The creator of OpenClaw: "I ship code I don't read"

22,893 words · about 92 min to read

0:00 What if you could merge 600 commits on a single day and none of it was slopp? This is what today's guest, Peter Stainberger, the creator of Claudebot, claims he's doing. Peter is a standout developer who built PSP PDF kit, the PDF framework used on more than 1 billion devices. Then he burned out, sold his shares, and disappeared from tech for 3 years. This year, he came back and how he builds and what he's doing now looks nothing like traditional software development. In today's episode, we cover why he no longer reads most of the code he ships, and why that's not as crazy as it sounds. How he is building Clawbot, his wildly popular personal assistant project, which feels like the future of Siri, the closing the loop principle that separates effective AI assistant coding from frustrating vibe coding. Why he says code reviews are dead and PR should be called prompt requests, and many more. If you're interested in how the software engine workflow could change in the coming years thanks to AI, this episode is for you. This episode is presented by Statsig, the unified platform for flags, analytics experiments, and more. Check out the show notes to learn more about them, the pragmatic summit on the 11th

1:00 February in San Francisco that I'm hosting with them, and our other season sponsors. Right, Pete, welcome to the podcast.

1:07 Thanks for having me, Gay.

1:08 It is awesome to meet you in in person.

1:11 Yeah, and I almost messed it up.

1:14 Yeah. What what what what happened? You lost track of time. Does that happen often? And how how so? Um, not usually. Not usually. Um, this is an interesting time for me cuz I cuz my latest project is is blowing up. Cloud, right?

1:29 Clawbot. Yeah,

1:30 Cloudbot.

1:31 I'm struggling a bit to get enough sleep, but it's it's it's interesting. I never I never had a community blowing up so fast and it's just incredibly fun to work with.

1:39 So, before we we get into Clubbot and all the fun stuff you're doing, I wanted to rewind all the way way back. You create a PSPDF kit which is used I think on more than 1 billion devices users. If if you see a PDF render you probably see

Tinkering with code from an early age

1:54 that. But even before that how did you get into tech? Oh my god. How did you get into tech? So I'm from rural Austria. Um always more being the introvert. So eventually I we always had like summer guests and one of them one of them was a was a computer nerd and then I kind of got hooked with with like the machine he had and begged my mom to buy me one and ever since then I this was in high school or so.

2:25 I guess I was 14.

2:27 Yeah. And ever since I started tinkering, like I can remember the earliest thing was like I stole an old DOS game from my school and then wrote a copy protection for the floppy disc so I could sell it. It took like 2 minutes to load. I was just always tinkering. Also playing a lot of computer games of course, but like building stuff almost feels like playing a computer game. Like definitely right now it feels better than Factorio.

2:54 Oh, when I started out, I I I I read like the equivalent of bash scripts for Windows and then I did like websites. So, I guess a little bit of JavaScript, uh, even though I had no clue what I was doing. And then the actual first language where I I had to learn how to build things is when I started university. And I never met my dad. And I come from a from a poor family. So, I always had to work. like I had to had to finance my own studies, right? So when other people were were having holiday, I just worked full-time at a company. So the the first real job I had was was in Vienna was supposed to be 1 months and then they kept me for 6 months. It was just a bridge between military and and my university. And I kept working there for like I think 5 years. And I remember the first day they they gave me this huge book. maybe that huge and say Microsoft MFC. Um I still have nightmares and I

3:56 got I got like I was like this is terrible like for for the next win I just I just silently used net I just didn't tell them and like a few months in I just told them yeah but the text I I did a few modernizations but then it was too late. I did this a few times in this company. I don't know why they kept me because because my [ __ ] worked. So uh I did net and actually I actually actually dig it.NET 2.0 had like generics. It took insanely long for the application to to to launch because like everything was compiled at first start and like your hard disk was like if you remember

4:33 so so you you how did you stumble into both iOS and where did the idea from PSPDF

4:39 come from?

4:40 Not even yeah the first one the first

Building a profitable app by parsing HTML

4:41 one wasn't even available in in in Austria. That's true. Yeah.

4:44 A little time goes went on and I was at university and a friend showed me the iPhone and I I think I I touched it for a minute and then immediately bought one like like this like it it clicked when I felt it and and to me this was like a holy f moment cuz it was just like so different and so much better. So I got I got one. I was still not thinking about building for it, you know. That was that was

5:14 when was this 2009 10 something like that.

5:16 Yeah. Yeah. And then I I used their browser. I I can see the story. I was I was literally driving in the subway.

5:23 Mhm.

5:24 And by the time I was using a gay dating app and this was iPhone OS 2.

5:28 Yeah.

5:29 So So I I

5:30 long time ago

5:31 I typed this long message. I pressed send and we were just going into a tunnel and the JavaScript disabled the send button and then an error message

1 Highlight · Scraping HTML to build an unauthorized iOS app

5:38 came but there was no copy paste. There was no screenshot. So I was just like and I couldn't scroll anymore because like scrolling was disabled. So like this long message was like a little bit emotional was gone and I was so mad. I was so mad. I'm like what the hell? I went home and I downloaded Xcode. That's that's where that's where the window came and I was like where is the ID? So it was like like I was like this is unacceptable. I basically hacked the website. I used regular expressions to like download uh to parse the HTML which is like totally not something you should do and I built an app and I used I used iPhone OS3 beta with like core data in beta regaxit light. I used a hacked version of GCC that backported the blocks compiler so I could use blocks in iPhone OS3. It took me quite a while until anything worked because like I had no idea what I was doing and I was like using all kind of like beta tech but eventually I I got it to work and I I I

6:38 wrote that company was like hey I'm making an app. What do you think about it? Got no response of course. So I was like let's just put it in the app store.

6:46 And this was for the dating app right?

6:48 Yeah.

6:48 So you just like you know you looked at you saw their APIs you could just like easily like build a client on top

6:53 API. It was HTML.

6:55 Oh

6:55 I was I was just literally parsing HTML.

6:58 Oh. So, so you kind of parse the HTML, kind of turn it into your own, you know, like you use it as an API. Oh, clever. I mean, this was back in the day where no one thought this this would happen, but I made I put it in the app store. I charged five bucks for it and I made like 10k in the first month. And I had no clue what I was doing. So, and there was like so many complex tech stuff. This was very early on where there was a lot of weird forums on Apple.

7:23 So, I just put in the the bank account of my grandpa. And then one day my grandpa called me, "Yeah, something is weird. Like I got this huge payment from Apple." I'm like, "This is mine. This is mine. Don't touch it." But the the funny thing was when I this blew off and I remember I was in a club one day and like I saw someone using my app and I was so proud and I wanted to like tap him on the shoulder and say I built this and I thought like that would be really weird. So I didn't. And then I I I went to the company I worked for for 5 years and told him like I'm going to pursue this. This is this is really exciting.

7:56 And my boss was like mocking me.

8:00 Oh, really?

8:01 They're like, "Oh, you're making a mistake. This is a fat. This will not go." Blah, blah, blah, blah, blah. And you know what that got me? That's what you call a chip on your shoulder. I'm like,

8:11 you know, one day I'm going to I'm going to have a company that that's worth more money than yours. Well, it took me eight years. So, I I got hooked. Like, I I worked I I'm I'm a little bit of an addictive personality. So which you see again right now. But I I worked a lot on this app. I learned I learned in in high speed and this was also the time where I started Twitter and that was usually hugely influential for my career. I made this app actually quite good and then one day I was at a party at at 3:00 a.m.

8:45 um slightly intoxicated and I got a call from a US number. The guy on the phone was like, "Yeah, hello. This is John from Apple. Yeah, there's a problem with your application. Like some people reported pictures and that was it. That was the end of my app. Um,

9:02 it was good until it lasted

9:03 and I was just I just quit my my job and was like, well, f you Apple.

9:10 I did freelance work. I was at dubdub. I was introduced to

9:16 dubdubdubc.

9:17 Yes. Sorry for the insider terms. I was introduced to someone as one of the best iOS developers in Austria at a bar at 2 a.m. in San Francisco and then basically got a job in the US and then I moved to the US for a while and then I I went to the Nokia development days. This is all like stone age by now. My god. And then someone came up to me and said, "Yeah, they built this app somewhere in Eastern

Rewriting a buggy viewer into PSPDF kit

9:42 Europe and it works but it crashes sometimes and it was like was like a magazine viewer, right? This was back when the iPad just came out and Steve Jobs hit that like this is the savior. So everybody was building magazine apps and I was like that sounds like an

2 Highlight · Rewriting unmaintainable code into a solid product

10:00 interesting short-term gig. And I I was like okay I I'll I'll I'll help you out. And I opened the app and it was like oh the worst code I've ever seen in my life. It was literally one file with like thousands of lines

10:17 of Objective C. Yes. Where they used windows as tabs. I I didn't know this worked. I was surprised this worked at all, but it felt like a a house of cards. And I I tried to I tried to surgically fix things, but like as soon as you would touch something, something else would break. Um so I got it I got it somewhat stabilized and I told him like, "Look, this is this is like madness. Um I'm going to rewite this for you." Yeah. But it took half a year. I'm going to do it in in a month. Well, it took me two months. Um, I wasn't that far off and and then here I was working on a on a PDF viewer. You know, on every technical problem, the domain is is I wouldn't say like completely unimportant, but you can always find interesting problems in every domain.

11:04 And there was a lot of interesting problems because you had a C call that would render a PDF that would maybe take 30 megabytes, but the whole system had 64 megabytes. So if you're not very smart and like very careful what you do in the background and when the OS would just kill you. I got really fixated at like making it good like when the rotation is like that the page would like animate and so so you know I I like I like those details. I spent way too much time on that. That's why I took two

11:30 months instead of one. But the end result was it's good. Um and then I I I worked with them for a while and then a friend texted me up. He's like, "Yeah, I'm working on this magazine app and it's really hard." I'm like, "Yeah, no way it's hard. I know." Oh, like I did it.

11:43 You just built one.

11:44 And and he was like, "Can you can you can you get me the code?" I'm like, "Sure." So, I sold him like I extracted the the part that was PDF from from this magazine app. And I I made sure I made sure like the other person was okay. And then I sold him that. I was like, "Well, if he's interested in that, why let's not try to sell it to other people." I used a a WordPress template and mutilated it to run on GitHub pages and then and then when you did the the fast lane flow at the end you got a Dropbox link to my personal Dropbox with a source code sip and I put this on one afternoon and I I tweeted it and and then in that week three people bought it and it was like I guess 200 bucks but back then and for me this was like amazing and not only I got like three people who just bought it and like 10 emails who comp 10 people who complained about uh because they wanted it but it didn't have the features they wanted. You know, it's like I got nerd sniped. I was like, "Oh, I didn't have text selection." Oh,

12:45 how hard can it be? 3 months later. Oh, yeah. It's really hard. Text selection in a PDF specifically.

12:52 Yeah. Yeah. Yeah. Yeah. You know the saying the saying like uh the companies are built by young people because they don't know how hard it is.

12:59 Yes.

