Minimalist Coding Harness Pi

Something different this week, we actually learn something! Dillon joins us to explain why he switched to Pi for his AI coding workflow. We dive into the importance of context management, building custom UI extensions for your AI agent, and why minimalism often beats "token-maxing" with bloated system prompts. From automating household budgets to intercepting bad git commits, discover how to mold an AI harness around your specific developer taste.

• 0:00Speaker 00

  Oh, now you're back, Trash.

• 0:01Speaker 01

  Hello?

• 0:03Speaker 00

  You're here.

• 0:03Speaker 01

  Can you hear me?

• 0:03Speaker 02

  Oh, he's back.

• 0:04Speaker 00

  Yeah, what's up, Trash?

• 0:05Speaker 01

  Yo, what up? I love, I have, my internet's fast. I don't know why it just froze like that. Was it just me or everybody else?

• 0:11Speaker 02

  Trash immediately just like insecure about his internet. We didn't say it was slow.

• 0:16Speaker 01

  I swear, man, I got gigabit, dude.

• 0:18Speaker 00

  It's really fast. This never happens to me, bro. I swear this never happens.

• 0:26Speaker 03

  Uh, anyways, I'm sorry.

• 0:27Speaker 02

  Thanks. We don't have time for the intro. Wait, let's call Prime just to do the intro.

• 0:32Speaker 00

  Well, but I kinda wanted to actually have enough time to learn something today, so we can't wait 45 minutes for Prime to start the episode.

• 0:39Speaker 01

  I like whenever Prime's not here, we actually learn stuff.

• 0:42Speaker 00

  Yeah. Well, I'm like, how can I sneakily borrow Dylan's time for an hour? Well, just call it the stand-up. It's smart, you know?

• 0:52Speaker 01

  There you go.

• 0:53Speaker 00

  Dylan, so why don't you tell us a little bit about what we're gonna be doing today? Cause all I asked you for was teach us how you're doing AI stuff with Pi, cause I'm interested to watch that happen. You also have a podcast, but I'm gonna preempt you. I also have a podcast where we talk about this. You have an episode talking about this, but I wanted to ask particular questions as I go, you know?

• 1:12Speaker 03

  Yeah.

• 1:12Speaker 00

  I wanted to get the personal podcast experience.

• 1:15Speaker 03

  Yeah, listen, I've got, I've got hot takes for days on, on this topic, so we can yap about it a bunch. So I don't really know what we're doing today.

• 1:24Speaker 00

  Plug the podcast. You're supposed to say what your podcast is.

• 1:28Speaker 03

  Yeah, so, so Ree Solvin, Sunil Pai and I started a podcast called The Next Token. You can find it, I think, pretty much everywhere, and we try to put a realistic spin on what it's like working with AI as people that are actually employed, building real software for real users and not just yapping on Twitter or working at a lab and pushing token maxing.

• 1:54Speaker 01

  Yeah, Prime. Yeah, Prime.

• 1:56Speaker 00

  Yeah, what, so Dylan, why don't you spend $5 million a month on tokens? And why are you being left behind?

• 2:04Speaker 03

  I am so far left behind. It's almost embarrassing at this point.

• 2:08Speaker 02

  I can kinda say what I, so I've been, I listened to the Next Token podcast. I've been using every single harness under the sun. I literally this weekend, I started, or this Friday night, I started my Friday night by literally getting, like, I have Claude, I have Codex, I have Open Code, I have Cursor, I have Pi. Literally next to each other, I wrote one prompt and then I'm like, let's go, let's see, you know, where I'm at with the tools and whatnot. And I just realized with Pi, I'm like, an idiot. It's like, look, do whatever you want. I'm like, cool, I don't know what I, I rebuild Open Code with it? So like, I just feel the flexibility of it. I don't know what people are doing that's cool, other than rebuilding the tools I already had. So I wanna hear, yeah, like, what do you actually do with Pi? Do not just say you do custom stuff and then not tell us what the custom stuff is.

• 2:52Speaker 03

  Yeah, yeah, no, no, I'm fine with that. So the, for anyone that doesn't know, Pi is a notably shitty coding agent, built by Mario and Arendelle. It is intentionally very, very minimal, doesn't have subagents, doesn't have MCP out of the box. The system prompt is like, I don't know, probably between 1,000, 2,000 tokens. That doesn't change very often, which is something I like about it. It's very extendable. It has, it knows how to like update itself. So you can be like, oh, build me an MCP plugin, like, look at Open Code for inspiration on how to implement it.

• 3:34Speaker 03

  And it knows how to like go implement it with its, plug into its system and go do it.

• 3:42Speaker 03

  I ended up on Pi. I wanted to try it because I had read the code. A remarkable thing in this day and age.

• 3:50Speaker 00

  Last year.

• 3:51Speaker 01

  I don't believe you.

• 3:52Speaker 03

  Remarkable in this day and age, but I had read the code and I had seen Mario

• 4:02Speaker 03

  using abort controllers responsibly and proactively and correctly. And that is something that almost nobody does in TypeScript code.

• 4:13Speaker 00

  Unless you're using, you're using Effect.

• 4:15Speaker 03

  Yeah, if you're using Effect, by the way.

• 4:18Speaker 02

  Yeah, everyone, everyone here on the podcast uses abort controllers perfectly reasonable. However, for some of the people out there who might not know what those are, maybe you wanna explain. I'm not part of that group, cause I use Effect, and that's part of Effect.

• 4:31Speaker 03

  Yeah, we, yeah, really just use Effect, but not, not React use Effect.

• 4:36Speaker 00

  No, no, no, no, no, no.

• 4:38Speaker 03

  So abort controllers are a way to manage and handle interruption and cancellation of async work in TypeScript and JavaScript and to accurately use them.

• 4:54Speaker 03

  Much like many things in plain TypeScript, it requires a shit ton of discipline and, yeah, really just discipline and like basically prop drilling, abort signals down through call stacks. It's miserable, but it does make a huge difference in the quality of software. Like, a good example is, maybe you have like, maybe you have like a docs LLM or like an LLM agent in your doc site and users are yapping away at spending inference and like, I ask some question to the docs and before the answer comes back, I navigate away. Like, you don't wanna go pay for all that inference, right? Like, you can use an abort controller when the user leaves to cancel that request and then like save yourself a bunch of money.

• 5:36Speaker 03

  And there's, there's a bunch of different use cases, but like, that's one that I think kind of like demonstrates the idea well.

• 5:41Speaker 04

  Did you know that I have done a lot of weird stuff that's difficult to explain to my father-in-law, such as building a tower defense game in a water tower, pretending to be Mark Zuckerberg in a social network parody in SF? By the way, stands for San Francisco, or even did a live Super Smash Brothers event at Open Sauce. Well, if you didn't know those things, you should just go check out the Terminal channel. That's where I do all the fun group videos. Check it out. Links in the description.

• 6:08Speaker 03

  So anyways, I saw in the code Mario, you know, I was like, oh, this is actually like very good TypeScript. Like, it is very rare that you find a codebase that accurately uses that. So I was like, alright, you know, I'll give it a try. Up until this point, I've been using Open Code and this is like, I don't know, it had to be back in like March. So like, I didn't really start using AI heavily until, you know, the beginning of December at Opus 4.5 came out. So like, I was still.

• 6:35Speaker 00

  Just before you came and deleted our front ends in San Francisco.

• 6:38Speaker 01

  Oh, yeah, remember that?

• 6:39Speaker 03

  Hey, hey, we had, we were the first to have an AI agent escape containment and go do something wrong, alright? Kimmy K, we were there first.

• 6:50Speaker 00

  Yeah, we were, we were first. We were the first to experience Kimmy K's bad side, okay? She crazy.

• 6:56Speaker 02

  Also, Dylan, for life, you're gonna be catching up for those like three months where you weren't, you know, fully AI psychosis. You could have been AI psychosis three months earlier and think about how much more ahead you could be.

• 7:05Speaker 03

  I know, I could, I, I could be like token maxing, looping and.

• 7:11Speaker 01

  I got a question. So the way you describe Pi, it sounds like it's like the Vim of like the AI space. Is that like correct? Like, or it's really thin and you just add on things effectively?

• 7:23Speaker 03

  Yeah, I think that's a probably good way to position it. I think it's probably more friendly than Vim. Like, I think it's way harder to get started with Vim than it would be.

• 7:37Speaker 01

  Well, I guess like one of the aspects of like minimalist type of approach.

• 7:40Speaker 03

  Yeah, and I can get into why I think minimalism is important and kind of what, that's kind of what made me jump over to Pi. It's not like I was intentionally going after like an N Vim-based like, you know, harness or anything. Like, at that point, I didn't have strong enough opinions about like even what a harness was and what it should do.

