got to move fast. But I had one question, I had one question for my agent this morning. I said, you're just modifying Agent Cookie, right? You don't have to like make a change to Vercel's Agent browser or to Playwright or to browser use. And it's like, nope, Just Agent cookie. This will magically work just with agent Cookie. Like, great. All right. CE work. Let's go build. And I. I think it built it, by the way. I think it finished a few minutes ago. I haven't looked at it yet.
- [05:10]
Interviewer (possibly Peter Yang)
Can you actually, like, go to your terminal and, like, show me what the plan looks like?
- [05:14]
Matt Van Horn
I said, you're not supposed to look at the plan, but sure, let's. Let's see it.
- [05:17]
Interviewer (possibly Peter Yang)
Yeah, I'm trying to understand the difference between C plan and just, like, regular plan mode. Like, what's so good about CPlan?
- [05:24]
Matt Van Horn
Let's see what happened here. So I said, CPlan, can you dig into the printing press and how the agents use agent browser. Browser for horror sniffing. Then dig into the agent cookie works and how the sync logged in. Are you sure you're logged in? Office says, I'm not. How can we fix that with Agent cookie? And then I said, no, sorry, I didn't mean that. Is it only agent cookie? And it's like, yes, I'm only modifying agent cookie. I was like, all right, great, go. So that was like my one question without ever looking at this plan. Obviously, I could have read this plan and maybe that would have created some value, but not really, because I just had one question. If I was submitting a bunch of PRs to agent browser and playwright and modifying their code bases versus just making a PR to agent cookie.
- [06:06]
Interviewer (possibly Peter Yang)
Yeah. So I guess to summarize the difference between. And I interviewed Kieran, too. The difference between CE plan, just like asking the agent to plan is it does a ton of research, your code base and online, and also, like, writes a pretty comprehensive plan. Right. Like, it goes into a lot more
- [06:20]
Matt Van Horn
detail, you know, So I don't even know because I don't read them is the honest answer. But. But the results are good. And. And my products often work and they don't break. Yeah. Right. So, like, for me, it's just about the iteration cycle. It's like, okay, I asked you to make a thing, and you made a thing. And then it's like, okay, does the thing work? Is it any good? And if it does, great. Thank you. Thank you, Kieran. Thank you, Combat engineering. But I think the big thing is I don't have to worry about it because I've used it so much that I'm able to just trust it. And I know what types of questions lead to success and failure.
- [06:56]
Interviewer (possibly Peter Yang)
Got it. So you do this for pretty much every new feature you want to build, and then you see build, and then do you read the code, like the PRs and stuff, or do you kind of. Is there like a command for that too? Like a skill?
- [07:07]
Matt Van Horn
So I certainly don't read any code. I don't even know how to read any code. I haven't written any type of code since high school. And that was probably HTML. I don't even think CSS was that big when I was in high school. And maybe some applescript language type stuff, but that's my engineering skill set, so certainly don't do that. And then pr. If I'm submitting an external pr, then I like to glance at it, obviously, and hit submit and make sure I'm saying things that are normal.
- [07:41]
Interviewer (possibly Peter Yang)
That's actually okay. Maybe I should have done more research. I thought you were actually an engineer, but it sounds like you actually. You're a gentech engineer, not.
- [07:50]
Matt Van Horn
I am not an engineer. I don't know what you call me. I've called myself a nerd my whole life. I remember when I got my first job at Digg in 2007, and I been a nerd my whole life. Like, everyone would call me to, like, fix their wi fi. Like, I used to camp out for Macworld to see Steve Jobs speak when I was a kid. And then I arrived in San Francisco in 2007, and I was like, I'm one of you. You're my people. Yeah, I'm so excited. And then they're like, do you have a CS degree? Are you an engineer? I'm like, no, I'm just a nerd like you. And they're like, not so much. And so I've always felt this. I don't know what the word is. There was always a negative signal. It's like, oh, Matt doesn't know. He doesn't understand. And again, I still don't try and pretend like I know. I literally say, I don't read the plan. I don't know what's happening. And so I obviously respect engineering and real engineers that I am not under any circumstance, that. But somehow I'm able to ship things of value, which is crazy and weird and still blowing my mind.
- [08:59]
Interviewer (possibly Peter Yang)
This episode is brought to you by Riverside. I've used Riverside for years to record my podcast because it records in 4K resolution each person locally, so the audio and video still comes through clean, even if a guest WI fi gets shaky. But the reason I love it now more than ever is what happens after we stop recording. If I go in here that I can use these AI tools to remove pauses, remove filler words, and. And just Clean up the recording and I can also edit the transcript directly and it will automatically generate clips with captions ready to publish to YouTube, Spotify and all types of social media platforms all from one place. As a one person creator business, that matters a lot. Riverside is the upgrade your content workflow needs. Try it at creators.riverside.com Peter Yang and use code Peter Yan at checkout to get one month completely free. That's creators.riverside.com Peter Yang now back to our episode. And it pretty much started like these tools started becoming good enough, which say last year, like late last year or this is like a pretty recent phenomenon, right, that you can just like ship all this stuff.