12:59 Yeah. Yeah. I had no idea what an insane madness this file format is. Peter was talking about how some problems look deceptively simple. PDF rendering is a good example. You look at it and think, how hard could it be? And then you spend months on edge cases that you didn't even know existed. This looks easy until you build it pattern shows up in other places, too. Internal tooling for feature flags and experimentation is a classic example. Teams often underestimate how much work it is to build infrastructure around these tools.

13:27 There's a reason big tech companies like Uber invested years into building internal experimentation and feature flagging systems. Which brings me to Static, our presenting partner for the season. Static gives you the complete toolkit without building it yourself. You get feature flags, experimentation, and product analytics all in one platform tied to the same underlying user assignments and data. In practice, it looks like this. You roll out a change to 1% of users first. You see how it moves the top pipeline metrics you care about, conversion, retention, whatever is relevant for the release. If something goes wrong, instant roll back.

14:00 If it's working, you can confidently scale it up. Companies like Notion went from singledigit experiments per quarter to over 300 experiments with static. They shipped over 600 features behind feature flags, moving fast while protecting against metrics regression. Microsoft, Atlassian, and Brex use static for the same reason. It's the infrastructure that enables both speed and reliability at scale. Static has a generous free tier to get started and pro pricricing for teams starts at $150 per month. To learn more and get a 30-day enterprise trial, go to stats.com/pragmatic.

14:31 And now, let's get back to Peter and why rendering PDFs was a surprisingly hard problem. But now, I remember there was a few weeks ago someone emailed me. They did something PDF and they wanted my help. And I just wrote them like, I'm sorry, like I I did my deed. I I know more about PDF than any any sane human person ever should know. And I went I went to therapy. Good luck. But that took off and and I just I while I was waiting for my visa, I I worked on this project and and it just kept on more people kept on buying it. And you know, it was like it was like summer. I was I was like lying at the lake and got another email that someone bought it for 600 bucks, 800 bucks. I just up the prices as it had more features. And by the time I I went to San Francisco to work at this company, it already made more than what I made there. But my whole life was I still I still thought like I have to be there, you know.

15:28 So I I did it. And also interestingly at this company, I had to

15:32 So what would you say that you moved to San Francisco?

15:34 Yeah. And and of course also it ended up being something where I had to build build something with my framework at that company too. But you know startups are not like 8 hours. They're a little more. And my personal project was also a little more. So my sleep was a little less. And then eventually after 3 months uh Sabine my manager came over and said this Peter are you okay? And they gave me a choice uh to either keep working at this company and drop my project or or vice versa. And I had one week to decide. The counter was one week to stay there or leave the country because I was on a on a complicated visa.

16:16 And well, the decision was quite easy. It's like, yeah, I want to do my own thing. And then and

16:21 and at this point, it was already taking off. You already saw that this is there's a big business here. It it will probably pay you as much as your US job would have paid. Ah,

16:31 it was never money driven either.

16:33 It was more about what what were you driven by? I want to make stuff that other people find amazing. Like I I I love tweaking the details. I love those little delights. It wasn't even that the space there were competitors in the space. But my angle was always like I built something as if Apple would have built it like like with with like all the love and care and and polish and and those little delights that a lot of people in the industry don't get. Uh, so even though we had competitors that had way more features and were around way longer, my company was more successful and my product was more successful cuz developers tried the the different ones and mine just felt the best. I think software is all about about how it feels much more so than the than the feature set. Like why did we buy Apple stuff? It has more features than than Windows. Um, but it it feels better. So, how did you go from like you you left this company and you were

17:34 building this PDF component that started to sell. At what point did you hire the first person realizing, okay, there's something more to this? when I went back to Vienna then I was like okay I have to go all in and that's where I I started working with freelancers a little bit and way too late to be honest also like I I could have could have hired much earlier but you know you know it's it's it's a big step and that's kind of where it it started having a life of its own and I spent pretty much 13 years of my career um building this product with this weird name that I never changed because I took me like I thought like five minutes about it and then stock PSPDF kit.

18:13 PSPDF.

18:14 They finally they finally did a rename, but I wouldn't have I wouldn't have renamed it. But now it it

18:19 it's it's a mouthful, but it's very unique.

18:22 Well, you get it if you do Objective C because it's just a name space.

18:25 Yes.

18:26 Um and by by the time it it made perfect sense, my marketing my strategy for marketing was always I only care about the developer. Like I know like upper management does the decisions but if I can convince the people inside the company they all do the marketing and lobbying for me that worked really well. We never did like cold cold emails or aggressive. It was all inbound. All we did was like make good stuff and write insertful technical blog posts that and I went to a lot of conferences like the for me important it was okay if if people understand that the people who built this product are like know what they do and and love what they do that reflects on the product and that that worked really well.

19:13 And then what was the text type behind PSPDF kit? Was it objective C? Was it later swift? Were there other technologies like C or or anything else?

19:20 We eventually expanded to to all the platforms. Um big shift was the switch out renderer which was and is still quite buggy to like a big C++ one that when then we used across all the frameworks. Um we we were we were really early with web. We were one of the first PDF frameworks that ran in web assembly. And I I did the most clever thing that it was in the very early days when Web Assembly was just taking off and we built a benchmark and that benchmark was eventually used by by Google and Microsoft and Apple and I basically had all these companies like working really hard on making my renderer faster cuz cuz they used their benchmark as one of their benchmarks and the benchmark was just like rendering our stuff with our [ __ ]

20:03 Ah nice. And then a as as a company grew, one thing that I remember about PSPDF kit, you did write a lot of blogs and one blog in 2019, so this was about like I think you know year nine or 10 in the company. It was about how the team worked and you you mentioned things there like every feature starts with a proposal. You mentioned that you are conservative because it's a it's a big API that that people use. You want to be careful things like the boy scout rule such a factor. How how did you kind of put together the the culture of of this

Navigating enterprise sales with custom pricing

20:36 now this team which was now closer to 30 or 60 people?

20:39 We were actually 70 when when I sold my shares and now it's almost 200. And I I knew right from the get- go it's like I'm not going to find the people that I need in Vienna. So it was always just like remote first and eventually we we landed up with some kind of hybrid model uh which made things a bit more complicated. I learned a whole lot on the go. Like I I never had the urge to be CEO. I always was coding. I I brought people in to people that that helped me a lot with other other parts. Uh and on the business side, I can do it and I I think I think I'm I'm quite good at it, but I just don't enjoy it. Like even on sales calls where you like have to like think of all the magic number how much it would be worth because that's how enterprise works. Gh worse. Peter just said, "Ah, the worst about enterprise sales. Because selling to large companies, enterprises, is as tricky as it gets." Not just because you need to get pricing right, but because of all the enterprise features that you need to

21:39 build. And this leads us nicely to our season sponsor, Work OS. If you're building with AI agents or automation tools, here's a problem most teams don't think about at first. Once an agent can take actions on your behalf, you need to control what it's allowed to do. And traditional O just wasn't designed for that. That's why work introduced MCP off which gives teams a way to authenticate AI agents with explicit permissions auditability and enterprisegrade security. Instead of sharing over scoped API keys, you can define clear boundaries for the data that agents can access and the agents they can perform.

22:11 If you're building AI powered features and want to share fast without compromising security, check out workeros.com/mcp. And with this, let's get back to Peter and enterprise pricing. But that's also the only thing that that really works on on a model like this.

22:26 Yeah. You mean enterprise sales

3 Highlight · Why enterprise pricing requires custom quotes

22:27 specifically right?

22:29 Meaning custom pricing. So So can you tell us for for for you know devs listening who go to a vendor's website and they're frustrated that there's no price. It says call us or schedule meeting. Why that is?

22:40 Oh, that's that's why because we going to look at your company and then just take the dice and and and think about the number that you're probably willing to pay. And that sounds horrible. But also when you have a product where you can't really tear it down to a specific number like it it's it makes a difference if uh a freelancer contacts us or one of the big Fortune 500s. Let's not say names

23:04 because the usage will be different. The value they get out of it will be different and charging the same. You would either exclude one or the other. If I if I if I go too low, they're going to see this fishy. It's like uh procurement for like 500 bucks. we're not going to even start the process. And if we we target it too high, we're going to lose those people. So So as horrible

23:25 and unfair this process seems for some kinds of products, it is the it's the most fair way after all. You know, you know, on software there is I would say there's like four axes. There's like easy and hard and interesting and not interesting. We were very much in the not interesting and hard part. If you build something that every developer wants to build, it's going to be a hard cell. It's it's a hard sell anyhow. Selling anything to developers is a hard sell.

23:51 Yeah.

23:52 But if it's if it's too easy or too interesting, good luck. But if it's

23:57 oh god, I don't want to do this and oh my god, this is hard. That's a good spot to be in. So So I found a really

Solving extreme edge cases in PDF parsing

24:04 interesting niche and there were just an infinite number of complex problems.

24:10 You you you need to tell me tell me one or two hard things about parsing PDF. How hard could it be? There's a specification. I'm an engineer. I know specifications. I What's so hard about it?

24:22 I mean, there was this one example

4 Highlight · Scaling data models for extreme PDF edge cases

24:23 where, you know, like PDF has have links. So, like there's like there's like a table of contents and you click on it and it goes to like page 37. So, I built this whole model with the assumption, oh yeah, maybe there's like a 100 or 400 links in there.

24:34 And then we got this one customer who like paid really good money. And then I was like, oh, it takes 4 minutes to load a PDF. What the heck, guys? And I looked at it and it was like a 50,000 page text bible from Canada and it had

24:46 50,000 pages.

24:47 It had like more than 100 links per page,

24:50 500,000 links.

24:52 My data model completely exploded because my assumptions were off by a number of what 1,000. But by then you have like a mature product with an API. So like how do you how do you completely redesign the internal part without breaking things for everyone? like suddenly everything has to be lazy where where before parsing 100 one was easy but now they were like this was like so difficult to keep it working for people um I think I spent like two months just on that completely redesigning like the internals and like making sure it's still easy for people they don't have to know what we what we load easy what we load lazy or if you copy this thing it it still has to like have to keep some connection

25:36 it needs to keep the references and and some of those things.

25:39 So, so I and I I love to do support and I think that that that also a confining factor why the company worked cuz if if you send a ticket and then the the the CEO replies and helps you out um that has impact and my my strategy was always like I always used to list in reverse cuz if you if you send a ticket and you get a reply within 5 minute that's magical. If you wait one or two days not much difference. Yeah.

26:08 So, yeah, this this was one of the problems where I worked two months and I finally got it down to almost like this.

26:15 Mhm. That must have been satisfying.

Engineering blogs as hiring and documentation tools

26:16 And and it was this was very satisfying.

26:19 And you were writing a lot a lot of the code or you were involved in in in a bunch of the code like obviously a big big team was now here, but you were still kind of overseeing it, right? You're in the details. I mean, of course, I had a a really great team and and some parts I was more involved. I was always more involved in mobile because that's where my my heart was, but I was always very deep in the tech and and the the marketing side, the business side. I had like Jonathan's help, I had marketing help. There was I I I found good people.