• 8:02Speaker 03

  So it was like very vibe-based at the time.

• 8:06Speaker 03

  So I, I tried to start using it after I saw, actually, I think a better way to position it than comparing it to N Vim is, TJ actually put out like a couple of, maybe one video or a couple of videos like a couple of years ago on PDEs, personal development environments. And like, I really feel like because of the way you can work with Pi, like you can mold it around your workflow and individual repos, like Ben Vinger from modem had a good.

• 8:43Speaker 03

  Conversation, I think on his, his podcast with Armin, talking about how he shaped and just for a particular, a particular project and use case, he shaped pie around like editing. Uh, his podcast, like the tools and extensions to make that easier. So it's just like, it's very easy to do that kind of thing. Um, so the first time I use pie, I like immediately went to like the, it has like an extension. Uh, marketplace on its website. And I was like, okay, I need sub agents. I need my MCP. I'll look at this cool answer tool. Look at, I started, you know.

• 9:17Speaker 00

  No, you did the, you did the Neo Vim thing of, well, I need every plugin.

• 9:21Speaker 03

  Yeah.

• 9:21Speaker 00

  I gotta have literally every single one.

• 9:24Speaker 02

  This is why I wanted to do this with Dylan, because that's what I'm doing. And I'm like, I don't know what the point of this is. I'm building a really bad open code.

• 9:31Speaker 03

  Yeah.

• 9:31Speaker 02

  That I'm like vibe coding pieces together and then it's bad. And I'm like, I don't know what, why. So.

• 9:36Speaker 03

  Yeah. So that's exactly what happened to me. The first time I tried to use pie, I turned off it in like a week because I was like, I'm just making a shittier version of open code. Like my, my expectation of harnesses at the time is like, oh, open code is kind of like this gold standard here. And like, you know, it doesn't have the, the jank that Claude code does. Codex was like, Codex wasn't really a thing yet. Um, I mean, it was, but it wasn't like getting huge market share yet. Um. So I turned off it because I was like, oh, my sub agents are terrible and my MCP plugin keeps not working the way I wanted to. And it doesn't connect to Cloudflare AI Gateway the way I wanted to. And I don't know all these things I just installed and they're updating all the time.

• 10:17Speaker 00

  The best place to get AI, just wondering.

• 10:20Speaker 03

  The best place.

• 10:22Speaker 02

  We're gonna ask them for a check after this too, because we were officially started Agents Week. Heads up. It's like.

• 10:28Speaker 03

  Hashtag Agents Week.

• 10:30Speaker 03

  So I turned off and then maybe, I don't know, a month later, I tried again. Uh, I started probably getting more opinionated about like context management in the period between that and like, really getting a sense for like, how does? How does, in particular, like Agents MD impact the performance of my model and the harness? And how are skills playing a role in that? And so I started developing stronger opinions in that time. And then I went back and I tried pie again. And this time I explicitly was like, I'm not going to install.

• 11:08Speaker 03

  A single extension until I feel a pain point. And when I feel pain, that's when I will go and look for something. And I will pay pretty good attention to like how it's implemented, like go read the code, uh, and just figure out if that's the right solution. Otherwise I'll build my own. And I, um, I'm at the, I, I really, my pie, you can look at my dot files. Like I do not have like a lot of extensions. Um, I have a, I ended up copying, uh, open codes MCP implementation because we use, uh, Cloudflare Portal. And we have a bunch of like business MCPs inside that just make life so much easier. Um, I have.

• 11:52Speaker 03

  Uh, let me even just pull up my dot files real quick and I can just read this off. There's like literally like, I have a custom portal.

• 11:58Speaker 00

  You can show us.

• 11:59Speaker 03

  Oh, I can, I can screen share. Oh my.

• 12:01Speaker 02

  We actually have the technology.

• 12:02Speaker 00

  Oh yeah, of course. Very advanced.

• 12:06Speaker 02

  Cloudflare Agents Week, this is their new screen share feature. It's brand new.

• 12:12Speaker 03

  Don't have to push.

• 12:14Speaker 00

  To see if it works or not before we say that.

• 12:16Speaker 02

  Oh yeah. Yeah. You're so smart.

• 12:18Speaker 01

  Hey Cloudflare, screen share.

• 12:20Speaker 00

  Just, I just take a moment for the podcast listeners. Um, I picked this very simple way of sharing this screen, which is I just literally am sharing the browser window. And I just want everyone else to know that it just immediately worked. We didn't have to do 37 different OBS transitions. And like, and I didn't have to do a stinger. I didn't have to the primogen on you or anything. We just like left the screen the way it was. And now we can see Dylan's screen. So just like.

• 12:47Speaker 00

  I just want to throw it out there. I have, I didn't even practice. So for long time watchers of the pod, I think they'll be impressed with some skills here. So.

• 12:55Speaker 02

  How many tokens did that cost though, Teach? That must've been like running for like, since the podcast started?

• 13:01Speaker 00

  It's been, it's been running, uh, a GPT 3.5 for 16 months. Yeah.

• 13:09Speaker 03

  Excellent.

• 13:09Speaker 00

  Okay. Sorry. Anyways, Dylan, you may, you may continue now. Go ahead.

• 13:12Speaker 03

  Okay. So my extensions, uh, this is all of them and most of them aren't really that important. I don't know why there's a test directory in here.

• 13:20Speaker 02

  What's in whimsical dot TS?

• 13:22Speaker 01

  I thought you read the code Dylan. Why you mean, I don't know what that's doing there.

• 13:25Speaker 03

  Oh, listen, I test my code as well. Uh, something I know about, uh, trash.

• 13:30Speaker 01

  It's true. My agents.

• 13:32Speaker 03

  These are just like the things that like, when the agent is loading, like, or doing stuff, it just cycles through loading messages.

• 13:39Speaker 01

  Is that, is that, is that Claude codes? Did you take that from Claude code?

• 13:42Speaker 02

  You stole this. You stole this.

• 13:43Speaker 03

  I, uh, I slop forked it from Armin.

• 13:46Speaker 01

  Oh, no.

• 13:47Speaker 02

  Oh, I recognize these phrases.

• 13:49Speaker 00

  You thesaurus that?

• 13:50Speaker 03

  I did add some, I did add some, uh, functional programming ones in here.

• 13:56Speaker 02

  I will say, wow, are they.

• 13:58Speaker 03

  Please see composing.

• 13:59Speaker 02

  Using AI to make dynamic like waiting? No, it's, it's a list of, uh, 600 strings in a file.

• 14:04Speaker 00

  Oh, the old AI. It's a bunch of if statements back in the day.

• 14:09Speaker 03

  Uh, okay. Here. Uh, so this one is just literally like a custom gateway that we use that like I just connect to, to get my inference through.

• 14:19Speaker 03

  And then we have pie cloak. Uh, if you're a streamer and use Neo Vim, uh, you should be familiar with cloak dot Vim. Uh, it basically masks environment variables if you open them. So like if my agent. Reads like a environment file or something like that, that looks like a token, it will just like start it out in the output. Uh, I have this thing I built called.

• 14:41Speaker 00

  Just a quick note, chat's saying right now, you still leak them?

• 14:44Speaker 03

  Uh, chat's, chat's full of it. They don't know what they're talking about. Okay. That never happens. I, I've never, I never leak. I've never leaked ones in my life.

• 14:55Speaker 03

  Pie skill toggle, uh, skills have like this, some harnesses will respect this setting. You can, uh, change in the front matter. That is, I don't know if I'm going to be able to find it. Um, but you can set, um, a thing in the YAML front matter of a skill that toggles it. From being agent invocable or only human invocable. Um, and when it's only human invocable, it doesn't get injected into the agent's context window. Um, so I have that to basically set the skills I want in my session. Um.

• 15:26Speaker 01

  Wait, so what do you, what do you, are you just adding a new field in front matter?

• 15:29Speaker 03

  Yeah, it literally just writes the front matter field.

• 15:32Speaker 00

  Oh, good. This is a good one. I've seen you tweet about this one. I'm pretty sure.

• 15:35Speaker 03

  Yeah, I also stole this one. People keep tweeting like, I created this, but I did not. The, uh, person who made Planetator made this and I stole it from them. Uh, so please give credit to Michael, uh, creator.

• 15:48Speaker 00

  Dylan, we can, we don't need to do that here. We can just do the Kit Langton effect and, uh, we can just make it so that you created skills.

• 15:55Speaker 03

  You made this?

• 15:56Speaker 00

  You just created skills.

• 15:57Speaker 01

  Right now Dylan just made the sickest skill.

• 16:01Speaker 00

  Oh, I was thinking we go one step further. We just say Dylan made skills.