- [10:08]
Matt Van Horn
November 24th, Opus 4.5 shipped and a few weeks later the latest version of Codec shipped. So I call it BC AC before Claude and Codex after. But it was that Thanksgiving last year where it stopped being a toy. Like I'd been playing and building and was an early adopter of Cursor when it first came out. But I couldn't build anything of value. It was like, oh look, I made a weather app that kind of works. And those sorts of things, they're like, look, I was able to build a small app where I can ask a question to ChatGPT or to Claude. Like that was like my. Yeah, like. And then I would spend like three days trying to get Google oauth to work and it didn't. And but there was that shift, that BCAC Thanksgiving last year where it stopped being a toy and started being real.
- [11:02]
Interviewer (possibly Peter Yang)
That's amazing, dude. That's amazing. And I think one of the most exciting things that you shipped is the printing press. So you want to tell us what it's all about? Sure. And maybe you can talk about what's so special about CLI vs. MCP or some of these other integrations.
- [11:18]
Matt Van Horn
So this is printing press.dev. how this started was, it was Sunday morning and I said, CE plan. I want to start a programming language from scratch for agents. And my agent was like, sure, let's get right on that. That's such a good idea. I'm going to make an intensive plan for how we're going to build the best programming language. Agents are always over eager to please. And after kind of did its research, it came back and told me about the other agent programming languages. I think there's one called mog. It was pretty much like tldr. This is a terrible idea. You should not do this under any circumstance. There is no value to this. And I was like, dang, it. But agents and CLIs and agent efficiency, there has to be something here. What Peter Steinberger is doing is incredible with openclaw and all the clis he's making what could be something interesting that we could make today on Sunday and ship in a day or so. And somehow that jamming back and forth CE plan, CE work ended up inventing this crazy. I wish I need to find the original plans because I'm curious, like what exactly was said? This is all a memory, but it's all written down somewhere. There's a plan I haven't read somewhere. That was the printing press, original plan and my original prompt. But the basic gist of where we ended up was I started studying Peter Steinberger is whenever his agent. So his openclaw would have a problem using a service. So like Google apps, like Google Sheets, Google Docs, Gmail, he would create a CLI for himself. And so gog CLI is made by Peter. And then he would find some frustrating something with something else. Like on Discord, there was no good cli. There was an API, but there wasn't a good cli. And so he's like, I'm going to make disc crawl. And he would do this, and he would do this for himself selfishly. But because Peter is incredible and lives in this open source for the greater good of people. So he built something for himself and then he ships it open source to the world. And so I had my agent start studying his patterns and there were two very, very interesting. So first, sorry, another thing is I ran last 30 days on Google's official CLI versus Peter's CLI and I said, okay, which one should I use? I should use Google's. It came out like three months after they got to learn from Peter's. And last 30 days says you should use Peter's under all circumstances. It is much better than Google's. And so what? Why? How is this possible? Google got to look at Peters and so I dug in and there's a lot of things he's done, but there were kind of two that really stuck out for me. So one, everything that Peter does, he thinks of the power user. So he's mindful of what is someone that is a power user going to do with this. Like imagine the agent is a power user train beating up something all the time. What are they going to do? And the second thing he does is about speed and token efficiency. And so one of the things that pretty much all of Peter's CLIs do is it downloads everything and creates an SQLite database. So for example, let's go with Discord. So OpenClaw has a Discord. There's a lot of activity there. And so Peter would be like, hey, what's the top feature request? If he just like asked the Discord API, hey, what's the top feature request that people posted in my Discord lately? It'd be like, okay, well, let me see what's been written down. Let me download all the messages from today. Download message 892, 893, 894. Okay, I've looked at these and I've analyzed them. But if he's already downloaded and has an SQLite database of everything that's ever posted in his Discord, because it's just built into his cli, that's brilliant. Like, oh my goodness, literally his open claw, however he hosts it, has everything ever written Discord. And that's fascinating. And so for the printing press, the idea was what if anyone could print their own CLI about anything? And it kind of does four things at its core. So one, it does what Peter does. So it creates an SQLite database and it says, who is the power user of this product? And so we actually have it. It creates Personas. Trevin did this, where it actually says, so if you were to print the Discord CLI using the printing press, it would say, let's create a few different Personas of who would be a power user of this. And it's like power user 1. Meet Jim. Jim owns two discords and runs a big open source project and he is overwhelmed. What features would Jim want? And then it literally defines that and it uses that as a data point. And so that's one of the things it does. The second thing it does is it looks for official CLIs or APIs that exist already and ingest those. The third thing that it does is it goes and it does a har sniff to look for secret APIs. So I'll give an example. So the Google Flights and Kayak Direct, they do not have an official API, but if you do a harsniff and click around, you can find all the secret APIs that exist. And so that's how the flight goat was invented. So the FlightGoat, you can just say PP FlightGoat. So it's a skill. So it's a skill for Hermes or openclaw or for Claude Code. I want to see Seattle long haul nonstop flights December 24th through January 1st for passengers cheapest first. And what it does first is it goes kayak direct secret APIs and looks for all the nonstop flights over 8 hours from Seattle. Then it takes that and goes into the secret Google Flights API and then runs it. And then in one command, you're able to get a result. Like, this is a real result. So London is the cheapest for Christmas to New Year's at 1200 per passenger, then Amsterdam, then Tokyo, then Paris. And it did this within one command, which is bananas.