26:44 The the thing is if you like the blogging and writing about how you solve interesting hard problems will help you hire interesting people that want to solve interesting problems. This is what I remember at PSP PDF kit that your blog was every now and then it admitted it to hacker news as well but it was just interesting to read and I couldn't name again I I'm not went into PDFs but if I had to say something a PDF I would have said PSPDF kit because they're the only ones where I read interesting engineing blogs about how you optimize your ship is still there by the way I I I myself also sometimes ask myself like hm interesting do more companies not see this or is is the question that you you need to be a developer who's either the CEO or or up there who just likes doing this. And by the way, did did you ever write this thinking this will be helpful or you just wrote because you got something out of it like putting out that you solved this hard problem?

27:39 I like sharing and and and like inspiring people. Um there was sometimes even conflicts where we were like should we write about this because it's like a little bit of secret sauce but I just never listen to those voices too much. Um I just you know there's also like when you when you write something down there it's this principle of like you understand it but then if you want to teach it you really have to understand it. So to to me it was also a little bit like oh yeah I worked on this really hard problem and now I want to like preserve it and like help others.

28:13 So, so, so I I got I got a gig of it of of course I liked the attention. Um, but really it it was this sometimes I just referenced a year later

5 Highlight · Forced technical writing for documentation

28:25 to my own post like yeah this this is a this is both company documentation. This is like my own lookbook. It's helpful on so many ways and a lot of those bigger companies. Oh, they put on too much red tape. There's a lot of developers who don't really like to write. So I I I forced everyone once a month a full day just to write a blog post.

28:48 But you gave them the time. You're like that day you don't need to do any other work but write something.

28:52 Yeah. You have Yeah. You have a day to come up with a post. Ah a day is is quite much actually. I mean when I'm nowadays when I write posts it still takes me

29:03 a few hours. I don't want to dwell too much on like the I think the the the starting time of the company is the most interesting. the then the the growth phase, you get more red tape, you get more people. It's much more gardening your product instead of like doing doing wild hacks

29:24 um and more iterative. So, so, so it got a little bit less interesting uh over the years and there was like more people drama cuz the more so the more people you have, the more issues there are and I didn't enjoy it that much and I was really really burned out. What what

The root causes of startup burnout

29:40 burnt you out, do you think?

29:42 I was just burning too hard. I was working most weekends. I I I tried to shuffle all my managerial needs. And you know, as a CEO, you're basically the waste bin cuz everything everything that other people don't manage or or can do or or mess up, you have to fix. And it's also quite lonely because you you can't openly talk about a lot of things. I mean I I I I structured the company to be quite open but still like you cannot you cannot be negative. You have to even if even if like even if like really bad stuff happens. I know there was like there was like one weekend where my my co-founder called me at at 5:00 a.m. and told me like yeah there's this big airplane company and their planes are down because our software is crashing.

30:28 That was a very interesting weekend until I could like I disassembled their their app and did proof that they messed around with our source code to triggering a triggering a license key fall back. Uh that eventually like caused issue they had. But that was like a if they sus company's gone and more moment. Um and that's just on top to all the additional stress and there were quite a few of those things. You can do

6 Highlight · Burnout comes from lost conviction

30:55 that for a while. And I I also believe like burnout doesn't necessarily come from working too much. It it comes more from or at least for me when you when you work on something but you don't believe in it anymore or you have like too many conflicts and and we also had a we did fight a lot in the team uh with like management team and by the time I I made this mistake and I thought you have to like lead a company more democratically.

31:23 Um so that was also something that burned me out. I wouldn't I wouldn't want to miss it for a while, though.

31:27 Yeah. So, you know, from from the outside, it seems you sold your shares,

Overcoming tech stack rust after a long hiatus

31:33 you made enough money to not have to work again, should you not choose. And for a lot of people, like, you know, people who are starting out their business or or one day want to start a business, this sounds like the absolute dream. Like, this is I guess what we know realistically that most people will not make it, but if you make it, I mean, you've kind of like I guess you know, checkbox done. You're kind of it's a little bit if you're like climbing on a wall and you ring the bell, you're done. And then what I noticed is from the outside again on your blog, the blog post completely stopped for several years. What what did you do uh in in this time uh and and what what what did you learn in this time, you know, before you came back to to where we are now?

32:09 I needed a lot of time to decompress. I I catched up a lot on the things I thought I missed. I I a lot. Um there were months where I didn't even turn on my computer. And for a while I was I just didn't had this feeling of like what should I do now? Like like I definitely was like why border? You know, you're not you're not supposed to to retire so early or like have so much have such a good exit that you never have to work again. That messed with my mind quite a bit. That

32:44 that was some that was some hard years. And then in in April I was like I there was this idea that I had years ago and even a side project that I started I was like oh yeah I want to I want to continue on that and then after after after more than 3 years I just sat back to my computer and and started hacking again. But the thing was this was like a a Twitter analytics thing and it was written in in Swift and Swift UI and back then I already knew this would have would be so much better if I would build as a website.

33:20 So So was this an existing idea that you kind of had at the back of your your mind something something Twitter analytic?

33:26 Yeah, it was just like something I wanted to build for myself because uh because it didn't exist and then

33:31 even three years later it didn't exist. It still doesn't exist. Uh it it kind of does but I got a bit sidetracked. So I I went back and I I wanted to build it in in with web tech. But web was really was always even at the company the one thing that I looked into the least because I had I had someone really smart who took care of of of that side in the company that I brought in Martin. So I never had to worry about it. That was one of the

33:57 You're not hands on with React or any of that stuff.

33:59 Yeah. And when I came back, I was like, what's a prop? You know, that that level where you really where and you know, this is like this is a a trap I see with many developers. The the better you get at one technology, the harder it is to jump somewhere else. It's not that you can't do it, but it hurts so much. You're like like I can I can program in in in in Apple stack. and program blind. But then in that stack, I have to Google the most mundane stuff and it it just like it just hurts. You you you feel you feel like an idiot again.

34:33 Yeah. And and I guess the more experience you have, it it kind of sucks feeling. I mean, I'm sure you say embrace and and all that, but it's it's not great. You're not as efficient. You know that you could be faster, etc.

34:43 Yeah. So So I came back and I was like, gosh, there has to be there has to be what is this AI? What is this AI stuff that every that that people are dismissing? let's look into this. Yeah. And in April, a lot of us were the specific probably for rightfully so, but and I and I and I like and I to a degree I I credit those three years where I basically didn't turn on my computer because in those years you guys checked out AI and learned that it's crap.

35:13 Yeah. the the people who like I was about to say so you missed out on you didn't do the beta of GitHub copilot you know glorified autocomplete which is GPC3 or or maybe not even there was then of course 3.5 which is a big jump and it it got incrementally better than GPT4 and so by the time you came back what tool did you first use when you cuz you missed out on like two years of like like devs us devs using dismissing finding some niche use cases for it

35:41 oh cloud code so you start with cloud mode that I think came out.

35:45 It just came out It came out in May, but there was a beta beforehand.

35:48 Yeah. Yeah. I think they had something Didn't they have something in February already?

35:51 They had a beta from February. Correct.

35:53 Yeah. So, so

35:54 so clock was your first you you come back after like a you know hiatus and you immediately turned on clock code and you missed everything else before. and and and you know, you know, it was like I I remember I took this big messy site project that I built and I have this browser extension where that that converts a GitHub repository into one big markdown that was like a 1.3 megabyte markdown file and I dragged it into into Google's CI studio with Geminina 2.5 or two to something and I typed write me a spec and it generated those 400 lines of spec And I dragged this back into cloud code and I was like build and then I continue continue continue and while I was like working on other stuff, you know, um and eventually told me like it's 100% production ready and I started it and it crashed. I'm sure we can all relate to the story of the AI saying the code is production ready then crashing. This is

36:55 a pretty funny and innocent story, but I personally don't trust code that AI generates without verifying it. And this leads us nicely to our season sponsor, Sonar. So, let's look at some data. A new report from Sonar, the state of code developer survey report, found that 82% of developers believe they can code faster with AI. But here's what's interesting. In the same survey, 96% of developers said they do not highly trust the accuracy of AI code. This checks out for me as well. While I write the code faster with AI agents, I don't exactly trust the code it produces. This really becomes a problem at the code review stage where all this AI generated code must be regularly verified for security, reliability, and maintainability. Sonar cube is precisely built to solve this code verification issue. Sonar has been the leader in the automated code analysis business for over 17 years, analyzing 750 billion lines of code daily. That's over 8 million lines of code per second. I actually first came across Sonar 13 years ago in 2013 when I was working at Microsoft Skype and a bunch of teams already use Sonar Cube to

37:56 improve the quality of their code. I've been a fan since. Sonar provides an essential and independent verification layer. It's the automated guardrail that analyzes all code whether it's developer or AI agent generated, ensuring it meets your quality and security standards before it ever reaches production. To get started for free, head to sonarsource.com/pragmatic. With this, let's get back to Peter and how AI agents cannot exactly be trusted.

38:21 Then I had then I added added an MCP so it could use the browser. I think the player with MCP was already there and it looped a few more hours and then I had a I had a Twitter login page and it it did something. I It was not great, but it did something. And to me to me this was my holy [ __ ] mind-blowing moment.

38:42 Yeah.

38:42 This and this was like in April or May this year, right?

38:45 Yeah. It was it was it was just good enough that I could see the potential and I I understood it's like

38:53 yeah this is this is where it's going and and from that moment on I I had a few months where I had really trouble sleeping and I in

39:01 I I remember because once on Twitter I

Shifting from line-by-line reading to architecture

39:05 sent you a direct message. I was up early for valid reasons, you know, my my kids or something like that. But it was 5:00 a.m. and I I sent you a message on Twitter and you replied immediately. And I was like, "Why are you up?" And he's like, "Oh, this is usual. Like I I usually I'm still usually awake." And and I asked like, "Why?" And you said like, "Oh, I'm I'm just like using Claude and it's really really addictive." And I was like, "Really?" And you're like, "Yeah, I'm not joking. Like it's really good." And I think that was the thing. You said something or wrote something like just one more prompt. like you told me how like what what made it so addictive or or what what still makes it so addictive?

39:40 Oh, it's the same economics as as you go to a casino. That's that's it's my little slot machines, you know? You you you press the trigger and ding ding ding ding ding and it's like nope. you you type in the prompt and it it will like and it does it does crap or it does something that actually blows your mind and it's this

40:00 and and you're saying it it blows your mind as like you're a really experienced developer like it's it's not easy to blow your mind, right? Like you you you've seen good code you can differentiate like crap code, decent code, good enough code like you have a bar, right? It's so funny, you know, in my company, I used to obsess over every detail, every spacing, every new line, the naming. I spend so much time bike shedding. And in retrospect, I'm like, what the heck? Why did I do that? Like, what's the point that the customer doesn't see the insights of of course like it has to meet certain certain standards. It has to work. It has to be fast. It should be secure, but like how much did that bike [ __ ] there is like stupid.

40:47 You say that, but then you also just said that people loved PSPDF kit because it was the most polished. It worked the best. Do you not think that that that amount of caring bike shedding as you call it? Being obsessed. It sound like you were keeping tech depth at bay, you know, like being obsessed with white spaces is is not going to be messy. And we know it's not just the white spaces. we know you're going to care about testing and all that. Like it sounds to me that PSPDF kit like you know like what I see is you were not just building a product that was great UX but you built something that had a really good hygiene and that's how it could be high performance and all that is how do you think about it?