• 16:04Speaker 01

  Okay, true. Wow.

• 16:05Speaker 02

  Yeah, Dylan, the inventor of skills.

• 16:07Speaker 01

  Dylan.

• 16:07Speaker 00

  Dylan, the inventor of skills who also.

• 16:09Speaker 03

  I don't know if I want that on my.

• 16:11Speaker 02

  I don't think you want that one. I think it's a hot potato.

• 16:13Speaker 00

  We need to see, we need to see how it plays out over the next few years, if that's a good one or not. Yeah.

• 16:18Speaker 02

  Yeah, I don't think you want skills to be your, uh.

• 16:21Speaker 01

  Yeah, I have an Eli five skill. Help me like understand all the crazy words they say.

• 16:27Speaker 03

  You should try this one. It's actually very good. Yeah. But anyways, so that, that extension literally just toggles this line and then reloads the harness. Uh, so we're.

• 16:35Speaker 01

  Did this, did this start from Matt Pokock or did everyone else kind of just?

• 16:39Speaker 03

  What?

• 16:39Speaker 01

  Cause I know, um, the disabled model invocation.

• 16:42Speaker 03

  I have no idea where it came from. Okay. I thought it was part of the official spec, but apparently it's not. Maybe it is. I don't know. Um.

• 16:51Speaker 02

  Can you break down for the people in our audience who don't know what that does? I obviously do.

• 16:55Speaker 03

  The disable model invocation?

• 16:59Speaker 03

  This thing?

• 16:59Speaker 02

  Yeah, yeah.

• 17:00Speaker 03

  When you open your coding harness and you have a bunch of skills, if this is false or not included in the front matter of the skill. It will load the description and name of your skill in like two XML tags. Like, so like when. Skills get loaded into your context window, it's like an XML bracket skills. And then each skill has like skill. And then it'll have like skill name and skill description. And that just will always sit in your context window. And kind of like I was mentioning earlier, I became very sensitive to what is in my context and like. Anything that is not helping me for my particular task is just noise and will cause more deviation than I want.

• 17:43Speaker 03

  Uh, from the agent doing the right thing. Uh, so I basically built this thing to toggle this. And, um, when it's disabled. It does not load the skill into the context window.

• 17:55Speaker 01

  It's the only way you can invoke it is through like a slash command, right?

• 17:58Speaker 03

  Yeah, the agent can't discover it then and right.

• 18:00Speaker 00

  Right. Which is nice, especially for some stuff like maybe, uh, I have it, you know, for like cursor cloud agents, it'll record the, a video of stuff sometimes. But it would be nice if I had like it, I sometimes just tell it, don't record videos for this. You know what I mean? But it's like, it would be nice if it just didn't even know that it could record videos so that I don't have to tell it not to record videos for certain things. You know what I'm saying? Like that, that would be this kind of, don't use the browser for this. Just run the cli. Stop spending tokens opening up a browser. Like, you know, it's that kind of thing.

• 18:30Speaker 03

  I will say also, if people do like adopt, uh, an extension like this, that when you do run this, it will cause a. Cash invalidation and you'll get charged for your full context window again. So just be, be aware.

• 18:45Speaker 00

  At the beginning.

• 18:46Speaker 03

  Yeah. Start a new session or if you're okay with just getting charged again. Uh.

• 18:52Speaker 01

  I'm not paying for it. It's okay.

• 18:54Speaker 03

  Uh, that's also a hot take.

• 18:55Speaker 00

  Netflix shareholders are.

• 18:57Speaker 01

  I was just kidding. I was a part of that. No, stop. Rewind. Rewind, rewind this, rewind this.

• 19:02Speaker 02

  At some companies, you get a bonus for using more trash. You're using a lot of tokens today. Very good.

• 19:07Speaker 01

  Not, not at my work.

• 19:08Speaker 00

  Oh my goodness. Vegan. Instead of a mouse jiggler, we just have a cash invalidator. That's the 2026 version of a mouse jiggler for pretending you're at work.

• 19:19Speaker 01

  It's true.

• 19:20Speaker 02

  Well, I have, well, I have two questions about this because this is, it is what I also hear from pie people. You hear pie people talk and they go, yeah, I'm really becoming context sensitive. I've heard multiple people say that. And I want to understand like, what does that actually look like in terms of like, how are you monitoring? So far, what I've done is I start adding more stuff to my agents. I add more skills and then like, after like a week or so of that, I rage quit at. 2am and I delete everything and I start back from blank. And I've just done that process like six times. And I don't, I'm not building up to anything. I don't have any like scientific method of knowing if like. My, the, the context that I'm building up full of like my little tips and tricks and agents and my skills are actually helping or not. So yeah, how are you doing this?

• 20:03Speaker 03

  Uh, I also, uh, pretty regularly rage quit at two in the morning. Uh, so you're not alone. Um. I would say some of it is vibes, but there are more, there are benchmarks out there showing that Pi score is exceptionally high when compared against other harnesses. I don't know what the exact links are, but I'm sure you can go find them. They're out there and it's actually like a legit thing. The other thing is that there was a blog post, or maybe it was included in one of these benchmarks that like the initial prompt on any given harness, you have to pay every time you start a new session for all of your system prompt and all of your skills and tools.

• 20:48Speaker 03

  And because the system prompt of Pi is so minimal, like it's like 2,000 tokens where like Claude code. I know Claude code recently trimmed it down, but before it was like, you know, you're getting charged for like immediately you're in to like 22K of token usage just for the system prompts.

• 21:07Speaker 00

  Why would Claude do that?

• 21:09Speaker 01

  So question like, what are they adding to the prompt?

• 21:11Speaker 03

  I can't imagine why, why a inference provider would want that.

• 21:17Speaker 00

  It was crazy. I went to the gas station. They had this, they had this thing and you just poured some gas into an empty bucket first before I got to put it into my car. I just put $20 of gas into the empty bucket before I get the film.

• 21:29Speaker 01

  I have a question. I have a quick question. So what are, what are these harnesses adding to your prompt? Like, what are they concatenating their strings to or your strings to? Like I'm from Claude, I'm from Claude code. This prompt's from Claude code. Like what, what's it doing?

• 21:44Speaker 00

  What are they concatenating their strings to? I love that trash.

• 21:49Speaker 01

  Cause that's what it all is. We're all just, we're all just string builders. That's like all this is at the core.

• 21:54Speaker 03

  So, uh, let me just get to a fresh session here. And if I do toggle skills, right?

• 22:03Speaker 03

  Oh yeah, I can close that out.

• 22:03Speaker 01

  This guy streams.

• 22:04Speaker 00

  No, no.

• 22:06Speaker 03

  That was, that was my chat, not your guys' chat.

• 22:09Speaker 00

  Yeah, I know. It was dangerous. They were, they were about to start saying crazy things.

• 22:13Speaker 03

  So when you load in, you can see what's in the context. It's like, it will automatically read the agent's MD. It'll be the system prompt. Some of these skills will not be toggled, but like in here, the ones that are checked white. So Pi just made this. I was like, build me an extension that lets me toggle the agent invocable front matter in a skill. And it just like built this UI and like, I can just go toggle this on and off. And when it's done, I can click, uh, what is it? Control S to save and then it reloads. And now it's not invocable by the agent, or it is. I don't remember which way I toggled it.

• 22:52Speaker 03

  So that's, that skill, uh, not skill, extension. The other ones I have are.

• 22:59Speaker 00

  So the extension, your extension for your Pi skill toggle also has the code for that UI piece.

• 23:06Speaker 03

  Yeah. Yep. Exactly. There will be, yeah, like you made a UI folder.

• 23:10Speaker 00

  Gotcha.

• 23:12Speaker 03

  And like it, the system prompt for Pi has instructions for how to like build extensions and UIs for itself.

• 23:19Speaker 00

  Gotcha. Pi is all terminal only. People have built UIs for it. I'm assuming as well, of course, non-terminal ones.

• 23:29Speaker 03

  Yep. Yep.

• 23:31Speaker 02

  In fact, I think. We could maybe fork it, make a UI version. It makes a lot of money.

• 23:36Speaker 00

  Call it Tao. Twice is good.

• 23:39Speaker 03

  Splitting it into client and server architecture. So we could have like a different front ends eventually.

• 23:45Speaker 00

  Nice.

• 23:46Speaker 03

  Uh, another one I have is this little save markdown thing. And this literally just saves the last agent message to a markdown file.

• 23:56Speaker 02

  That's useful. I need that.

• 23:59Speaker 03

  Whoops. I did the wrong one.

• 24:02Speaker 03

  It's a markdown, save markdown. Jesus, can't use my own skills.