- [17:58]
Interviewer (possibly Peter Yang)
A secret API is basically like, you know, if I go to kayak.com and start browsing around, there's like a bunch of APIs that it calls through the browser. You just found out what that is?
- [18:07]
Matt Van Horn
Exactly. Yep. That's the feature I was working on earlier today, was to make sure that the cookies are synced so you're always logged in when doing a hard sniff when printing something with the printing press. That's the third thing we do. The fourth thing we do is we go on GitHub and we look for community projects, people that have already been hacking and finding these secret APIs and building off of them. So we found this really cool project. It was a Python library for Domino's Pizza ordering. And apparently Domino's.com hadn't changed any of their order pizza endpoints in like 10 years. And this someone built a Python, a pre LLM Python library to order by a Python script, Domino's Pizza. And it had, I don't know. I don't know the numbers, but a couple hundred stars. So it's not like a major project, but. But it worked and it was really cool. And so we're able to learn from all the community projects, which gives signal of what people want to do, but it also gives signal of where the APIs exist. And what's funny is, literally the night before we launched the printing press, Domino's changed their APIs. Literally the night before Trevan tried to order a pizza. He's like, it's not working. And so he reprinted it and we shipped it and they changed it for the first time in ten years the night before our launch. But now you can order pizza within the printing press.
- [19:26]
Interviewer (possibly Peter Yang)
He reprinted it, as in, like, he found the secret APIs that Domino uses.
- [19:29]
Matt Van Horn
He just. You literally go into Claude code and you type printing press here. I can.
- [19:33]
Interviewer (possibly Peter Yang)
Yeah, maybe you can demo, like, both, like creating CLI and also like some of your favorite clis on printing press.
- [19:40]
Matt Van Horn
Yeah, so literally I just typed in Flash printing press, and I might be a few days behind because it's shipping a lot. I actually don't know what to print. Do you have anything you want to print? Any web pages that you use? Let me show the library for a second just so you can see examples of stuff that's in here. So in AI, there's 11 Labs Midjourney. Midjourney doesn't have an API, so that's pretty cool. Someone printed a Mid Journey CLI for commerce, Amazon ads, Amazon seller for developer tools, agent capture, airframe, all recipes.
- [20:19]
Interviewer (possibly Peter Yang)
I think a substack CLI will be useful personally.
- [20:22]
Matt Van Horn
Yeah.
- [20:23]
Interviewer (possibly Peter Yang)
To see what kind of posts people are writing.
- [20:25]
Matt Van Horn
There is one already. Oh, there's one printed by Cherentan. So there already is one. And what's crazy is I think we launched with maybe like 40 or so clis. Now there's 200 plus. Like people are just printing these and it blows my mind how cool it is.
- [20:46]
Interviewer (possibly Peter Yang)
And to print, like, what's the process to print one? Just be like, hey, can you, can you just print one for me for this product? Isn't that what it is?
- [20:55]
Matt Van Horn
Yeah. So literally you just type in. I think it's just checking to make sure it's updated.
- [20:59]
Interviewer (possibly Peter Yang)
Why don't we try Papa John's Pizza? Papa John's Pizza is there.
- [21:03]
Matt Van Horn
Sure. Yeah. So Papa John's Pizza. So I'm not logged into Papa John's, but let's see what it does. It'll at least start the process. We definitely do not have a Papa John's. Omar from Microsoft who's running open cloth strategy at Microsoft, he printed the Jimmy John's sandwich CLI using the printing press.
- [21:28]
Interviewer (possibly Peter Yang)
Nice.
- [21:28]
Matt Van Horn
All right. And then it says anything you want to share before we begin a vision for what the CLI should do. Specific features you care about. And then so the options are, let's go start research. I'll ask for API keys, browser author, other context when I need them. Do you have vision? Specific features. Let's just go see what happens.
- [21:46]
Interviewer (possibly Peter Yang)
Yeah. Okay. So I guess like a lot of the value is actually.
- [21:49]
Matt Van Horn
Yeah.
- [21:50]
Interviewer (possibly Peter Yang)
Finding the relevant GitHub projects and the secret keys. Yeah, Like a lot of stuff that I use are like, just really annoying. Like I have a scale that I use to get my weight and stuff.