41:25 Yeah. Yeah. Yeah. To a degree. Yes. And and even now like I I I mean like my my last blog post was a confession that I I ship code on read and

41:37 yeah we have to talk about that

41:40 and at the same time I spent so much time to like restructuring I mean I mean like even even today like I I really wanted to get this PR in where it was like 15,000 line change

7 Highlight · Most app code is just data plumbing

41:55 where in my I moved everything over to a plug-in architecture where I was so excited about and I care a lot about the structure. Did I read all the code? No, because a lot of code really is just boring plumbing. Well, what are most apps? Like data comes in from an API in one form. You like you parse it, you package into a different form. Hey, you store it into database and it's a different form. It comes out again into a different form. Then it's like HTML or whatever. And you type in something it's a different form again. And all you do is like you're massaging data in different forms throughout your app.

42:30 This is what most apps are. We are pretty chasing printers. And the and the the really the hard part is solved by Postgress 30 years ago by some neck birds. Uh that's that's really what a lot of software is like. There's always some interesting parts, but I don't have to care how this button is aligned or which tailwind class is used or or like many details are boring and many other

Orchestrating multiple AI models simultaneously

42:52 details are interesting. But it I think it's much more about system architecture than having to read every single line.

42:59 Right now jumping forward, what is your workflow like? Like like when you're working on on cloud bot, are you using a terminal, multiple terminals, which which tools and and how are you you know like you said you're not you're kind of like not reviewing the the code, but you're still thinking about architecture. Like what does your average day look like in terms of tooling? You know, you have to explain to a developer who might join the team. you know, at at one point you didn't get like what does it look like?

43:24 It's interesting. Let's let's let's go a little bit. We were we were in in in April with Cloud Code and then I got really hooked and then I did some I had a phase where I did cursor and then I did I I used Gemini 2.5 a bit. Then we had this phase with Opus 4. I hooked up a lot of my friends like I know I know both Armen and and Mario from Vinner.

43:49 They got they got AI pill because I I was addictive. You know, my my enem was like confusing them and then they tried it out and then and then eventually they also were up at 5:00 a.m. and I called it like the black eye club. I mean, there's a reason like I I I I started a meet up in London that I called called uh Cloud Code Anonymous because because it's it's a little bit like a drug because it's so it's so much fun. Like to to me what what what blew my mind so much was this realization that I can build everything now. Before you had to really pick which side project you build because software is hard. Yeah, it's still hard. But now like I I am this this friction that I talked about where I'm so good at this at this technology and I'm like so bad at this and I'm like oh let's make the CLI in Go. I have no clue about Go. But I have I have a good system understanding and once you have that is like you you develop a feeling

44:50 what's right what's wrong like it's it it is a skill in itself. I remember there was this tweet where someone said, "Oh, when you write the code, you you feel the friction and that's why that's how you make good architecture." I feel the same friction when I prompt because I I see the code flying by. I see how long it takes. I see if like the agent pushes back. Um I see if what it creates looks like messy or like makes sense.

45:15 When I prompt, I I have a hint already how long it's going to take. If it takes much longer, I understand that I messed up somewhere.

45:21 You kind of feel the model. you you know you know

45:24 yeah usually it's like this or if it runs

45:26 I I feel it's very much a symbiosis like I I learn to to talk

45:32 may I even say dare or that language more so it's like my my knowledge how to use those things improved and also the models improved and then and then like over the time between between April and now I would say yeah the inflection point was summer where it just got so good that you could you create software without actually writing code by hand. But the real that change that like sold it for me is was again GBD 5.2 that was again I think it's underrated. I don't know I don't know why why all these people still use cloud code. I I I kind of get it. It's it's a different way of working but whatever OpenAI cook there is insanely

8 Highlight · Using multiple AI agents in parallel

46:18 good. pretty much every prompt I type gives me the result I want which is insane like like on on on cloudbot the my latest product I use between five and 10 agents in parallel if you're very much cloud code build you have to forget quite a lot of the the silliness that the things that you have to do to create good output with cloud code I mean I also met that team and and they created a whole new category. Like cloud code is is is a category defining product and it is amazing for general purpose computer work and it is is really good for coding and I I I I still use it almost every day. But for writing code in complex applications, Codex is just so much better because it it it takes 10 times longer. um Claude would read three files and then be confident enough to just like create code and then

47:20 you really have to steer it and push it so it reads more code so it gets it sees a bigger picture of your codebase so that it it it weaves in new features better and Codex will just like be silent and just read files for 10

47:34 minutes and if you if you only work on one terminal I completely understand how you how you find this unbearable But I rather have something where it's also you don't tell it what to do. You know this is this is also something that people don't get like I have a conversation with the model. It's like oh let's look at this what what what what options do we have for this structure? Did you consider this feature? It's like because every every session is like the model starts from having no understanding about your product and you have and and sometimes you have to just give it a little bit of pointers. What about this and this? So it explores different directions and you don't need plan mode like I'm just having a conversation until I say build this it will not build this. There's some trigger words because it it is they all are a little trigger hungry but as soon as I say let's discuss or give me options they will not build things until I say build.

48:29 So so a lot of a lot would you say a lot of your prompting or a good part of it is this conversation where you are pretty much planning together with the agent.

48:38 Yeah. It's like what about like I say okay then you remind them it's like okay we need documentation what would be a good spot it would like give me some recommendations I say no this should really be its own page do we need a configuration how how does this fit into this other feature it's like I am designing the system because I have this I have this system understanding about how how is my my product how are the shapes looking I don't have a line by line code understanding that's that's what Codex does for me but I'm the architect you It sounds a little bit like you're almost, you know, for years back this this totally came got out of style. But there was this idea that you would have the architect with a capital A who used to be a software developer, but they're not hands-on anymore because they spend a lot of time understanding the business and they have these developers working underneath them. And some companies still kind of work a little bit like this, but most modern companies don't. But some banks, etc. I met people there who are capital

49:38 architects. They do the system plan. They talk with fellow architects. They have the blueprint and then they literally pass it down to the team and everyone hates this model obviously because you know again like I I think as people you kind of want more. The architect is never on call for for this stuff and so it just kind of breaks down in practice and a lot of large companies just move to the staff engineer model where you're kind of all working together. Of course, there's people who make who might have more input, but sounds like it's almost like this world where you are the architect who kind of, you know, you have your little agents who who do the code, except in this case, you are of course fully responsible because you're still an individual contributor. You're not you're not like a okay, you might say you're a manager of agents or whatnot, but the code is is yours. It's your responsibility. You're going to be on call. If you know, if you push out code that takes down CloudBot, which it did just recently, you're on the hook for it, right? like and I think the the difference in this system when when it was in companies it was the architect was kind of shielded from the output of their work because there's so much

50:39 people and so much process etc. Well, I wouldn't say architecture. I I like the word builder.

50:44 Builder. Yeah.

50:45 And and and I think also that's there's a few categories that I see for people that are highly successful using using AI and people who really struggle.

50:55 I care more about the outcome, the product. I very much care about how it feels and everything, but how the plumbing works underneath.

51:04 I care structurally, but you know, not to the not the biggest detail. And then there are people who really love to to code on hard problems like think about algorithms don't really like the I'm building a product with like all the marketing all the they they more like they like to solve hard problems and those are the people who really struggle and and and often reject AI or get really sad because that's exactly the job where that AI does like it solves the hard problems. Now sometimes I give it some pointers but many times I learned I learned more this year than last five years around around software architecture and designing. I the there's so much inside those monsters um on knowledge and everything is just a question away but you have to know what question to ask. Of course I also built this Twitter thing and it's still not done and I and I I really hope I I'll get back to it at one one time.

51:57 everything worked but if I used it more at some point things got really laggy and weird and then it worked again and I just couldn't figure it out and it was like really difficult to debug because it was not easy to reproduce. It was just like you use it more and things get really slow. I basically had like software in in in in Psql like in Postgress that would be triggered when certain inserts were were doing and then the database would would get really busy and the model couldn't see it because it was it was it was so far abstracted from all the you know like those those models are really good at tracing through but this was a side effect that was so hard to see because it was only in this one file a function that had no connection to anything else. um with a name that was not easy grabbable. I just never asked the right question until I was like do we have any side effects for this and this and I found it and I fixed it and it's like but I everything is just the right question away.

52:53 Yeah. But you you need to have like knowledge expertise.

52:56 Yeah. You experience

52:58 you you I mean so so so these are the people who rejected and then the people who who care a bit less about how it's being plumbed internally but are just excited to build things. They're really successful. And one thing that also helped me is, you know, when you run a company and then you hire people, you can't breathe on everyone's neck and like make them have the line of code exactly that way. And there's a lot of people who who didn't manage a team, they didn't had this experience how to how to relax a little bit and understand that yes, this maybe is not exactly that code that I want, but it will get me closer to my goal. And for anything that is like not perfect, we can always make it better and like put more time into it. I very much believe into this iterative improvement. I had to learn to let go a little bit at my company. So, so, so then when I when I had cloud code, it kind of felt like I have like I have like imperfect, sometimes silly, but sometimes very brilliant engineers that I have to steer and where we where we

53:59 work together on a common goal. It felt a lot like being the boss again.

54:04 Yeah. And and interesting now you know you you built kind of software I guess the traditional way you know pre AI for 15 years or even more than 15 years and you got really good at being also leading a team and then how to have high standards. You really cared about the the craft there as well. You've now kind of been I guess vibe coding or working with agents for a year. You're comparing

Closing the feedback loop is essential for agent success

54:25 the two. What do you think? What do you think really really changed? And what do you think are things that kind of stayed the same despite all

54:32 First of all, I don't like I don't like the term VIP code.

54:34 All right. How should we call it?

54:35 I I think I think I think VIP coding is by now almost a I I call it I tell people I do what I do is agending engineering with a little star. VIP coding starts at 3:00 a.m.

54:46 Now like because all the the mundane stuff of writing code is automated away, I can move so much faster. But also means like I have to sing so much more. I'm still very much in the flow. Like it is it is completely the same feeling as for me as as I I very much get in this

9 Highlight · Agentic engineering as multitasking

55:03 flow state but it is mentally even more taxing because I have I don't have one employee that I manage. I have like five or 10 that all work on things and I switch from this one part to this other part to this other part to this other part. mostly because of I'm designing this new subsystem or like this feature and then I know that it will probably take Codex like 40 minutes or or one hour to build. So like I want to like have the plan right and then I build it and then I'll I'll move on to something else but then this is cooking and then I work on this and then this is cooking and then this is cooking and then at some point this is cooking and then this is cooking and then I go back to this one. So like I I I I switch around a lot in my head. I wish I wouldn't have to do that. Like I'm sure this is a transitionary problem and at some point we have we have models and and systems that are so fast that that I can paralyze a little less. But to stay in the float flow state I need to massively parallelize. So that that's that's how it work. I go back to there and and maybe tweak it a little bit more. But

56:03 usually just like try it out and maybe then this is ready because this only took like 20 minutes. So like I constantly jump around. Usually there's there's one main project that has my focus and I have like some satellite projects that also need attention but where I can make maybe I spend 5 minutes it does something for half an hour and and and I try it and it doesn't need uh so much capacity up there.