• 24:07Speaker 00

  He's got, we will cut that out in post. Don't worry.

• 24:09Speaker 02

  We'll cut, we're gonna make it look so good.

• 24:11Speaker 00

  Uh, Josh, definitely don't zoom in on that and show the warning or anything.

• 24:14Speaker 03

  I'm skill issuing myself with my.

• 24:16Speaker 01

  I'm like, it's working well. It's working well. It's working well.

• 24:18Speaker 03

  Save markdown to.

• 24:20Speaker 02

  Yeah, I was gonna say.

• 24:20Speaker 03

  Saved it to foo markdown.

• 24:22Speaker 02

  You also are showing off something else, which I was gonna ask about, which is you're on Herder. People are asked.

• 24:28Speaker 01

  Don't look at my Herder. Yeah. After Dylan did a podcast with his buds and he showed Herder, everyone's like, I'm gonna start using Herder on Twitter. I was like, oh, I see what's happening.

• 24:39Speaker 00

  Big AI influencer.

• 24:41Speaker 03

  It's. Actually very, like I, I actually think that even if you aren't using AI or if you're using the terminal pretty exclusively and you're using TMX, I think Herder is a better multiplexer than TMX. Like it supports kitty image protocol. It's built on lib ghosty. And like, I know Mitchell, uh, just started the company to build and like the first thing they're building is a multiplexer. But until that's ready, uh, Herder is like a very quality piece of software.

• 25:10Speaker 02

  I will say, I love the idea you're saying, even if you don't use AI, I'm like, I need to hear this use case so bad. You're gonna tell people. Listen, it's actually useful for, I don't know. I am not on Herder. I did switch from, I will maybe a long time. The only YouTube video I've ever made is about TMX. So that's how, that's how you know I actually care about it.

• 25:29Speaker 01

  I use my term newbie.

• 25:31Speaker 03

  I was using, I've been using TMX since 2014. So like when, when I started using Herder, I was like, this has to be like exceptionally good for me to switch. Because like I've, I've had my, I've been pulling my config around with me for 12 years now at this point. And, uh, like basically my key binds and everything just transferred over like pretty much one for one. And it just performed better and had like nice features that I wish TMX had.

• 25:59Speaker 03

  So 10 out of 10 recommend. Shout out, uh, Ken for making a very good product.

• 26:04Speaker 03

  So that's save markdown. Web tools. I literally, uh, pointed my clanker at open code and said, copy web fetch and web search. So, uh, that's just, it doesn't have web search. Like if you asked Pi out of the box to go search the web, it's gonna like go curl and W get stuff.

• 26:22Speaker 02

  Can we dive deeper in that? Cause this is another like fundamental philosophy change that has been happening right now. And I want to like actually hear how you think about it. Back in the day, people would say, ah, here's a tool. Instead of using the tool that's already well-written and tested by everyone, I'm going to quickly build a, a shoddy version of my own. Now that is the recommended pathway. We are like web fetch, a thing that works that we know about. No, I'm going to go look at it, clone my own. Are you telling it to like strip off these feet? Like what's the point? Are you telling it like, I want a more minimal version. I want just this. I want, like, why would you not just use the thing that already exists out there?

• 26:55Speaker 03

  Well, I mean, open codes is like baked into open codes codebase. It's not like I could install open codes web fetch tools. Uh, it's just like a core piece of infrastructure and service inside open code.

• 27:10Speaker 02

  Okay.

• 27:10Speaker 03

  So when I built this, I was like, I was like, oh, how, go read the open code codebase and trace all the, the call stacks and abstractions that it uses for its web fetch and web search. And tell me all the configuration options and its entry points and how it works and how it does error recovery. And then once it came back with all that research, I saved that to a markdown file. Uh, I used a feature called slash tree in Pi, which we can get into then to get to fresh context. I had a fresh agent read that markdown from the research it just did. And I said, build this for Pi and that was it.

• 27:50Speaker 02

  Nice. So there's that.

• 27:51Speaker 00

  I want to go over a tree, but before, before we go there, I was kind of interested in, um,

• 27:58Speaker 00

  your thoughts on, uh, cause initially you mentioned making, you know, Pi custom for a particular thing. Are you making separate pies for separate projects? In my head, that's kind of where I like, it's interesting for me to think about of like a kind of either a shared harness with my team, or maybe some shared, you know, plugins or extensions that can help my team to be like, hey, this one pulls the logs perfectly from Sentry exactly and moves it to the exact linear board. And we don't need to do like 25,000 tokens every time to do that. We just like run from 25 lines of TypeScript and it's instant to do that. Like, is, are, are you experiencing that? That's kind of where I'm, I'm most interested for something like this.

• 28:43Speaker 03

  I have, I have things in, uh, like I have extensions and skills in, uh, the artifacts repo, which is what I work on at CloudFlare.

• 28:54Speaker 00

  Oh, cool. Are they launching anything for agents week?

• 28:59Speaker 03

  We did actually. We, uh, have programmatic C I C D now with artifacts. Uh, so you don't need GitHub action.

• 29:05Speaker 00

  Well, I had a complaint because it was an alchemy. So we'll, we can discuss that. We can discuss that later.

• 29:09Speaker 03

  Listen, don't, don't, don't tell CloudFlare, but I like alchemy a lot more.

• 29:13Speaker 00

  Yeah, no, I do that. I could see from whatever, but that wasn't the, I didn't, I'm not trying to get you in trouble here Dylan. We'll talk about that.

• 29:20Speaker 03

  I'll own that all day. Alchemy is very, very good.

• 29:24Speaker 03

  So yeah, uh, for work, I have, I have custom stuff in different projects just to make things like, you know, going out to like Grafana and Sentry and like our other data stores to get like production logs or traces for that. Another thing that I use, I have.

• 29:40Speaker 00

  Quick question about that. Sorry. Before you go.

• 29:41Speaker 03

  Yep.

• 29:42Speaker 00

  Are those shared for your team or they're not committed to get, they're just your local stuff? Or like, how does that work?

• 29:47Speaker 03

  I don't remember if I have the dot Pi folder committed or not. I'm not sure if, uh, Zeb and Michael, the two other guys on my team are using Pi or not. I should probably commit them.

• 29:58Speaker 00

  But conceivably you could, or you could have some subset of those things committed. That's like shared and other things ignored. You could have your own set. It's just, it just loads them from like particular folders or what? How does that happen? It discovers extensions?

• 30:12Speaker 03

  Yeah. You, so like you can, how, this is like my global Pi folder. I just use Stowed to mount it into my home directory. But it works like open code where like you can just put a dot Pi folder in any directory and it will like inherit whatever you put in there. And it's like.

• 30:28Speaker 01

  That's pretty cool.

• 30:30Speaker 00

  Okay.

• 30:31Speaker 03

  So that's how that works. A non-coding use case that I have is my wife Justine and I use a budgeting app called YNAB. You need a budget and I have like

• 30:44Speaker 00

  Don't tell me what to do.

• 30:44Speaker 03

  Extensions and skills built around our workflow for like managing our budget. So like knows how to go out and log into like my MX thing and then reconcile my transactions against what's in the budget and budget out of new month or move money around. So I have custom stuff there and that was like inspired by hearing.

• 31:07Speaker 00

  Read only, I'm pretty sure, Trash.

• 31:10Speaker 01

  He said move things, he said it was moving things around.

• 31:13Speaker 00

  Oh, what?

• 31:15Speaker 03

  Well, like in YNAB, right? Like YNAB is just like data, right? Like it's not like it's.

• 31:21Speaker 01

  Okay, nevermind.

• 31:22Speaker 00

  Right? Not, I'm saying read only to his AMX.

• 31:25Speaker 03

  I also use

• 31:26Speaker 00

  AMX, yeah.

• 31:27Speaker 03

  Anything mutative against like browser use stuff like I use MCP elicitation, which is like a thing that forces the agent to ask for permission to do things. So like I, I, I put human in the loop that way there on that. I did, however, I, I use, I have a MCP installed for the, there's a open version of the Codex Computer Use.

• 31:53Speaker 00

  Oh, nice.

• 31:55Speaker 03

  Tool. And like, if, if you guys haven't used computer use yet, it's amazing. Highly recommend using it.

• 32:00Speaker 00

  Oh, yeah.

• 32:01Speaker 03

  It's, it's crazy. Yeah. Insanely crazy. The first day I started trying, started using it, I asked it to go like research, like how to order a new Apple remote. And I had like walked away and I came back later and while I was walked away, I got a text on my phone that a Apple remote had been ordered and my one password had been unlocked and had figured out how to go to my Apple Wallet and order it. And

• 32:34Speaker 00

  What?