- [21:59]
Matt Van Horn
Yep, Yep. Yeah. So Trevin, actually he had a wacky idea, let's say yesterday, to build a Bluetooth har sniffer. Because he had a scale. Sorry, not a scale. A treadmill that he used. And he wanted to be able to write a CLI to control his treadmill so his agents could control his treadmill. And he built it and it works. And he Shipped it yesterday, which is crazy.
- [22:35]
Interviewer (possibly Peter Yang)
Wow. And while we're waiting, maybe you can just give a quick primer of CLI versus MCP or what's so special about cli.
- [22:44]
Matt Van Horn
Yeah. So I'd say for me, I often find for companies where I'm using their tool where they only have an mcp, I feel like I'm always authenticated and logged out. Like, it just seems to always happen. I don't know why I probably need to do a CE plan. Why am I always logged out? Is it a me problem? Is it someone else? But I know a lot of people around me feel like they're always logged out, and it's like, oh, you got to refresh your MCP and renew your auth token. Right. And it's quite annoying. And what I found is CLIs are very clever and thoughtful about how they re off. So a lot of the. The clis, within the printing press, there's a mode that the agent knows about. It's like, yeah, okay, I just need to go load the browser and refresh the webpage, and then all the CLIs will work again for this specific use case. And I know how to refresh my token. Like, it's built in. How do I refresh my token automagically?
- [23:46]
Interviewer (possibly Peter Yang)
Okay.
- [23:46]
Matt Van Horn
And what's interesting is what we actually print with the printing press is we don't just print a cli, we print a clique and an agent skill and an MCP for everything we print. So, for example, so I used the flight goat before, right? So if you install the flight goat, you get a cli and you get a cloud code skill. Yeah, the codec skill, you get an Hermes skill, you get an open Claw skill, and you also get a full mcp. So you can use it actually in Claude desktop, which is crazy. So you can actually download that and install it into cloud desktop, which is wild. It blew my mind when Trevin showed me that one that he built. And so the answer, the magic for me is if I just have a cli, I feel like I don't know how to call upon it. Like, it's up to the luck of my agent to maybe find it. But if there's a skill plus a cli, that's the real magic. And so if I'm in openclaw, if I'm in Hermes, if I'm in Claude code, then it can just. It knows how to go.
- [24:50]
Interviewer (possibly Peter Yang)
It knows how to trigger. Or, like, you can trigger it manually.
- [24:52]
Matt Van Horn
Yeah, Yep.
- [24:53]
Interviewer (possibly Peter Yang)
Makes sense.
- [24:54]
Matt Van Horn
Exactly.
- [24:55]
Interviewer (possibly Peter Yang)
Well, did Papa Jonathan have anything, or is still working.
- [24:58]
Matt Van Horn
Oh, it takes about 45 minutes to an hour. By the way, we are not going to finish the cli. And I also need to register for Papa John's. But what it's saying is it says there's no CLI right now, but two pizza ordering peers. Do. Do you want to continue building Papa John? Of course we want to keep building Papa John's. So. Well, so here's what it's going to do. It's going to resolve the spec, library lock, registry checks, API key check gate, research the brief for Papa John's Auth intelligence browser, sniff gate, crowd sniff gate, absorb gate reachability, generate cli, build absorb transcendent features.
- [25:33]
Interviewer (possibly Peter Yang)
Got it.
- [25:34]
Matt Van Horn
And so right now, look, it googled Papa John's ordering API reverse engineering, GitHub API endpoint, store locator menu, JSON, Papa John CLI, Python, pizza order automation. So this is, this is how it's kicking off, which is fun.
- [25:48]
Interviewer (possibly Peter Yang)
Got it, got it, got it. It's trying to find all the secret keys and like, you know, get our projects.
- [25:52]
Matt Van Horn
That's where starting. Look, Papa John's has a postman collection plus several unofficial wrappers. Let me dig into the ground truth endpoints. So now we're at GitHub. Yo. Contra Papa John's postman, Papa John's team GitHub, Brett Jurgens, Papa John's Thin readings. Let me read the actual wrapper source for ground truth endpoints and probe the live API directly.
- [26:10]
Interviewer (possibly Peter Yang)
So there is an API.
- [26:11]
Matt Van Horn
Maybe everything has an API. There's no official Papa John's ordering API. It's talking about the secret APIs. Not like you go to Papa Johns.com API. They're like, here's our documentation. Here's how to order pizza in your application.
- [26:24]
Interviewer (possibly Peter Yang)
Got it, got it. Okay, got it. Dude. Yeah. The world's going to run the APIs, man. No one's going to browse around the UI anymore. Just get the agent to go to the API and order pizza from me or do whatever I want.