56:27 This almost sounds you know like two things come to mind. One is there's these like games where you have to manage a kitchen with the employee and and you see like the recipes or

56:36 something come out and you need to jump and do it again.

56:38 It's like Starcraft, you know, you have like your main base and you have like your side bases. They give you resources.

56:43 That as well. And also one thing that just came to mind as you said like I go there and I watch this and I make a decision is when I see the chess grand masters play multiple boards at once. You see see sometimes they 20 boards and they always all you they go there they kind of you can see that they just like see what's on that board. they make a decision and for some boards they stop for longer I guess better players or better opponents. It feels, you know, both they're occupying 100% of of their brain. You're occupying your bay and you're you're kind of scaling yourself as long as you can context switch. The difference the difference was up until with with cloud code I you have to work a little different because it is much faster but then the output often doesn't work on the first try. So like it makes something but then it forgot to update three other things. it crashes or you give it the good thing how to be effective with coding agent is always like you have to close the loop. It needs to be able to debug and test itself. That's the big secret. Um that's also something I I

57:44 think that's part of why it got so much more effective. Um but yeah, with with with clock code you I often had to go back and like fix up the stuff um or it just takes a lot of iterations. So in the end it's not that much faster. It's just more interactive. And and these days with Cordex it just almost always gets it right. My my general strategy is always I I build a

10 Highlight · Closing the validation loop for coding agents

58:08 feature of course you and and of course you always let it write tests and you make sure that it runs it.

58:15 It runs them. Yes. So even even when I write a MAC app, I don't know like I I just yesterday I debuged this feature where the MAC app couldn't find a a remote gateway but like the the same code in Typescript could but makeup is kind of annoying to debug because like it builds it you have to start it you have to look at it you have to like say no this is not working. So now I just said it like you know you're going to build a CLI just for debugging that invokes all the same code path that you can call yourself and then you just iterate and you fix it yourself and then it will just cook and it just cooked for an hour and it was done and it told me like there was a race condition here and here and like a misconfiguration blah blah blah and like yeah it sounds sensible. I don't need to I don't need to see that code. It's like but but you don't need to see it because you set up the validation loops and you trust that because it ran it. I mean this is I guess I guess it's not too dissimilar to like sometimes when you work on a large project in a large company when all the tests pass I mean it doesn't mean 100% it's there but it's it's a pretty good

59:16 and and all the new new code has test as well you know someone thought about it and tested it and and all that. So even even on my on the very latest project we always had bugs but like anti-gravity has like a certain a certain weirdness with how it takes tool calls in the loop in the in the format. So you have to do like some filtering.

59:36 Yeah.

59:37 And that broke a bunch. And it actually took me way too long to realize like what am I doing here? I just need to automate this. So I was just going to codeex. like design life tests that spin up a Docker container, install the whole thing, spin up a loop, use my API keys from this and this file and then you tell the model to read an image, create an image before and then look into the image and see what it sees. So I I don't not just tell the loop, I still tell tool calling, make it work and then it solved itself. It took forever, but it it it it tested all my API keys like from from Entropic over SEI over GLM like everything and it fixed all those little indicies where where sometimes the tool calling didn't work or the ordering was wrong because I closed the loop and and that's that's

1:00:28 and closing the loop you mean just have a way to to have have the agent be able to validate its work?

1:00:33 Yeah, that's why that's the whole reason why why those models that we currently have are so good at coding. But like sometimes mediocre good at writing creative because there's no easy way to validate right but code I can compile I can lint I can execute I can verify the output if you design it the right way you have a perfect loop like even now even now for for websites I built the core in a way that can be run via a CLI so it's like I have this I have this perfect uh execution loop because the the browser loop is insanely slow you want something that that loops fast. So it sounds like one thing that is not really changing from like before is we had this before like backend or business logic heavy he heavy thing could easily

Verifiability forces cleaner software architecture

1:01:17 be or more easily be verified that it's correct.

1:01:20 Surprise actually using aentic coding makes you a better coder because you have to have to think harder about your your your architecture so that it's easier verifiable because verifying is the way how to make things good. Well, then remember back back even before AI for complex systems like once you got someone who built these things before what they started with how do we make it testable right like you you need to design interfaces classes testable you need to think about like am I going to fake things will I use mocks will I use end toend testing which will be long etc but these are like really hard architectural decisions and once you make them they're I guess harder to change in your in your word you know like the model would cook a lot longer if you asked it to make a massive refactor acture and you know if you have test it'll get it right but you know now these we we still have these trade-offs

1:02:07 yeah it's still it's still software it's it's I I would say I write better code now that I don't write code myself anymore and I wrote really good code but but like even back at the company sometimes testing was so tedious and you come up with all those edge cases and and and the branching

1:02:25 I mean outside of Ken Debbec who I deeply respect and he was on the podcast and we We we we talk like he he still writes test first and he tells me that he's not mad at me for not writing it but if you want to write like you know poor quality code it's on you. Uh but I don't know many developers myself included I never liked writing tests and even even when I pretended that I did I I just never did. It's a little bit like writing documentation and writing tests to me it was never a creative expression. It is so good now like I I would say for my last project I have really good documentation and I didn't write a single line myself like no I don't write the test I don't write documentation I explain the model the trade-off so like why we did something like this and then tell it like like write that write the entrance section beginner friendly and then add more technical detail at the end and it is so good I never had a project with that good documentation just by every time I design a feature this is a part of this a part of the process And also like testing I was like okay we built this

1:03:27 how are we going to test this? Yeah we could do this and this and this. What if we build it this way and oh yeah then we can test it better. So it's like this is now part of my singing because I I always think like how do I close the loop? How do I the model always needs to be able to verify the work itself um which automatically steers me to better architecture. So why do you think there's, you know, a bunch of like

Dismissing AI coding from single prompts is a skill issue

1:03:50 experienced devs who are still pushing quite a bit back on on just like the idea that AI can do a lot of this?

1:03:57 That was a week ago. I I stumbled over a a blog post by Nala Coco with love that I deeply respect and learned a lot from. And this blog post was just was a dissing of the current way how models work. And and what he did was he he tested like five or six models including some that make no sense like the the Open EI 120 billion open- source one that is not good enough to write good code you know it's like and he just he he wrote a prompt as far as I understand it there was there was not a lot of information on the website but to me it sounded like he wrote a prompt he put it on claude web and and he pressed send and And then he took the output and ran it and it didn't compile and he was disappointed. But he's like, "Of course it will not work. Do you think I can write buck-free code on the first attempt?" And those little those

1:04:58 models are ghosts of our collective human knowledge. They work very similar in many ways. Of course, you don't get it right the first time. Like there will be mistakes. That's why you have to close the feedback loop. And also you don't just send a prompt to the model, you start a conversation. Hey, this is what I want to build. It's like you he complained that it used old API. Yeah, you didn't specify the Mac OS version.

1:05:23 So it so so it made an assumption to default to like old API because that information was missing and it is it is trained on a lot of data, not just the last two years and there's just more old data than new data. So this is like the more you understand how those little beasts think, the the better you get at prompting. Um and then and then and then he he spent maybe I don't know a day or so on on playing with it and then just decided that this technology is still not really good. But to be effective you have to spend significantly more time.

1:05:59 You know, it's it's like it's like you know how to play guitar and I put you on the piano and you tried a bit. It's like, "Oh, this sucks. I go back to my guitar." No. No. It's like it's it's a different way of building. It's a different way of thinking. You have no idea how often I screamed at like 3:00 a.m. to cloud code because it did something silly. I slowly started to understand why those things do what they do with like exactly the way I tell it to do things. And sometimes you can literally ask you can even even last year like I for this project I the last project like clotboard I feel like a human merge button because the community is like blowing off and all I do is like reviewing PRs. I I I I have very little time to actually write code myself anymore.

1:06:42 And in the beginning it would often like just cherrypick things and would close the PR and I was like so annoyed. So was I'm like why are you doing this? Yeah. When you say this and this I interpret this this and this. It was like ah like I I learned the language of the machine a little bit more. I tweaked my my prompting and now I get exactly what I want because it's it's a skill like any other skill.

1:07:06 Yeah. And this is like Simon Wilson has been saying the same thing even though he's been using it for for years and I think everyone I think once I start to use it I also realize like I'm not I'm okay at it but I I I I could do better. What if we put this to a real test? cuz I think it's fair to say that right now

Complex multi-agent orchestration layers are waterfall traps

1:07:21 you're building CloudBot which is a you know it's not something that generates revenue there's a lot of users and it's blowing up and it's it's a really cool tool but it's not PSPDF kit which is a business that it's a lot of revenue is hinging over it if today you know we just wiped PSPF kit does not exist you need to rebuild PSPDF kit you now have these agents how differently would it look how much would you trust it what would you delegate what you would you validate and and when you know you built up a team around it because now it's a profitable business at the very least you need to hire sales people whatot how do you think the team would look different today with that same product cuz you you know exactly what it took to build it and you also know what these tools can do today

1:08:02 I could easily run a company with 30% of the people it would probably be quite difficult to find people on that level but you you you want you want to have really senior engineers that really understand what they build but that are also comfortable in in in delegating and know which parts are actually important to to work on and which parts I can vibe. That's still something I don't see. I don't see a lot like like especially in the AI world, there is so much crap on Twitter, there's there's so many people that are loud but clearly have no clue what they're doing. There's there's so many there's so many dumb concepts around like I'm sorry, but the Ralph Wigum one gh like this is again another another silliness people use to work around uh model limitations of of of of Opus

1:09:05 that you don't even need when you use Codex. There's there's maybe a few cases where you have a really long list of individual tasks that can be automated,

11 Highlight · Why elaborate multi-agent orchestration layers fail

1:09:13 but that's usually not how software building works. So there's these people who I see so many people building up this elaborated orchestration layers and then you have like beats that automatically creates tickets and then your agent does tickets and then your agent emails the other agent and then you build up this this elaborate mess. What for? Oh yeah, they did they design the the spec for like a few hours and then you just like the machine builds it in the whole day. I don't I don't believe this works. Like like this is this is this is the waterfall model of software building. This we learned long ago that this doesn't work. Like yes, people work differently and maybe it does work for some. I just I just I don't see how this how this could work for me. Like I I have to start with an idea and often I purposefully underprompt the agent so it would do something that would give me new ideas. You like maybe like 80% of the things I assumed

1:10:13 were like crap but like there were like two things like oh I didn't think about that way.

1:10:18 Mhm.

1:10:19 And then I I I iterate and and and shape the project and I have to I have to click it. I have to like I have to feel it. I feel I feel to make good software I you know one thing those those things often lack is taste. I have to feel like how does this feature feel and and the beauty now is that features are so easy I can just like throw it away or like reprompt it. My building model is usually very much forward. It's very rarely that I actually revert and have to go back. It's just like okay no then let's change this. No let's do this.