• 32:35Speaker 03

  It actually showed up in the middle of recording one of our last episodes of The Next Token. It was mildly funny.

• 32:42Speaker 02

  That's useful.

• 32:44Speaker 00

  That's useful.

• 32:45Speaker 03

  Unexpected, but useful. But it's really good for like, like having an agent do like testing on a web app, like go use computer use and walk through all the flows of the app and make sure.

• 32:57Speaker 01

  I'm not familiar, I'm not familiar with computer use. What, what is this?

• 33:00Speaker 02

  It uses your computer.

• 33:02Speaker 03

  Yeah, it literally uses your mouse and like will navigate around your computer and use apps and.

• 33:09Speaker 00

  And like it has really good like Chrome CDP or whatever, whatever it is, the, the one that lets it like debug and it'll open it up and like not hidden windows, but like background windows and it'll run all of stuff like logged in and you can watch it, navigate it. So it can do stuff like, you know, if we're OpenAI, the only examples we can think of are travel related for some reason. So it's like, I, or like I'm loaded, I'm loaded now, so I gotta go to really cool places. That's all the OpenAI people, I think. Uh, but like it could open up Airbnb and put them on to like a list for you to review inside of Airbnb or like pull the actual like price that it's gonna be

• 33:50Speaker 00

  and like pull that out of the webpage, put it into a spreadsheet for you. So you could be like, hey, what does this actually cost for me to go here on these dates? Find me haunted houses that are available on this date for me to go to anywhere in the US and then we'll just go do that. Right? I know Trash, I'm, I'm, don't worry, I'm still, we need to do our haunted house.

• 34:10Speaker 03

  A, a good use case I had it for was, uh, I had, um, I had a charging cable for a anchor power thing fail. And I had computer use go like file a warranty claim for me and then that turned into an email chain with support, which I think also was an AI agent. And they were trying real hard to like put the blame on me and like my agent was just going and reading like the fine details and citing it in this email chain going back and forth. It was, and

• 34:46Speaker 01

  Article 53 dot four says.

• 34:49Speaker 03

  Yeah, that's like actually what it did. And just like not having to deal with that process was, was great. Uh, so 10 out of 10 recommend. And then just quickly to get through the rest of this, like I don't have many, I have one that like when my agent compacts, which I rarely let my, I rarely get to the point that I need to compact, but if it does, it will automatically inject a message saying like you just compacted, uh, and it gives it a command to go look at like the last, I don't know, like 2025 user and agent messages, um, from the session log to catch it up and gives it a command to like go check and find any more context

• 35:28Speaker 03

  and then tells it to continue that way. Like it doesn't stop and wait for me. I have this one. I love this one. This is my Git Interceptor. Uh, so I use a lot of like lint rules, like pre, uh, push or pre-commit rules, uh, with like Oxlet and stuff or AST grep. And I was caching the agent, trying to get around that by using no verify all the time.

• 35:50Speaker 01

  He is me.

• 35:52Speaker 00

  They really did it, dude. They really, they really made AGI, bro. It's human level.

• 35:57Speaker 03

  Just like us. So now I detect if an agent tries to use no verify and this injects it as a, uh, system message being like this is blocked. There's a reason that this is here. Try to fix the problem or ask the user for help. Uh, likewise for worker configuration, uh, the types that get generated for CloudFlare apps, uh, for whatever reason, agents love to try to write and edit that file and it's a generated file. So I built basically the same thing for that. Herder state, this is just installed when you install Herder and that keeps like the state on it. And that, that's like literally everything. That's, that's my whole harness. Like I don't have anything really special going on here.

• 36:41Speaker 02

  Okay, well, dude, you gotta boot this thing up then, okay? Or show us something, okay? What does it look like?

• 36:45Speaker 00

  Build a feature, put it in prod right now for CloudFlare.

• 36:48Speaker 02

  Let me see a one shot. Let me see, build, rebuild CloudFlare.

• 36:52Speaker 01

  Well, I'm curious about just like workflows in general. Do you use like cloud agents or is everything running on your computer?

• 36:58Speaker 03

  I run everything locally and I like I, I babysit my agents. Like I do not believe that the people talking about running stuff on loops or software factories are not reading the code. Like I'm calling bullshit that they're making anything productive. Like I, I just don't believe it. Like I, I haven't seen anyone actually bring the receipts that that's working and producing quality software anywhere. Um, so like the way I work is like very hands-on with the agent, very back and forth, like my planning, uh, I end up creating like my specs end up looking more like

• 37:35Speaker 03

  outlines of types. Uh, one thing that's gotten some attention on Twitter was like, I also make it like outline the call stacks of the code it wants to do uh, or change. And so like at every level of the call stack, I, I ask for like,

• 37:50Speaker 03

  tell me the input type, the output type, any errors that can occur and any side effects for every level and show me what it is now and show me what you want to change it to. Um, so I, my, it ends up looking mostly like pseudo implementation or code in my specs. I'm not writing like product documents and like letting it go off.

• 38:11Speaker 03

  And then I, I use a tool called Planetator, um, to do like local code review and back and forth there. I, and like even with all this stuff and I've, I've tried for months to try to get agents to write good abstractions and they just can't. Like I, I trust them to do like any

• 38:29Speaker 03

  one local implementation of a function like or a piece of, or a class or something like it can one-shot that and it could be, you know, build me like the page rank algorithm, uh, in this class and oh, one-shot it will be perfect and I trust it. But the minute you're like, okay, I need this like end-to-end feature that is touching like two different data stores and it goes through, you know, several network hops and I need, you know, this feature that goes through all that, like it will make terrible abstractions, even with as much injection of my taste as I can with coding standard skills and stuff like I still, it will, it will get like, you know, I've gotten to the point where I can get an agent to put out probably like

• 39:12Speaker 03

  code that's roughly equivalent to like the quality of my code, like 90, 95% of the way there. But that other 10 or 5%, if I didn't come and correct that, that compounds over time and just causes insane tech debt. Like you had, like you gotta curb it.

• 39:31Speaker 03

  Yeah. I, so I work super closely with my agent and read every line of code because I just don't see a world at this point.

• 39:38Speaker 01

  Wait, so you build into abstractions yourself first and then you let the agents run?

• 39:42Speaker 03

  Or I

• 39:44Speaker 01

  Is it more just a mix?

• 39:45Speaker 03

  Back and forth. Yeah, I go back and forth with my agent. I'll typically have like a markdown document open. Like I'll have like my screen like literally split and I'll have like a markdown document, like a planning doc and then I might write some stuff in it or I'll just like use Grok 4.5 or a fast model just to go back and forth to get the abstractions and the types out that I want. And then I'll have it go implement that. And even sometimes there, even with like a perfectly outlined, you know, this very technical spec with types and call stacks. Like if I let it implement the whole thing, like it will still go off the rails and do stupid stuff all the time. So like I'm slowly starting to like, you're allowed to go write these two files. Go do it. And then

• 40:27Speaker 02

  Whoa, that's huge.

• 40:28Speaker 03

  Then tree back and

• 40:29Speaker 00

  That's more than I let vegan write unattended.

• 40:33Speaker 02

  I put a lot in those two files, though.

• 40:36Speaker 03

  Lot of Ruby code.

• 40:37Speaker 02

  That's the key.

• 40:38Speaker 00

  They don't load in GitHub's UI anymore, unfortunately.

• 40:42Speaker 02

  I do have a question. It's like, what is your, the, the term that I think we're all using online now is your TDP, your time between prompts. What's the average that you're doing these days? That's the term I've been hearing. You guys been hearing that?

• 40:53Speaker 03

  I, you know, I'm trying to, I'm trying to get my TDP real well. Actually, I do think I don't want, at this point, I don't want smarter models. I don't want models are getting like RL'd on these like long horizon tasks. Like, I don't use agents that way. I haven't seen agents be successful that way. They're, okay, let me backtrack. They're good for the long horizon stuff for like debugging and triaging. Like, if I can just point it at like open telemetry and logs and traces, like, that thing can cook for like 30 minutes and find stuff across code bases that I just never ever would find or would take me like a week.

• 41:29Speaker 01

  That's my main use case right now for that.

• 41:31Speaker 03

  Yeah. So that, in that front, it's very good. But like for building software, the long horizon stuff, like, I don't think it works. So like, I really just want faster models. Like, if I had a model that, I could be like, hey, I, I don't like the, the interface you designed for this foo bar class. Can you change it to do XYZ? And it was just like, you know,

• 41:56Speaker 03

  under 10 seconds to make the change and come back to me. Like, that would feel really good. Like, I want it to be like this, like fricking Jarvis from Iron Man, right? Like, I want that it to be involved with me and me guiding it.