- [26:34]
Matt Van Horn
Yeah, I get so frustrated when something doesn't have a cli, but then I just build it for myself. So I'll give an example where I trying to figure out how to talk about this without outing my company. That doesn't exist. But I want to generate a lot of AI videos from the command line while working on a different workflow. And Open Art is one of my favorite places where you can just choose, okay, I want to use C Dance, I want to use this video model. I want to use this image model. If I want to do Nana Banana. It's just like they have access to all the great video models. I have an account and image models, and so I have a cloud code skill that it's. The goal is not to generate videos, it's an unrelated thing. But one of the outputs I want is a generated video with prompts that are automated. And so I was like, all right, Open art, do you have an API? No. OpenArt, do you have a CLI? No. I was like, all right, printing press, OpenArt. And so it built it, it found all the secret APIs by sniffing and it worked. And I use it every single day now. Like, it's built into my workflows entirely agentically. I never have to go to openart.com anymore. It just literally downloads videos for me. Another one is Suno. So for like making music and you never have to go to suno.com again, you can just do it entirely through the CLI. And there is no API for Suno, so it's just completely sniffed and found by the printing press.
- [28:07]
Interviewer (possibly Peter Yang)
And these CLIs, are they also on GitHub or they're just on the printing press?
- [28:11]
Matt Van Horn
Yeah, they're all on GitHub, they're all on printing press. Everything links to GitHub, they're all open source. And people make PRs. So if things change, like dominoes, where they change their APIs, someone could just make a PR to the printing press library and then everyone gets the new domino's update.
- [28:26]
Interviewer (possibly Peter Yang)
Or you can build like some sort of routine or something. Just check, check for any changes, APIs or something. Yeah, that's amazing. You build like a platform to make cli, basically. And by the way, all this stuff is open source and free, right? So it's just like, it's a big service, man, to the world.
- [28:41]
Matt Van Horn
Yeah. And we keep hearing people like saying, like, hey, I was about to have my engineering team start on building our mcp because that's the cool thing to do. But then we discovered printing press. So we're just going to use you and make that our official CLI and mcp. I'm like, go for it. It's open source. Have fun. Which is amazing. We had a company, Kuala Lumpur, they put out a press release on the wire officially announcing their official CLI and skill. And it was literally, they just used the printing press. And I was like, this is awesome.
- [29:14]
Interviewer (possibly Peter Yang)
Nice. Nice, man, nice. First of all, I didn't realize you were not an engineer. But second of all, how do you come up with these crazy ideas?
- [29:19]
Matt Van Horn
You.
- [29:19]
Interviewer (possibly Peter Yang)
You Just like Saturday morning, you just wake up, be like, hey, I have like three ideas. CE plan, do three ideas.
- [29:25]
Matt Van Horn
That's it. That's it. The cost to testing and building something, it's like, so, so easy. I'll give an example of a stupid one that I spent probably too many tokens on for. For 24 hours. I said, CE plan. SQLite is 21 years old. I want to redesign SQLite for agents from scratch. Can you make a plan for that? And then it's like, yeah, of course. This will be great. And then spent a really long time, like 8ish hours, with very limited. I felt like I did slash goal, which I didn't even do. Like, it was just very aggressive. Like, we're going to make it better. We made benchmarks. Oh, this is faster, this is slower. I'm like, I don't know what's going on. And then it made it to the end and it's like, okay, here's the honest truth. Like, these three things are faster than SQLite, but these three things are actually much slower. I was like, okay, well, let's play devil's advocate. If I said to Trevin, who I built the printing press with, hey, I'm ripping out SQLite on every single CLI and replacing it with this today, what would he say? And he's like, yeah, he would kill you and this is trash and you should probably end this project. And so I did. I didn't. Again, I never read the plan. I never sent it to Trevin. I just. I felt confident in its answer, that it built something that was not good or useful. But I've also done that to make it better. Like, for example, with Agent Cookie, it's literally taking all your cookies and secrets and sending them to another machine. That sounds scary. I'm not hosting anything in the cloud. It's all through tailscale. Which I think was a smart security decision. But I said, CE Plan. One of my friends, who's one of the leading security experts in the world, just said that Agent Cookie is a complete nightmare and I shouldn't even launch it. And by the way, no one said that. Like, I was just trying to play devil's advocate. And it was like, you're totally right. And then, actually, no, it's fine. And. But I make this one change. And so it was like, it's kind of fun just to think that way.
- [31:28]
Interviewer (possibly Peter Yang)
So maybe you can show us the way that you actually work with agents. So it sounds like you actually just like, uc plan on several things at once. And maybe go play with your kids or something. And then like five hours later you come back and you ask it like an adversarial or stress test question to see if it actually makes any sense. Is that kind of horrible?