1:10:51 It's like it's like shaping. I I love how this like you start with a rock and then you like sisle away at it and like pick different areas and and and then slowly like this statue emerges out of out of marble. That's that's how I see

1:11:05 that's how I see building something. I I guess reflecting on how software engine is changing. This seems like a change because before before we had AI or any of these agents, upfront planning did make a difference. you know, writing at PSPDF could you insisted I think to have a proposal where people put a lot of thought up front to specify and do all because it was expensive to I guess to to build to to do you think this is changing because of the the cost of just writing code is is going down or

1:11:36 I mean I still I still I still plan and I but

1:11:40 you still do yes

1:11:41 but I don't put as much into it because I it's now so much easier to just like try and look at the results and then see if oh yeah this this shape could work or no we have to like the tweaking and even even like oh no we have to like do it a completely different way isn't so much cheaper that it's to me it became much more playful.

1:12:00 Yeah. Yeah. I guess cuz cuz like you know when you're working even if you have like a new grad on the team or an intern you know you give them something they work it for a day or two now you give them another it's another day or two you know and and we're not talking days here we're talking minutes or like if it's a longunning task like 10 20 minutes at at worst. Plus, you're not just waiting on that thing. You have parallel things running. So, it's not that much of a waste, if you will,

1:12:22 in in in in in cloud. At the beginning I had this assumption of like one agent and then eventually changed to multiple agents and there was the assumption of like one provider like WhatsApp and now it's multiple ones and changing that was like such a pain would would have been such a pain if I would have written it myself because you have to weave in literally everything through the whole logic of the application and yeah it took Codex like three hours it would have taken me like two weeks you know so so that upfront planning. I I could have realized that in the beginning, but now I I know that like I can just change things and it's it's much it's much easier to work down your technical depth or your you know, you evolve how you think about a project as you build a project. That's why I don't believe in I don't know things like Gas Town where like you write up the spec and then it builds itself and then it's done. How can you even know what you want to build before you built it? you learn so much in the process of building it that will go back into your thinking of how this how the how the system

1:13:24 actually will end up being to me this is very much it is very much a circle um until I you don't I don't you don't walk up the mountain like this you go you go around and sometimes you like you you you stray off a little bit of path but but eventually you you you reach the top that's that's how I feel so

1:13:42 then you know you've been building cloud bot for what like two months three months non-stop upish or or like like how long?

1:13:50 Let me let me let's switch a little bit

Building a hyper-personal AI assistant with proactive triggers

1:13:51 gear. So, one of the ideas that got me back was even even in in April May was I I wanted to have this hyper personal assistant and not like not like one that sends you a good morning email. Oh, these are your three tasks. No, one that has a really deep understanding of me and doesn't just I don't know I meet a friend and then and then when I go home it would ping me, hey, how was how was that meeting? Or one that would wake me up one day and say, "Hey, you haven't texted Thomas in 3 weeks and I noticed he's he's in town right now because I checked his Instagram account.

1:14:36 Do you want to say hi?" or something that says, "Hey, I noticed every time you meet that and that person, you're sad. Why is that?" Like something something that is deeply personal. Um like almost the the anti-orem. It's kind of like the movie Her, but but that's where the technology is going. Those models are really good at understanding text. the the the bigger the context is, the more patterns they see. And even though they're like matrix calculation without a soul, it very often feels different. So this was like one of these ideas and I even created a company I called a mant machina like the loving machine. But in summer when I explored it, the models weren't quite there yet.

1:15:21 I got some results that it was like okay this is like I'm a little too much on the edge of what I need right now which it was very exciting because I know that that the the state of AI goes so fast that oh I can just revisit that in like a little later and and one of the ideas also was is that I assume that all of the big corporations right now are very much working on personal assistance in the future. Yeah,

1:15:47 everyone will have you will have your best friend who is a a freaking machine that will understand you, that will know everything from you, that will can do tasks for you, that will be proactive, that will require a lot of tokens, but everyone who can afford it will have one. And of course, this will democ democratize and and trickle down to like more and more people as as we learn how to build more efficient systems and and and hook up on on on chips. No question this is where the things are going. You see like the first things with like OpenAI who who launched pulse with some productivity but we just don't have enough compute yet to offer this as a feature and also it's it's quite difficult. My idea always was like ah I kind of want something that runs on my computer and where the data is

1:16:35 it's yours

1:16:36 is is actually mine and not and it it's also quite scary that like you you give openropic access to your email your calendar your your dating apps I don't know if you talk to to to your normie friends but a lot of my friends in they use that a lot to basically have a a therapist And it it does work incredibly well. Like it's it's a really great listener.

1:17:02 It understands your problems and unless like some of versions of 40 that are like sure this is a great idea. I want to put French fries into a salad. It it works really well. And I did that too like to like re I mean part of it just is like the the act of reflecting already is helping you. So it would even work if the machine would only repeat exactly what you wrote to a degree. But it actually gives insightful questions. It's actually it it got really good. So I had this idea of this like assistant but the tech wasn't there. So I did other part and I I built a whole bunch of fun stuff with like of course like I built VIP tunnel this in your career to become like an authentic engineer you have this phase. It's a trap phase where you you're looping and building your own tools to like optimizing your own workflow. But this idea of like this hyper personal agent stuck a little bit and then over the last few months I I really

1:18:02 started I built it but finally initially I didn't even had the the scope that it has now. Like I called it WhatsApp relay. I just I just I just wanted to do to trigger stuff on my computer with WhatsApp. So, I built like a WhatsApp relay where I had an agent that could do stuff with my computer and then I I was traveling to Morocco for a friend's birthday and was out most of the day and just used WhatsApp to to talk to my agent and I was kind of hooked. It it it was guiding me through the city. It was making jokes. It could text other friends via WhatsApp from me.

1:18:45 And I remember I I was blown away because I in the beginning the tech was very scrappy but I I built in something where I could send it an image. Didn't even use the proper thing to send an image. I just gave the the a string and it could do the read tool to like read the string. And then I was in Morocco and was just like just like not seeing it and saying it a voice a voice message but it didn't build that. And then like like 30 seconds later it replied to my voice message. I'm like, "How the did you do that?" Oh, yeah. You sent me a file and and then I looked at the header and I found that it's OG. So, I used FFmpeg to convert it. And then I I looked for Visper on your computer, but it's not installed. But I found the OpenAI key. So, I did a curl to open server, let it translate, and I'm like, "Holy cow." like this was Opus 4.5 and it's so incredibly resourceful like you just did this you know other people say oh you need a skill or some system no

1:19:46 just like it just figured it out I slowly got hooked on the thing I I used it I used it to wake me up and it was running I it was running on my Mac Studio in London and was connecting over SSH to my MacBook in in Morocco and was turning on the music and making it louder louder is because I didn't reply. And to make that work, I I added a heartbeat. So, which which in a way is insane from a security perspective. You have a model that you prompt with do something cool and surprise me that you send every few minutes to make it proactive and like go through your task list. Uh like probably the most expensive alarm clock ever. But it was just hilarious. And also the text it sends like cuz I I I was I had a balloon fart and it it knew that I had to wake up very early and I didn't reply and it was like you could see the reasoning Peter's not responding but Peter has to wake up. No, no, no, no, no, no sleep.

1:20:45 Like I it was bitching to me and then I I showed it to the the the friends I was with and everybody was like hooked. Like this is something magical and I was hooked too. Um, and then I I I went on Twitter and I got the most muted responses cuz nobody would get it. I feel it's somewhat of a of a new category of products. Um, that a little bit like your story with like you know when you didn't get the iPhone uh from

CLIs beat MCPs by enabling data filtering and piping

1:21:14 the marketing campaigns on TV and anywhere and then you had to use it.

1:21:18 Yeah. So I I worked on it but only the last two months and it the name changed from V relay to at some point a claude uh said like then what is this name like it doesn't fit the feature set anymore because like I had like in there and other features so I renamed it to to Claudius because it's an inside joke because I like Doctor Who. I felt cloudbot is is a better name has a better domain and explained the product better. So I did on all the domains and then I I also quietly built up my army because to make this work you want you

12 Highlight · Why CLIs beat MCPs for AI agents

1:21:53 want everything to be a CLI. So I was just building CLI for everything like for Google for my bed for lambs for music.

1:22:02 Why CLIs? Why not why not MCPs? And what do you think about MCPS anyway?

1:22:06 As a crutch it's it's I think that the the best thing that came out of MCPS is that it made companies reync to open up more APIs.

1:22:14 Mh. But the whole concept is is silly. You you have to pre-export all the functions of all the tools and all the explanations when your session loads and then the model has to send a precise blob of JSON there and gets JSON back. But surprise, models are really good at using bash. And like imagine imagine you have a better service. So the model could ask for list of available cities and then get like 500 cities back and then it has to pick one city out of 500 city. But it cannot filter that list because that's not part of how MCP works. And then you say, "Okay, give me the weather for London." And you would get like the weather temperature wind rain and like 50 other things that I'm not interested in because I just want to know is it raining or not? Probably raining because London. But the model needs to digest everything and then you have like so much crap in your context.

1:23:13 Whereas if it's a CLI, I could use just it could use GQ and you could filter for exactly what it needs. But does does not seem like a limitation that everything is loaded around the MCP in the context. That seems a problem. Like it sounds like it could work if MCPS were not in the context and there was a way to discover or decide which one to use.

1:23:33 That's what that's what companies are building now. But there's still the problem of that I cannot chain them. I cannot I cannot easily build a script that says, "Hey, get me get me like all the

1:23:43 all the CDs that are over 25 degrees and then and then filter out only that part of information and like pack it in one command." That's it's all individual MCP calls. I cannot I cannot script it.

1:23:54 Yeah. But but I guess this is just a matter of time because if we think about like you know when when I'm building a weather app right now, I know that you know even without AI, I know I need to build up this thing. I need to I needs to fetch the data. So I will search what kind of APIs are available, which one do I like, which what kind of trade-offs for pricing, for covering, etc. And then I choose that API and I could chain APIs because I I could get that result and look up a etc. So I I guess this is, you

1:24:20 know, like it it sounds pretty much we've solved this. So as free AI, we're going to solve it. It'll just take some time and who knows what the format for it will be. I mean I mean I built make porter which is a which is a a small TypeScript thing that converts an MCP to a CLI. So so you can you can just package it up.

1:24:39 Basically you're saying CLIs right now are a lot more efficient.

1:24:42 Yeah. Yeah. So so my in in in cloud but I don't have MCP support but you can we make portal you can use any MCP you can you can literally be on your phone and say hey use the use the versel MCP to do this and this and it will go on the website. It'll find the MCP. it will load it and it'll use it all all on demand. Even right now if you use MCP you have to restart cloud code which is like very user unfriendly. So I quietly built up my army to like automate everything which was a lot of work. Uh I think t did a video a few days ago where he told me like this guy is insane

Giving an AI agent full system access

1:25:18 because the list is really long by now but like I I as I was playing with my my my agent I just I want him to do more and more stuff you know. I felt it really hard to convey what it does. It's still hard to me. In January, January 1st, just a week now, I did. Okay, let's let's try something. Let's let's do the ins really insane thing of like making a Discord and then adding my agent to Discord. There was somebody who contributed Discord support to it. And uh even though I wasn't sure if I should merge it and I eventually did. So I put on my agent who has full read write access to my computer in a public Discord.