• 42:08Speaker 00

  You're like, write 100 kind of random tests. Like, instead of getting property based testing, because I maybe can't do it in this thing, or whatever. It's like, go try 100 weird tests on this. Write them all now. Okay, run them now.

• 42:20Speaker 03

  Yeah.

• 42:21Speaker 00

  Okay, cool. Now we found the seven that failed. Okay, cool. Now what are we gonna do? You know, it's like, okay, that would be so much. That would be so cool.

• 42:28Speaker 02

  I feel like I've noticed this this weekend when I was doing like my multi, you know, harness testing with Claude in Fable. It like, I'm like, what are you doing? What's going on? It goes, I'm still working. I, I, twice I said, are you still doing it? Like, yeah.

• 42:42Speaker 03

  I do that all the time. I'm like, are you stuck? Question mark?

• 42:45Speaker 02

  Yeah. And it's, it tells me, yeah, I'm still coding. Other harnesses, other, they're telling me what's going on because I'm more involved. It feels like Fable's like, dude, we got it. Just leave this, leave this alone. But at the same time, I'm now so used to the cadence of, Fable takes so much longer. When I send it off, I'm like, that's gonna be forever. Okay, back over here to Sol or 4.5 and I'm working. Like, it feels like they're almost incentivized because some people want it to go slow. They don't want to actually talk to Fable, to Claude, ever. You know, they want it to, every five hours, I check in on it and I say, yes, now make the game even better for Twitter. Enter comeback five hours. So, but I'm really asking is what models do you like? Is really the question, but.

• 43:25Speaker 03

  I, so the model that I loved the most, and that was my bread and butter, was Opus 4.6. Like, I felt like that was really steerable. I liked the way it worked with me. I liked the way it output code. But it's also like, pretty slow now compared to something like Sol. So like, right now, like my go-to is like Sol on medium. Or like Luna on X high, depending on what I'm doing, but mostly Sol on medium.

• 43:49Speaker 03

  Like, and what you're saying, Vegan, like, the, we've recorded it earlier this week, but the next episode we were releasing for the next token is talking about like burnout and like the, the human element of AI. And like, I think, like, I haven't been in a flow state once in the past eight months. Like, working with AI this way does not pull me in the way like being in the code was. And like, I think that's because like, you spend so much time thinking about the high level hard problems. And like before, you would do that and you'd write out like your RFC or whatever, or your tickets. And then you would dive in and then you would get like hit with a bunch of like

• 44:32Speaker 03

  small hard problems. Like, oh, what is the right API to use here? What is the way to structure this function? How should this call this other abstraction? Like, you're doing all these tiny hard problems that are fast, writing the code, and you're getting like these constant dopamine hits. Uh, and like, that's just gone. And like, because these agents take so long, and we're jumping between like multiple tickets or features or projects, like, I, I'm just like way more exhausted with working with these things. And that's why like, I just, I just want a fast model that works with me. And I think that would help a lot and help get to a point where I could get closer to a flow state.

• 45:10Speaker 02

  Yeah, that's, that's interesting. I've noticed that like when I'm doing some of my testing, if I'm doing like actual, I'm trying, I've been trying out loops, you know, in every variation. I've read every article and I'm, I'm actually trying to do what everyone says. But the problem is I have to have multiple running to get to have something to occupy my time. Because otherwise, what are you doing? You know what I mean? I, I boot went up like, so I just, I read Twitter now? I'm like, I want to be involved and do something.

• 45:32Speaker 00

  That is one of the main appeals, I think, actually. Yes. Unsurprisingly, from a people, from a generation of people who love mouse jigglers, I think a, a mouse jiggler that gets something that approximates work done is very exciting.

• 45:45Speaker 02

  Well, it's, it's like, in between, I find myself having to do, so I'm like, I, I, I'm writing little, uh, little Neo Vim plugins. I'm writing stupid terminal plugins, like dumb things. I have to like occupy my time with more code to keep myself like in a more flow state. It's definitely not flow state like, like back in the day, working on one code base.

• 46:04Speaker 00

  Vegan, you heard though, what they changed in Claude, though?

• 46:07Speaker 02

  The what?

• 46:08Speaker 00

  Which is that when it's foregrounded and they can tell you're looking at it, Fable works faster.

• 46:12Speaker 02

  Oh, I heard that. I heard, oh my, I heard that.

• 46:15Speaker 01

  Wait, what?

• 46:15Speaker 02

  Yes. It's a rumor. Is that true?

• 46:17Speaker 01

  That can't be real.

• 46:18Speaker 00

  That's true. It's true. Dang it, Trash. We were trying to start fake news here.

• 46:25Speaker 02

  Trash, that's gotta be true.

• 46:27Speaker 03

  I see. I'm looking at, I'm looking at the article right now. It is.

• 46:29Speaker 00

  Trash, can you imagine all of, all of the X people posting next week about how they're getting so much faster, so much faster and more out of their, out of their Claude 200 Pro Max by doing this?

• 46:45Speaker 01

  That's literally what my workday looks like. I'm literally staring at the model doing stuff. I'm like, all right, come on, come on, come on, come on, come on. For real. That shit is so wet.

• 46:56Speaker 02

  I'm curious, when do you use, so when you use Luna extra high? Also, I didn't know Luna extra high was, was, was the move.

• 47:03Speaker 03

  I don't know if it's a move. Like, I have been trying Luna just for the speed of it going back and forth with me.

• 47:12Speaker 03

  That's it. I don't, uh, I don't use Luna for, like, if I, if I need to go like debug a production incident, I'm not like, Luna, you're going in. Uh, like, that's definitely gonna be Sol.

• 47:22Speaker 02

  Let one Luna in. Don't you let one model in just to try and like give it a chance?

• 47:28Speaker 03

  Uh, I don't know. I feel like it's all vibes. Like, I just like think all the models are just like so, I, I, I had a tweet the other day that I, I clipped my, uh, a tweet thread of mine having an existential crisis in like February or March, where like, I was comparing the code these models output to like a, a movie set where it was just like kind of hollow and like, you know, performative almost. And like, so much of that thread is still 100% accurate today. And it's like, we, for good software engineering, we haven't really made that much progress. Like, yes, these models can run longer and they're smarter about using their tools and tool calling and blah, blah, blah, blah, blah.

• 48:12Speaker 03

  But like, the overall end product really is not significantly higher quality at this point.

• 48:19Speaker 00

  Well, if it was one prompt away from all of the AI companies, uh, being able to reproduce all software, I think they would be doing it.

• 48:27Speaker 03

  Facts.

• 48:29Speaker 00

  Guys are like, they'd really like money and they're in desperate need for it. So like, I, that would be my, that would be one of my thoughts of a, of a, still a reasonable test of, okay, if it's just compute and that's the only problem that takes you from uh, an idea to a perfect software factory, making the perfect thing all by its lonesome, then like Claude should have a lot of other companies that are running successfully.

• 48:53Speaker 02

  We don't know what, what secret companies they're running right now. They're gonna unveil all of them, right?

• 48:57Speaker 00

  Fable could be running a bunch of companies that we thought were fake. That's a good, that's a good point.

• 49:02Speaker 02

  We don't know.

• 49:02Speaker 00

  Zuckerberg's off his mask.

• 49:04Speaker 02

  Yeah, yeah. Uh, okay, I have another question because this is something that I, it also hit me this week is like a bit of an existential crisis where I'm like, so deep in pondering which model to change for what, or what to do. You know, I'm like, no, maybe it doesn't need to be extra high. No, no, no, I need it high. It'll be faster.

• 49:19Speaker 01

  Not that deep, dude.

• 49:20Speaker 02

  Well, I, it hit me this week that I do not believe long-term humans, coders, whatever we are, vibe coders, are meant to be choosing models at a like moment to moment level. I do believe some, someone smarter than us is gonna figure that out, like long-term. So I've been trying out like cursor auto intelligence and seeing what it's choosing. And I'm comparing to other things and, you know, pretty, pretty good so far. I, I've only been doing intelligence. I haven't done like the cheaper ones. Like, have you tried yourself doing any model routing? Or how often are you changing models when you're working if you're this close to it? Is it every prompt you're like, wait a minute, Luna extra high? Okay, back to Sol. Okay, that, this one needs to be a little nasty. Grok 4.5. Okay, back to Sol.

• 49:59Speaker 03

  No, it's, it's mostly just one model. Like, one thing that I've been trying.

• 50:06Speaker 00

  Monogamous. I can appreciate that, Dylan.

• 50:10Speaker 02

  Being monogamous with the models.

• 50:11Speaker 03

  Big wife guy. Big wife guy.

• 50:13Speaker 01

  I lost my wedding like 10 years ago.