- [31:45]
Matt Van Horn
I mean I usually just have like six or so windows open, so there's always something to do. And I'm always carrying a laptop around with me. Like I figured out how to close it at least and like keep it in my bag. So my wife is like less like open laptop at school and now it's under my arm or in my bag at school, but it's still running, the agents are still, still going. And I just have a lot of these going all the time. And so I quit everything actually before I did this because there's some stuff I cannot share my screen on. So you only see like four tabs in there right now when I'm sharing my screen. But yeah, I always have good ideas, bad ideas, fun things, stupid things going at all times.
- [32:32]
Interviewer (possibly Peter Yang)
Got it. Okay, so maybe you have a couple of things that you're working on constantly printing press and then you have a bunch of zero to one stuff that you think about.
- [32:41]
Matt Van Horn
Yeah, yeah. And a lot of open source stuff too. So I've gotten really, really addicted to open source. So I'll share my screen for a second.
- [32:50]
Interviewer (possibly Peter Yang)
Tell me, because I want to be open source contributor too. How do you become a great open source contributor? I guess as also a not engineer, I kind of worry about shipping slop or like putting slop PRs in there. But like how do you get over this fear just to do it?
- [33:03]
Matt Van Horn
Yeah, you just do it. You just start. I remember I tweeted about it my first open cloud PR that got merged. Peter Steinberger was like, thanks M. Van Horn for the good pr. And I remember I tweeted and I was like trying not to be starstruck, trying to keep it cool, but I can't keep it cool. I can't believe Peter just said thank you to me. And my code now exists in openclaw forever. Like ah, like was like overly excited and that, that, that moment like I, I can ask an agent to look it up but I like, I would love to look at the date of that tweet compared to the velocity of my open source contribution of getting addicted to that adrenaline.
- [33:48]
Interviewer (possibly Peter Yang)
Yeah.
- [33:49]
Matt Van Horn
Has been crazy. So let me, I'll just scroll through here a bit. So these are my, my projects last 30 days which has pretty much 20,000 stars printing press for 3,000 printing press library 1.4 and then agent Cookie, which just launched. But then here are the projects that I've been merged to.
- [34:02]
Interviewer (possibly Peter Yang)
Python, dude. Yeah.
- [34:04]
Matt Van Horn
So Python go. OpenCV. I had one get merged to OpenCV this morning. Vercel. I'm the number five human contributor to agent browser Zed. I've had many verged, openviking, paperclip. I'm the number three human contributor. Superpowers. I'm number seven. And so Gemini CLI have had a bunch merge too. And this list. GStack. I'm the number three human Contributor on GStack. And this list doesn't end, by the way, like post Hog. Continue. Light Panda. It keeps going. Em Dash.
- [34:43]
Interviewer (possibly Peter Yang)
Let's take the GWS cli, the Google cli, right? How do you find what to contribute to? You just use it and find it annoying and then you just is like, hey, can you put a pr?
- [34:53]
Matt Van Horn
That's how it started, but now it's very, very automated. Like, I've built massive systems around how this happens because I have a company, I have my projects, and so literally most of my open source contribution happens while I sleep. And I've built up these systems over time to do this. And again, it was very manual in the beginning. Like, you asked the question how you could get started, right? Like find a problem or a feature you want and then say CE plan. I want to build this feature and how could I do that? And again it's. It's like I've been merged to hundreds and hundreds of projects and I've also been banned from projects. Like I'm. I'm banned from Rust, which makes me very sad because I love Rust. My last company, we did a lot of Rust, pioneering on the embedded front at June, and I'm banned and I appealed and they said, no, thank you
- [35:50]
Interviewer (possibly Peter Yang)
because you were merging Too many PRs. Oh, you. Too much AI. Okay, got it.
- [35:53]
Matt Van Horn
Yeah. So they're. I don't know if it's changed, but I literally got banned the week that Peter's thing got accepted me in openclaw. So I got overly excited and then like banned from Rust immediately. I don't think I've been banned from anything in a while. Like, I want to add value. But what was really fun, I can go with Python as an example. I can pull this up quickly. But so I had some very small minor things get merged into Python. And then I said to my agent, like, I want to build a major Python feature, like, similar to, like, I want to start a programming language from scratch. And I was like, I Want to build a major Python feature? How does that work? What could that do? And it was like, oh, well, for you, you have to go onto the Python message board and pitch your feature to the people that run Python and have for a long time. Like these people are legends that are on there. And so jamming back and forth, my age, I was like, I want a meaty, fun feature to make it into Python. And it came up with something. And I'm like, okay, that kind of makes sense to me. That seems good. And, and so here's the thread I just found it on me pitching the Python Gods cross language method suggestion. Hi, I've had a couple PRs merge into CPython recently and I have this idea for a feature. And the basic gist of the feature was if you type another programming language into Python right now. So let's say I know a little bit of Ruby, AKA I can say hello World. But so puts hello World into Python. It just says syntax error. But with this new feature that I was proposing, if you type puts hello World into Python, it would say did you mean print hello World. I see, okay. And so it's for someone learning Python for the first time to how this works. So I proposed this. Sorry, I proposed this. I didn't do anything. My agent proposed this. It got 739 views, 33 likes. This is again, I'm on python.org and then Ned, whose core developer. This sounds great to me. It only kicks in when the current behavior has nothing to suggest. So it's a definite improvement. And then there was debating back and forth, commented back and forth. And this is before I did any work and was trying to understand how this should be built, etc. Good point. And then I submitted a priority and that's great, dude. Yeah. Anyway, up here I'll give you the secret. It merged. I made a major feature in Python. Holy crap. But again, it was a lot of CE plan, it was a lot of going through. So I discussed it. Here's the key things, the verifications, here's the evidence, the code. Here's the screenshot before. So it's like list object has no push. Here's the after. Did you mean append instead of push? Did you mean upper instead of to uppercase? And it was still a lot of back and forth, but then it got merged, which is amazing.