1:26:00 What could possibly go wrong?

1:26:02 Yeah. It's like this is absolutely insane. And then of course like some people join the Discord and then they saw me they saw me using the full power of this thing like checking my cameras doing home automation. It playing DJ for me. Like I was in the kitchen and I told him like look at my screen and are my agents done cuz it has full access of my clean and it can click. So it can actually click into the terminal and type for me and like it can tell me your codex say this and this because it just sees the screen. I mean I'm working on optimizing that like I I actually want to stream out it because would be much would be much better if it's text but it works already like it it's it's in the background it look at my screen and like make some rants if I do some [ __ ] And everybody who experienced it for a few minutes got hooked. Like this was this was the craziest blow up from 100 stars to like what 3,300 stars in a week. Um and I think I merged 500 pull requests already. That's why I feel like I even merge button. So, so that that's why

1:27:02 that's why I'm I'm a little I'm a little all over the place these days because this project is blowing off and and and you know the beauty of it is the technology disappears. You just you just talk to a a friend on your phone that is infinitely resourceful, has access to your your email, your calendar, your files, can build websites for you, can like do administrative work, can scrape websites, can call your friends or can call a business. I'm just about to to to merge the call feature. It literally can call a business and like make make a reservation for you and you don't have to think about compactation or or any all of that context blends away. I have like a I have a memory system that will remember um not perfect. Nothing's perfect yet, but it's already feels magical. cuz all cuz cuz now I I walk around, I see like this event, I I send Claude a picture, it will it will not

1:28:03 only tell me the reviews of this event, if there's a conflict in my calendar, if like friends talked about it or, you know, it has so much context that it the responses that it can give me are like so much better than like what any of the the current tools that live in their own little box can give me. Well, sounds like you built whatever Apple was hoping Siri to do, but they've been unable to.

1:28:26 I honestly I built the best marketing tool for Entropic to sell them more subscription. I don't know how many people signed up for the $200 subscription because of Cloudbot and like many people already had one and used a second subscription because of that because it's so token hungry. It's not is not that it's token hunter. It's just that people love it so much that they use it all the time. And because the technology blends away, they don't see that it spawns sub agents and does like a whole bunch of things in the background to just make it feel easy.

1:28:58 But like there's some actual engineering like there's a lot of work in the back uh to make it feel easy. You know, this is like the hard part. Like you hide complexity to a degree that it it it feels magical. Well, but yeah, but this is interesting because like I I I can sense from we're talking, you know, you put so much thought into architecting this thing and right now like you've been building this for a few months and yes, it blew up, but in your head like do you have a structure of how CloudBot is structured like what parts you need to modify, you know, like like you kind of you can get your your mindset into it and you you know where modification needs to do. You know what you want to refactor because it's not going to be efficient. Are you thinking about like things like memory

Architecting simple text-shuffling and self-updating bots

1:29:41 consumption, token consumption, efficiency, those kind of things?

1:29:45 I mean, token consumption is more like how do how do you structure the prompt and memory? It's it's TypeScript that shows Jason around in the end. Let's be honest. Like like like I get text from from an LLM,

1:30:00 I save text to to disk. I send text to WhatsApp or to now we have like MS Teams, Slack, Discord, uh Signal, iMessage, WhatsApp, and there's there's two more that are landing like Metrics that will will expand this thing even further. It's like it's it's really poly by now. But but mostly I I again I I move around text in different shapes and maybe maybe it goes to different providers or there's like now it's different agents and there's like the agentic loop and there's like a lot of configuration and there's it's it's a lot of plumbing but nothing's there's nothing in there that is really difficult. Yeah. Well, but it's it's a lot lot of small things, right? Like I I feel in software, right? Like we we know for software even before AI there was not much difficult. Of course, you need to learn and understand the language and all that, but

1:30:53 the difficulty is how do I how do I make it so that it feels magical. So, so what I worked on a lot is now you have you have this oneliner that you type in that you python your command. I will I will check if you have node installed homebrew installed. I'll I'll install the mpm package. I do some check if you have any existing stuff just to like y

1:31:12 just to make it work simple even if you already used an older version and everything. and then I I'll guide you through uh setting up a model. But again, I will I will predict or or claude installed. So you can just press enter. So you don't have have to think about it. Mostly just press enter and then you want a WhatsApp, you type in your number, it will just work again.

1:31:38 And then and then I'll ask you do you want do you want to hatch your bot? And you can press yes. And then and then like a TUI a TUI comes up because you're still in the terminal, right? You want a good experience.

1:31:49 Yeah. So just a toy basically for that and where you where you to see wake up my friend and then the the I programmed the model I added a bootstrap file and the to explain the model that it is now being born to like create an identity and a in a in a in a soul where like the values of the user are in and then the model will be like hello like stretches who are you um who are who I am what's my name you know and this this is is like I've watched people do it and that's where the magic starts. That's where that's where they're like they no longer think about I'm talking to to GPD4.2. No, I'm now talking to my friend created Vajorn like a a unicorn with part of his name or like I'm talking to Claude. Uh and then it's like what what's important to you? What do you do?

1:32:40 It's like curious. I I programmed it to be like curious and then go through this bootstrapping phase and then it will actually delete the bootstrap file and create a a user.md with like information about you a soul.md with like all the core values and an identity with the like what's his name what's his core emoji what are the things that are like inside jokes and and but it's like evolving documents that it will maintain and like tweak as you interact with it and then you it will just like send you a message on on WhatsApp and you just like suddenly you talk on WhatsApp like making this flow easy. That was hard.

1:33:13 Yeah. Also like even even get coming up with the idea of you know you you're not you're not editing the configuration because the agent can edit its own configuration. You don't have to update anything because the agent can update itself. You can literally ask your bot update yourself and it will fetch itself and update itself and come back like hey I have new features. Planning the technical giveaway so far that's the magic. That's why that's why I

1:33:39 But it feels it's very similar to what you would with PSP PDF cut, right? You kind of blended away the complexity of a PDF. So it was just there. You could rotate, you could do.

1:33:48 Yeah. Yeah. Even at the API level back then,

1:33:51 but it's a it's a bit bizarre like what what you described reminds me of this Black Mirror episode I just watched, which is called Play Thing. uh where it's a it's a digital little uh creature that creates of course it's black mirror so it has a black

1:34:06 bit bit of a dark ending but but it it had it was also a game. It also kind of feels you know we talked about how you you don't play as much games cuz you like but this also feels a little bit like a game, right? But it it's it's kind of like more connected with reality. Just fascinating how we're we're we're here pulling back into the the realm of of software engineering. So you built this this product and it's now it's a production software you're merging porocas people are using it now thinking back to PSPDF kit and and companies that are like that which which have you know like like tens or hundreds of developers working on on production software knowing what you know with how you're building cloudbot and the tools that you're using how do you think software engineering at those larger companies could change because one one thing I see is is for individual people like you it's like AI I is really really hitting a fit like you're making you way more productive. You're in control at teams or at companies that are you know have existing code. It's just a lot kind of slower. It's it's not really okay people use it for this or that but but

1:35:08 it it seems a huge divide between the two worlds. And you've kind of been you know CEO for this company. What what might that be or is it just more of a timing thing where every new technology often comes with with hobists you know

Redefining corporate software engineering for AI

1:35:18 pick it up uh earlier? I think companies will have a really hard time adopting AI efficiently because this also requires to completely redefine how the company works. You know, you know, you know like you know like at Google they they tell you you can either be an engineer or like a manager but or you want to also like define how the UI looks. That role doesn't exist because either you like you you build it or you you design it.

1:35:44 But this new world needs people that that have a a product vision that that that can be able to do everything and you need like far fewer of them, but ultimately just very high agency and and and high competency people. But you can you can like probably like trim the company down to like 30%. Which is very scary because like I mean economically this will all this will all lead into a fiasco. Um, and a lot of people will like have trouble finding a a place in this new world but I'm not the least surprised that current companies cannot very successfully use AI. I mean, they do to a degree, but you you have to do a big refactor first, you know, like not just on your codebase, but also on your company. I design even on code bases I design the codebase not so it's it's useful it's easy for me so that it has to be easy for the agent I optimize for different things not always

1:36:45 the things that I prefer but the things I know work the best and and and have the the least friction for those models because I just want to move faster and

13 Highlight · Treating pull requests as prompt requests

1:36:53 ultimately they have to deal with the code not me I had I deal with the the overall structure and architecture and I can still do that in the way that I like everything has to be resold you know pull requests uh I see them more as prompt requests Now, like I don't I I somebody opens a pull request, I don't I lot I do say thanks and I think about the feature and then with my agent we start off with the PR and then I'll design the feature as I see fit. The agent rarely reuses like maybe I reuse some code but it's more like it gives the agent a good understanding of what the goal is and sometimes it's very useful because it's it's tricky bucks, right? But I I basically rewrite every pull request and and and v it in also also a lot of people let's just say the overall code quality of of PRs went down a lot because people vibe code and and building a successful feature still needs a lot of a lot of understanding of your overall design and if you if you cannot do that you will have a harder time steering your agent and the output

1:37:54 will be bad.

1:37:55 Yeah. And if you don't have the feedback loop to close it etc.

1:37:57 Yeah. Yeah. So, so I found it highly effective. Like I know at my at at PSVDF kit sometimes a pull request was like a week in the work

1:38:05 and you comment on it and then somebody has to context switch and you wait for CI for 40 minutes. No, I have the discussion. I see okay how would this affect something like I I I let the model review. They will already bring something up. I have some some ideas as well. We're going to reshape it into a form that fits my vision and then we weave in the code. It's literally

1:38:25 there's so many new words I use for like writing code now with those models which is so funny like weaving in code into an existing structure. Um and sometimes you have to like change the structure so it would fit. Now imagine that you would hire one or two people to make it a small team. How do you think in this world and you keep want to keep doing what you're doing? How do you think

Abandoning traditional CI for local agent gates

1:38:43 things like code review CI CD would change?

1:38:47 I don't I don't I don't care much about CI. I

1:38:50 Why why not? you used to care a lot like a PSPDF kid. You used to care a lot, right?

1:38:55 And and I still do this. There's value,

1:38:58 but I I have local CI. I'm I'm a little bit I'm a little bit of DAH now that

1:39:05 because because the agent runs the test, right?

1:39:07 Yeah. And it's just way faster. I don't I want to push on the on on the on the back on the P and then wait for 10 minutes to wait for for CI

1:39:15 because you waited 10 minutes on the agent already. I if if the tests pass locally, we merge and then yes, sometimes main slips a little bit, but it's it's it's usually very close because maybe sometimes I forget the and the agents call it gate. I don't know where that's coming from. Should I run full gate? So now I call it gate with g full gate is like linting and building and and checking and running all the tests. And I almost think it like because it's a it's a wall, you know, like it's like he calls the llinter and like the builder and the tester. It's almost like a gate before my code goes out. So I know obviously like okay once you're done like commit this run full gate like I'm I'm slowly adopting their language but but and if you hired like one more person to work on this you probably wouldn't do code reviews either. That's what I'm sensing. You would you'll probably trust this person to run like like pick up your working style right?