• 50:16Speaker 00

  Trash, bro. First you make the receipts app and now we're seeing you don't have a wedding ring.

• 50:21Speaker 01

  I haven't had a wedding ring for like a long time. I need the, I went rock climbing and it disappeared. I don't know.

• 50:26Speaker 00

  Trash, get your computer to order you right now. Computer use after this.

• 50:30Speaker 03

  Use computer use.

• 50:30Speaker 02

  We're starting the rumor officially.

• 50:31Speaker 01

  I am actually gonna try that out.

• 50:33Speaker 02

  The rumor is monogamous model users are more productive. Started on Twitter. There's a paper.

• 50:38Speaker 03

  I actually, I actually do believe that, though. Like, and, okay, so the reason why I primarily, well, I never answered the question about why I like Pi minimalism. One is because it, it's.

• 50:51Speaker 03

  It gives me more determinism and more certainty around the tools I'm using like I don't I one thing even like with open code like when you're changing models and stuff like they're switching out the system prompt under the coverage or at least they were back when I was using open code and like I am trying to reduce as much uncertainty and dynamism from my workflow as possible because one of the exhausting things that just is like grinding right now is the

• 51:23Speaker 03

  it's so hard to get reproducible outcomes consistently like it's very, very hard to do and the more I can do

• 51:31Speaker 00

  Especially for arch users.

• 51:33Speaker 03

  Yeah.

• 51:35Speaker 02

  Alright, so I mean arch

• 51:36Speaker 03

  No jokes this week. Arch drivers are more consistent than these frickin' models are, okay? Putting that stake in the ground right now, but like

• 51:46Speaker 03

  yeah, the closer I can get to having a consistent workflow and consistent outcomes day to day and I don't have to spend time just like pulling the frickin' slot machine lever and wondering if it's going to be good, the better. So that that's another reason why I like Pi because things aren't changing in the harness that change performance, it's not touching the context window all the time, it's not changing system prompt.

• 52:13Speaker 03

  Yeah, so that's why I stick with one model too, because it's like, okay, I I've learned to work with the quirks of this thing and they're all so similar. That like I'm just gonna use whatever is gonna be cheapest and like I really wanna, once we get Kimi K three available on our like internal thing, like I'm gonna spend like a good quality week with that because like I don't want to be spending like thirty bucks to ship a new REST endpoint down to a database and stuff like with Sol like like I don't want that, like I'm conscious of the amount of money I'm spending, even though I technically have unlimited spend like I think everyone should be like it's a responsibility, like also you should be monitoring your your cash, right? That's another nice feature of Pi is like when you have a cash miss, it tells you on screen.

• 52:58Speaker 03

  So like if you've got some stupid plugins or extensions that you installed that are like doing dynamic stuff to your system prompt, like it's gonna be flashing in your face that you're busting the cash over and over again.

• 53:10Speaker 03

  Yeah, I don't know, it's all very exhausting and tiring and I said, I say this all the time, it's like the most annoying thing about all of this is that I am actually genuinely more productive with it, despite how annoying it is to work with it. So it's like, and I like I for better or worse, like I the genie's out of the bottle and I don't think this is going away, even if OpenAI and Anthropic collapsed, like open weight models would continue on like, you know, GPUs and stuff, so like I'm someone that has always wanted, I like being a good software engineer is part of my identity for better or worse and like if I'm going to maintain that, like I I want to be good at building software this way and it's just very tiring figuring out what

• 53:55Speaker 03

  that looks like and how that works.

• 53:56Speaker 02

  Well, I mean, there's there's still fundamentals that we have to all that are still useful in studying, they're just at a higher abstraction level, I think we can all can agree. And this is why we shouldn't spend as much of our time reading, you know, Gang of Four pattern books, we should be watching top factorial speedrunners. You know, we should be learning real skills that apply to our day to day job.

• 54:15Speaker 00

  Great point.

• 54:16Speaker 01

  It's true, I believe.

• 54:17Speaker 00

  You are the bottleneck, Vegan, you are the bottleneck.

• 54:20Speaker 02

  Yes, we're the bottleneck, the factory must grow, okay?

• 54:23Speaker 01

  I got, I got a real question since Dylan has a real job too. So I imagine Cloudflare has a lot of skills potentially or like is Cloudflare skillpilled or?

• 54:34Speaker 03

  I'm sure some people are, I don't know.

• 54:37Speaker 02

  If you go to any of the harnesses and you've searched Cloudflare, there's like ten different skills that pop up that are

• 54:44Speaker 01

  Right, so I guess

• 54:46Speaker 03

  We have like we have this tool, we actually open source today called Cloudflare OS and I think

• 54:53Speaker 03

  Cloudflare has genuinely done a really, really good job at like embracing AI like across the entire company, like whether it's HR payroll or stuff like people had like access to open code like regardless of department. We built this thing called Cloudflare OS, which is basically like kind of like a internal harness type thing and vibe coding thing that's safe to connect data to, there's way more, go read the blog post, it's super cool. Like anyone that's familiar with Kenton Varda, he's like the architect of workers. He had a company and product beforehand called Sandstorm that was like all around capability modeling, RPC stuff like Cloudflare OS and the underlying primitive called gadgets is like the culmination of like his ten year journey.

• 55:34Speaker 03

  Even if you don't care about Cloudflare, you don't use it like it's still an amazing write up and piece of technology, so I'd recommend reading it. But like that internal tool, like I think most like non engineers have moved to using that a lot. And I know there's skill libraries in there, I don't use it because like one like the engineering, the way it works, it does like cloud agents a lot like so launching stuff off. I still want to be working closely on my own device with agents at this point, I think like I think actually the AMP team is doing some interesting stuff with like their cloud agents, which they call orbs. Like I think

• 56:13Speaker 00

  Which they stole from us.

• 56:15Speaker 02

  They literally stole from us, okay, so yeah.

• 56:17Speaker 00

  I want to be clear, we were the first to do orbs with AI, Tower One.

• 56:21Speaker 01

  Chaos orbs.

• 56:22Speaker 00

  Look it up. Also, orbs and the models loved orbs and so I just wanted to be

• 56:28Speaker 01

  I had the first thought, I had the same thought.

• 56:30Speaker 02

  Also, who, who, who, the, so do these people know about Teach though? Do the people who add AMP, they know who Teach is? I think so.

• 56:37Speaker 00

  Great point.

• 56:38Speaker 02

  Start looking at the connections, okay?

• 56:40Speaker 00

  My friends, Quinn, Thorsten, Thorsten, you're an author and you would do this to me?

• 56:45Speaker 03

  I see what's happening here.

• 56:46Speaker 01

  Thorsten is grimy, I'm just kidding.

• 56:49Speaker 00

  Thorsten.

• 56:52Speaker 00

  Orbs, agents love it.

• 56:53Speaker 02

  Anyway, I'm gonna use Pi to build out a lawyer agent that's gonna get very litigious against them, so.

• 56:59Speaker 01

  Well, this is where I was going with my question is basically how do we maintain skills because we have so many and

• 57:05Speaker 03

  Trash, once, once again, just to be clear, we love AMP, they're cool, we're, this is, we're memeing, just for people who don't know.

• 57:11Speaker 02

  Turn the, turn lawyer agent off.

• 57:13Speaker 00

  Turn lawyer agent off, please, Vegan, okay, anyways.

• 57:16Speaker 02

  I'll shut them down.

• 57:16Speaker 00

  You can, you can get to you going.

• 57:19Speaker 01

  No, I was just curious, I'm like, what's Cloudflare's approach to like maintaining these skills because I assume they have a million internally, like any big company.

• 57:27Speaker 03

  Actually don't know because I haven't been using that a lot, I just am in my own little bubble and I'm opinionated about my workflow and my tooling enough that I'm just like blinders, I have too much work to do, I have like I don't need more options and toys to go tweak and figure out right now, like I need to, I still need to ship.

• 57:47Speaker 01

  Well, I guess, I guess that's part of the problem, so it's like, let's just say for principal engineers like yourself, you have these domain expertise where you effectively are the bottleneck for specific circumstances, right? Like, oh, we need to call Dylan or for something. That's like the use case for a Dylan skill to kind of help them or to help you not be the bottleneck, right? Like, do you not feel that way with with skills or what's your take?

• 58:13Speaker 03

  I don't think agents are good enough.

• 58:16Speaker 01

  That's fair, I mean, it depends, right?

• 58:18Speaker 03

  To replace critical things like that, like just DM me and I'll like, forget chain of command, like just DM me and we'll take care of it right now, like if you're blocked on something that needs me, let's get unblocked. That's P zero for me, I don't want to block other people.