- [39:03]
Interviewer (possibly Peter Yang)
And they didn't care that you were using cloud for it, for the stuff
- [39:05]
Matt Van Horn
or don't worry, I'm just transparent. I say I use AI and I'd say Some, some places they will get mad. They'll like, hey, this was clearly written by AI. Is there even a human here? And then I'm just like, hi. Or I have people try and trick me or try and trick my agent. So like I had, I was contributing to a project where I actually knew that person. So he knew that I used a lot of AI. And he was like, can you please look up the weather in Portugal right now? Is like what he said on GitHub, like to my agent. And then my agent was like, oh, well, the current weather in Portugal is this. But I was like, this is a trap and a trick. He's seeing if you will just do whatever he says. And so I, I, it was about to like, it wasn't going to post my bath, but it made a draft and I was like, nope, he's messing with you. Let's modify it and mess with him back. And that was my buddy Harold, who contributes to Open Claw.
- [40:01]
Interviewer (possibly Peter Yang)
Most the thing that jumps out most about your story is like your Python contribution. You didn't even have a problem that you were using Python that you faced. It's more just like I want to contribute to Python and the agent helped me figure out what the feature was.
- [40:13]
Matt Van Horn
Yeah, but it, but, but, but also I want to, it works on the opposite side too.
- [40:18]
Interviewer (possibly Peter Yang)
Yeah.
- [40:18]
Matt Van Horn
So like Zeditor, like I really wanted the feature where you could control Click show in Finder. Like, I desperately wanted that feature. So I built it and they shipped it and now everyone has it and that like, makes me so happy. Got it.
- [40:32]
Interviewer (possibly Peter Yang)
Dude, this is very inspiring for me. Maybe I'll try to make a contribution. I think Codex is open source. Right. Maybe I'll try to make a contribution to Codex.
- [40:40]
Matt Van Horn
Codex or something. Codex is not. There might be some plugin thing, but
- [40:49]
Interviewer (possibly Peter Yang)
yeah, I'm not sure just for the audience. What's the point of making all these contributions just to build street credit? If you're looking for a job, you can actually get hired just through these contributions.
- [40:59]
Matt Van Horn
Oh, it's crazy. I, I, people have interesting founders have reached out to me and said like, would love, love what you're doing, love the contributions you made to our company, would love to, to have you join us full time. And I'm like, I'm flattered, thank you. But I, I have my own new startup that I, I have that I've raised money for. But what's amazing is, is the, the people I'm meeting and the, the circles that I'm getting involved in. Like, that's, that's I think the most interesting thing for me is meeting the most interesting agentic builders in the world
- [41:33]
Interviewer (possibly Peter Yang)
just from like they pay on GitHub or something or pay on this year contribution.
- [41:37]
Matt Van Horn
Like it's just, just like there's like a secret network in any world. Like the most interesting founders all know each other. The most interesting PMs all know each other. Like the like all these like secret WhatsApp groups that have happened over the last decade like that exists in amongst the most interesting agentic engineers. And so what's amazing and I've been meeting so many incredible people and by, by being in the discords of interesting projects and like I, I call the Peter. Peter Steinberger's his crew. I call them Peter's disciples. And so like, like the, the. The stuff that they're doing is crazy. And. And I've. I've become friends with a lot of them which is really neat and very impressive and we're always sharing tips and tools and tricks and sharing each other's projects and contributing to each other's projects. So even though it kind of looks like it's just for the fun and the glory like this is making my future company more successful by meeting the most interesting people. And it's also incredible for recruiting like being able to know the best agentic engineers in the world right now. I've taken pride in my career as recruiting is one of the reasons that I've been successful so far in life and now I have have aligned to the people moving the fastest in the entire world at shipping.
- [42:55]
Interviewer (possibly Peter Yang)
It's amazing how many of these agentic engineers are also they have a decade of experience in engineering or the people like you who thinker stuff is it kind of a mix.