1:40:06 Even in Discord, we don't we don't talk code. We talk about architecture like big decisions like you still need to have style. Like there was like this one pull request that adds voice calling. So now like literally I can tell Claude, hey, can you call this restaurant and and and and reserve your seats and and it it it can do that. But it it it's a big it's quite a big new module that like touches a lot of places. I'm like you have to have this feeling is like this like I got a little this is like I want to merge this but oh this is becoming bloatware. So I so I I had this idea of like uh let's my typical way let's make a CLI out of it and I already had a project where I tried to solve something like this but I'm finished. So I opened up Codex and said, "Hey, look at this PR. Look at this project. Could we weave this feature in?" I again say weave. I'm like, "So I'm

1:40:57 let's keep it.

1:40:58 Could we weave this feature into the CLI? What are the up and downsides?" And then would like tell me like, "Oh yeah, I could do this and this and this." They give me honest opinion. Would this to me this sounds like it actually it would fit into the project? I was like, and was like, "Yeah, you would get this and this benefits that we cannot do if the next CLI." Okay, but I don't like this. This is getting blowware. could we build a plug-in architecture and then I will and do you know like one of the the secret hacks on on on using effectively is you reference other products like I constantly tell it look into this folder because I solved it there and I solved that there and all the previous thinking I did to like solve a problem well AI is so good at this still to like read the code and understand my ideas I don't have to like explain it again uh or if I explained again I might like make mistakes that like wouldn't get across was exactly the idea that I have in my head.

1:41:49 So in this case, I know that Mario who does like shitty coding shitty coding agent which is like a actually very much not shitty coding agent. It's called Pi. I know that he had this plug-in architecture that would load

1:42:00 code via via GT and because it's all TypeScript. So I was like, can you look into this folder and this folder? And then it just came up with this really insanely good plug-in architecture again by like being inspired by the people. And then that's why, you know, I have this feeling and then I I came up with Yeah, that's what I built last night basically. I mean, sounds like this is going to be completely different like you know PRs are are like in your workflow, you're not using PRs that much as CI is just different. It's it's tests are still doing there's an more important feedback loop. you're using things more like weaving instead of code. You're talking more about architecture and taste. It sounds like a pretty big shift to me. Now, in this world, let's assume you get to the point where you're going to you're hire the next one and two and three developers on this team. Let's imagine that this thing gets a life of its own and you know, maybe it's a business as well. What skills would you look for? And what would you advise a an experienced engineer right now? Who would you be excited to work with? kind of either expertise projects would you look for for someone

1:43:01 who sounds like who can work in this way or can pick up this way of working?

1:43:05 Someone who's active on GitHub and does open source and and someone where I have the feeling that they they love the the game. The way you learn in this new world is by like trying stuff and it very much feels like a game where you improve your skills as you get better like like a like a music instrument. you have to like keep trying and I and I that I'm now this efficient and this fast and I don't know like I I think like the other day I had like 600 commits in a single day. this is like completely nuts and it works like it's not it's not like there was there was a somebody did a code review and said like oh this is actually not slop and like yeah uh

1:43:46 there's a lot of skill that went into

1:43:47 yeah it's it's it's it's a lot of hard work but you need to play with the technology and learn and then you will get in the beginning it might be frustrating I don't know kind of like you you know you start going to the gym it's going to suck it's going to be painful but very quickly you like you get better and you and you feel that that your workflow gets faster and then you feel the improvements and then you slowly get hooked. So So but yeah, play and and yeah, also work hard.

1:44:19 Yeah, I I mean you're putting in more hours into this thing.

1:44:22 I've never right now I've I've never I never worked more even even when I had my company. I I've never worked so hard as I do now. Not because I have to, but because it's so addictive and so much fun, but also because right now I'm like using the moment where where this has traction and and there's a lot of people who like are pushing me. Um, and I I feel could it be cuz I think you have pretty good business sense. Not not not necessarily in the business business, but seeing when there's an opportunity there is an opening for to get traction, right? Like what you said for people to work in the open right now it seems novel. You're telling me you don't think you could, even if you wanted to hire, you don't think you could hire people cuz there's not many people working in the open, clearly using these things.

1:45:08 Fast forward 2 or 3 years from now, once a bunch of people start to do it and everyone does it, it's kind of like moot a little bit. So there there's also that a group that a lot of lot of people are worried about is the the new grads, the the people with with no experience who are either in school or about to graduate because of course you've been an experienced engineer by the time this came around. you know, you have a lot of things to build on. Putting that yourself into shoes of of someone like that and knowing what you know now, what would you recommend of of activities that they do, things that they build or

Cultivating infinite curiosity as a newcomer

1:45:39 try and you know like do would you recommend on focusing on the fundamentals of software engineering on this on the agents kind of mixing the two?

1:45:48 I would I would recommend them to be infinitely curious. Yes, it's going to be harder to enter this market. It's it's absolutely going to be harder and you need to build things to like gain experience. I don't think you need to write a lot of code but you need to I don't know you know there's a lot of open source that is complex that you can like check out and learn and you have an infinitely patient machine that is able to explain you all the things so you can ask you can ask all questions why why was it built this way to like gain system understanding but it requires real curiosity and I don't think universities right now are set up to teach you that in a in in a really good way this is usually something you discover through pain it's it's not going to be easy for new people, but but they have they have the benefit that they are not tainted by all the experience like they they they use agents in ways that we don't even think about again because they don't know that it doesn't work and by then it probably does

1:46:43 and also their friends use it all all the time

1:46:45 especially like like the other day I I have this little menu bar app for cost tracking on on on on cursor and and and cloud code and everything and it was a bit slow. So I I was like, "Okay, let's do let's do performance measurement." And and my old way is like I open instruments and click around. And it would just call and do everything by the terminal. It blew me away. I was like I didn't even have to open instruments anymore. And it just like made it faster. And then did like some recommendations. I'm like all of that sounds good. Do it.

1:47:15 Yeah. I I think we might be underestimating both like how resourceful people entering tech have been also how young people if I think about some of the great companies started they were very young and obviously very inexperienced but had a lot of passion so so that's there as well yeah it's it's a big opportunity I I'm especially taking in like it's I have to take it in like but all the things you mentioned about just your way you know weaving code in not caring about PR not caring about code reviews it's a big change because these things have been us with like again for like 15 plus years of your life they have been in fact you know a lot of it has been kind of you know solid building blocks of PSPDF kit right yeah we need we need a lot of new things even you know even when I get a PR I'm actually more interested in the prompts than in the code I I ask people to please add the prompts and some do and I I read the prompts more than I read the code because to me this is this is a way higher signal of like how did you get to

1:48:15 the solution what did you actually ask how much steering was involved then the actual out to me this gives me more idea about the output I don't have to read the code or like if someone wants a feature I ask for a prompt request like write it up really well because then I can just point my agent to the issue and it will build it so because because the the work is the thinking about how it should work and what the details are and if someone else does it for me I can literally say build and it will work uh and I and then yeah of course I think about But it will will re really or if someone sends me a PR that are just a few fixes. I told people please don't do that. It takes me 10 times more time to to review that and just type in fix in in in codex and wait a few minutes. So there's there like all these insane things that are like would have been completely different

1:49:03 even even at the beginning. uh now we

14 Highlight · Automating repository onboarding with agents

1:49:05 have a oneliner but for the last two weeks like when it got really traction I told people to just just point an agent at the repository to configure it so so I didn't had I didn't have an onboarding but we had cloud codebased onboarding where claude would like check out the G repository read the things and write the configuration for those people and set everything up so it works like set up a launch agent that didn't have the manual setup because it was not a priority anymore because because agents can now do that for you and the and the and since the the product was built by agents. They structured it exactly the way agents expect things to be named and things. There's certain ways that are encoded in in in the in the weights how they expect things to be named and everything exactly like how they expect.

1:49:50 So they are really good at navigating their product. So it was not a priority to work on onboarding as much. I mean eventually I wanted this magical experience but it was more important to like make sure that your message arrives

Recharging through physical gadgets and gym routines

1:50:01 and that things don't explode. So onboarding was literally like type this prompt into your agent. Um which is would have been mind-blowing even a year ago.

1:50:10 All right. So to to wrap up, we'll do some rapid questions. So I'll just ask and you tell me what's on your mind. What's a tool that is not a CLI, not an ID, it can be physical that you you use you like you would recommend? I buy a lot of gadgets and many of them dust away. But there's this one kind of crappy thing that was not expensive that gives me almost unlimited amount of joy.

1:50:37 And it's like this Android powered photo stand where I can upload pictures and where like it has an email address and friends can send pictures and it will just show pictures. And I and I put a few a few in in my house again. And I mean even the animations are a little croppy because it runs nward and it's terrible from the technology but but it gives me infinite number of joy because it is low tech that just shows pictures and reminds me of happy moments in my life and it was like 200 bucks and I don't know to be honest that gets me more joy than the latest iPhone. I bought the iPhone 17. I still haven't unpacked it because I just in my head I wanted it but then I couldn't get around to it because it's just a hassle to like move the Sims around and I said like basically no no no no feelable benefit but like this little this little device gives me infinite joy.

1:51:25 What's something that helps you recharge outside of tech like just or just moving away from from tech and screens? What keeps me sane, even if I work crazy hours, is going to the gym, even better, working with a a coach and leaving my phone in the locker. And then I really have like a good hour where I just feel me and I and I'm like in the moment and I'm not distracted by notifications or tempted to like touch my phone like we need more time for this. Or even sometimes I I I go for a walk and I I leave my phone at home and it feels very scary. It's almost like a it's almost like an organ now, you know?

1:52:14 It's like your your body knows where it is and if you don't know where your phone is, you freak out.

1:52:18 I I'm having I'm having a blast.

1:52:20 Love it. This is great, Pete. Thanks very much. Well, this was a super interesting conversation and it feels to me that how one person teams built software with AI is already completely different to what we've been used to. One thing that really caught my attention is how Peter thinks in prompts and not pull requests and how he weaves in the code and no longer merges the code. He doesn't find poll requests all that useful and would rather get prompt suggestions even on GitHub. I do think we might have to rethink the importance of prompts or at the very least sharing of prompts in software development the more we use AI and AI agents. Another thing that struck with me was Peter's emphasizing how important is to close the loop. As Peter explained, the reason AI is so good at coding but often mediocre at writing is because you can validate code. You can compile it, run tests, check the output. So, the secret to making AI system development work well is to design your system to close the loop and have the AI run the test.

1:53:13 Finally, I was wondering if Peter is in the flow as much even when he's not writing code. Turns out he is. He's in the flow more than ever. And he told me that it's mentally more exhausting to juggle several AI agents in parallel than it was just to write code. My feeling is that someone who was a great developer without AI can be an excellent kind of code architecture or carding person with AI. This is just a gut feeling I've had so far, but Peter seems to prove it. Finally, we should note that Clawbot is more of a YOLO project than most production apps. So, take the approaches that we discussed with a grain of salt. At the same time, I do think that a lot of what Peter does could well spread to building production code, except review and validation will become a much more important step in those projects. If you enjoy this podcast, please do subscribe on your favorite podcast platform and on YouTube. A special thank you if you also leave a rating on the show. Thanks and see you in the next one.