- [43:06]
Matt Van Horn
I'd say I'm the most incompetent person that I've met in terms of actual skills to success ratio. But what I am finding is a lot of people. I'll use Trevin as an example and I'll use him as an example then it's a funny story of how he got involved in printing press. But Trevin at least studied computer science in college. But for the last I don't know how long he's worked 20 years, 15 years he has been a product person VP of Product like chief product officer and now he's running compound engineering and number one contributor to compound engineering and he has not written to the best of my knowledge not written a line of software since college and look at the lines of code that he's doing. And again I'm incompetent. He's competent. Right. If you look at our base skill sets. But like, I'd say that there's a very different type of person that can be very, very successful right now.
- [44:04]
Interviewer (possibly Peter Yang)
That's amazing. Yeah. Because I think a VP of product, you definitely don't contribute engineering code. So it's good that he made the career transition.
- [44:10]
Matt Van Horn
Yeah. Yeah. But again, you need to be careful and thoughtful about not creating slop. And I think tools like compound engineering allow you to do that.
- [44:19]
Interviewer (possibly Peter Yang)
Last question. Have you ever launched something and I'm sure you have that just no one cares about. People think it's slop. Like, how do you get over that fear of embarrassment to get started with this stuff?
- [44:29]
Matt Van Horn
Yeah. So it's interesting. So I have two projects that are 100% shippable and ready and I'm afraid to ship them and that makes me sad. And I'm afraid to ship them because what if they're not a hit? Right. Like, I've been very lucky and fortunate that my launch videos have been getting a million views on last 30 days launch. A million views on printing press launch. I haven't looked up the agent cookie numbers. Definitely much smaller on agent cookie, but it serves its own niche, powerful purpose. But I have a game, like a video game. It's a sim game, like a simulation game for being a founder and starting an AI startup.
- [45:15]
Interviewer (possibly Peter Yang)
Yeah, you should ship that. That sounds good.
- [45:17]
Matt Van Horn
It is good. But the early feedback that I got from friends, and again, most of my friends are like crazy busy founders. They were like, okay, yeah, I'll play with this tomorrow. And then they never played with it or they played with it for five minutes. The amount of friends of mine that actually finish the game, which you can do in like 40 minutes, was. Was very low. And so like that, that made me, I don't know what the word is, just like sad. Like, I think it's great. Like, it's so fun, but like, does anyone else care? And like just that, like, fear of rejection is real. And so I'm. I'm not sure I probably should just launch. There's one brilliant game mechanic in it that I feel like if I launch it, even if it's a total flop, that the game industry should steal and be inspired by from this game. So maybe I should just launch it for that because I invented something crazy that I think will be a standard in the game industry once it's out there. But I'm fearful that no one's going to care that it's only Going to get X number of retweets and we'll see.
- [46:25]
Interviewer (possibly Peter Yang)
Okay, well, send it to me. I want to play.
- [46:27]
Matt Van Horn
Play it.
- [46:28]
Interviewer (possibly Peter Yang)
Yeah, that's how I feel.
- [46:29]
Matt Van Horn
Yeah, sounds good. Maybe you will inspire me to ship it.
- [46:34]
Interviewer (possibly Peter Yang)
So I guess, yeah. So I guess the lesson is like, even after all these successful projects, or maybe because of these projects, you still have the fear. The fear of failure is still there for sure.
- [46:44]
Matt Van Horn
But also, just build, just launch something. It's okay. Even if you build something for yourself. Even if I had no users of Agent Cookie, if no one used it, I get value out of it. It helps my agents be so much better. Like, that's magical.
- [46:58]
Interviewer (possibly Peter Yang)
Yeah. I do think it's very important to
- [47:00]
Matt Van Horn
solve your own problems.
- [47:01]
Interviewer (possibly Peter Yang)
So, like, at least one user is satisfied. Yeah, that's what we need to do too. Cool, man. Well, dude, this is super inspiring, man. This is super inspiring. Where can people follow your journey?
- [47:13]
Matt Van Horn
X M. Van Horn on X. Yeah.
- [47:16]
Interviewer (possibly Peter Yang)
X. And also your GitHub and all your crazy open source contributions.
- [47:19]
Matt Van Horn
Yes, M. Van Horn, anywhere but X is where. Where my stuff happens.
- [47:25]
Interviewer (possibly Peter Yang)
Cool. All right, man. Well, you definitely inspire me. I'm going to make some open source contributions this afternoon.
- [47:31]
Matt Van Horn
Great. I have projects. We will take your PRs. I will look at them personally.
- [47:36]
Interviewer (possibly Peter Yang)
Okay.
- [47:37]
Matt Van Horn
Don't submit garbage or slop, please.
- [47:40]
Interviewer (possibly Peter Yang)
I'll use a plan to make sure it's not garbage.
- [47:42]
Matt Van Horn
Yeah. Yeah.
- [47:43]
Interviewer (possibly Peter Yang)
All right, Matt, thanks.
- [47:44]
Matt Van Horn
Amazing.