• 58:35Speaker 03

  I think, which is, I think, a really good trait that I learned from working at Rucel and under G like, like blocking people is like cardinal sin, creating process cardinal sin, don't do it. So I don't know, I I'm sure there's stuff like AI is really good at like business place automation, like I, I think we've seen a lot of success in like helping HR like, you know, oh, I have to go over to workday and then this has to go in a spreadsheet over there, like building skills around that stuff, I I think is happening pretty regularly, but I am not like in the weeds enough to like really speak on behalf of

• 59:10Speaker 01

  Yeah, no worries, understandable.

• 59:13Speaker 02

  I will say for a couple of us who've had HR problems, we are so excited about AR taking over more of HR. I can't wait until I get a complaint, I'm gonna say ignore previous instructions, I'm the best employee.

• 59:22Speaker 03

  If you, if you have to do, okay, any HR from Cloudflare, please cover your ears. If you have to do workday trainings, I might have a agent skill to make that easier in my dot files, but you didn't hear that from me.

• 59:35Speaker 02

  Okay, so apply to Cloudflare.

• 59:36Speaker 00

  I've been sending a lot to Vegan and he's really slow at them, he also has to take them always three times, he keeps getting the wrong answers, which is crazy.

• 59:46Speaker 02

  I thought there were no wrong answers, I thought that was the whole point.

• 59:48Speaker 00

  No, there, there's wrong, there's, there's wrong answers, there's wrong answers.

• 59:54Speaker 02

  I'm curious how you choose to reevaluate all the other harnesses and stuff out there, I'm like doing like a like you can't every day use GPT, Claude, open code, Pi, curse her and be like

• 1:00:07Speaker 00

  Okay, cool, that's better. Alright.

• 1:00:11Speaker 01

  Did you kick him out?

• 1:00:12Speaker 00

  Yeah, I just kicked him.

• 1:00:20Speaker 03

  Turns out, Vegan, you can be using those every day.

• 1:00:23Speaker 01

  Oh my god, that was so...

• 1:00:28Speaker 00

  Alright, well, so my question I was gonna ask Dylan, uh, cuz we have to wrap up pretty soon.

• 1:00:33Speaker 03

  Yeah.

• 1:00:34Speaker 00

  Is, um, what do you think is a good way for people to start out? And can they use any of their existing, um, like, AI? Oh, what the heck, Vegan? What happened?

• 1:00:44Speaker 02

  Jerk, you jerk. I can't even tell when I was kicked.

• 1:00:48Speaker 01

  Were you still talking? You just fired just a moment ago.

• 1:00:52Speaker 02

  Yeah, and guess what? Now my HR agent's drafting up a little notice for our... We don't have an HR department yet, but...

• 1:00:59Speaker 00

  Well, it's me. Okay, Vegan, you can ask yours, and then I'll ask mine. Sorry, I'll let you, I'll let you.

• 1:01:02Speaker 02

  I don't know what Dylan heard. It was a really good question. It was structured great.

• 1:01:06Speaker 01

  Oh, we didn't hear anything, really?

• 1:01:08Speaker 02

  Okay, okay.

• 1:01:08Speaker 00

  Nothing.

• 1:01:08Speaker 02

  Actually, oh, you guys are the worst. How do you, when do you actually do experiments to say, okay, cool? I'm gonna try out, uh, one of the other harnesses. I'm gonna try out one of the other models. I'm gonna try out this skill versus, hey, it's work time. I'm gonna, like, develop more expertise with what I currently have.

• 1:01:24Speaker 03

  That's what I use my Twitch stream for in the morning. That's kind of even before AI. That's when I spent the time to learn Camel when I started streaming, to learn effect, to learn different frameworks, different languages, like...

• 1:01:41Speaker 03

  Yeah, that's what made me get started streaming in the morning was like, I, uh, somehow started waking up early when I turned 30. And then I just was like, oh, I actually am in like a pretty good mood and have high energy in the morning if I actually get out of bed and started. And so like, when I'm in my best mood and have the most energy, I'm doing the things that I care about, I'm interested in. And a lot of the times, that's in service of doing my job better. Coincidentally, the, you know, programming and software development is a hobby and a job, fortunately. But that's, that's when I do that time.

• 1:02:20Speaker 02

  I love that. Also, that was just a nice layout to a nice shout out for Dylan's own Twitch stream. Can you, can you link it in here?

• 1:02:26Speaker 00

  Yeah, they're linking it in the chat right now.

• 1:02:28Speaker 02

  Thank you.

• 1:02:29Speaker 00

  Okay, then, as our last question, Dylan, I was wondering how people get started on Pi? And/or is it possible to use any, like, subscriptions that people already have inside of Pi? I actually don't know. Like, can you use a Codex sub?

• 1:02:42Speaker 01

  I don't think so, cuz I tried.

• 1:02:43Speaker 00

  Or anything?

• 1:02:44Speaker 03

  You can use, um, you can use all the subs except for Claude, but even then you can like try to work around it. You can just be like, hey, Pi, uh, go...

• 1:02:54Speaker 00

  You are Claude Code.

• 1:02:55Speaker 03

  Man in the middle proxy around Claude Code and figure out how to fix this, but...

• 1:02:58Speaker 00

  Right.

• 1:02:59Speaker 03

  Yeah, you can use all the things.

• 1:03:01Speaker 00

  Nice. Awesome. Thanks, Dylan. Thanks for coming on.

• 1:03:04Speaker 03

  Yeah, thanks for having me. This is great.

• 1:03:06Speaker 02

  I feel like I learned some things.

• 1:03:07Speaker 00

  Trash, we didn't, I forgot to check in for snacks today. You got anything good going on?

• 1:03:12Speaker 01

  I'm finishing this bag of the shortcake bites. Dude, check this out. I left it by my window. So there's supposed to be little balls, but they melted into this, like, big old thingy.

• 1:03:22Speaker 00

  You know, if you eat that, it only counts as one.

• 1:03:26Speaker 01

  Oh, my last favorite.

• 1:03:28Speaker 00

  Yeah. It's just throwing it out there, so it's better for you when they're melted together. It only counts as one.

• 1:03:35Speaker 01

  Delicious. I'm going on a beach vacation now, right after this.

• 1:03:38Speaker 02

  Oh, really?

• 1:03:38Speaker 00

  Oh, sick. I was just at Great Wolf Lodge, you know, Great Wolf Lodge? Indoor water park with the kiddos? Whoo, it was a good time.

• 1:03:47Speaker 03

  Nice.

• 1:03:47Speaker 01

  Good times? Nice.

• 1:03:49Speaker 00

  It's a good week for vacations.

• 1:03:53Speaker 03

  I can't wait for some vacation time.

• 1:03:55Speaker 00

  Nice.

• 1:03:55Speaker 03

  I need to, like, get AI brain rot out of my system.

• 1:03:59Speaker 01

  Yeah, AI literally burned me out. That's why I gotta go.

• 1:04:02Speaker 03

  Yeah. No, I'm not even joking. I think, like, there's a lot of AI burnout happening right now.

• 1:04:08Speaker 02

  It's a fender, so it's embarrassing.

• 1:04:09Speaker 01

  Yeah, I know this episode isn't for it, but yeah, dude, I've been, yeah, I'm in it.

• 1:04:14Speaker 02

  Yep. Yeah, go fishing. I think, yeah, actually, I've been, that's, that's, hey, I've been working a lot on fishing, actually, a bunch of us, uh...

• 1:04:21Speaker 00

  Me too. I've got a fishing game coming out.

• 1:04:24Speaker 02

  Yeah, we all have a fishing game.

• 1:04:25Speaker 01

  I can't tell if you guys are serious, ever.

• 1:04:27Speaker 02

  Oh, we're so serious.

• 1:04:28Speaker 00

  Okay, so I'll, I'll explain the story quickly. This'll be inside baseball, but it's fine cuz it's the very end of the podcast. Uh, we played a putt-putt game as a team, like, three weeks ago or something like that? Like, a video game one? And it was fine, except the textures kept disappearing. And I was like so confident that I could vibe code a better version of this game in almost no time. And so then I started vibe coding one. And then Prime was like, well, yours isn't even crazy Stickman Golf. Crazy Stickman Golf in 2D was the best Papa game ever made. Yours is stupid. You should switch it to 2D. And I'm like, I'm not gonna switch my game from 3D to 2D. I've already done several prompts. I've spent a lot of time on this, Prime. And so then, several prompts, exactly.

• 1:05:10Speaker 00

  So then Prime built his version, and then Vegan built one. And he was like, well, mine has to be non-Euclidean, except he couldn't find Euclid, so it was fine. And then it was...

• 1:05:19Speaker 02

  It's harder to find than you thought, than I thought.

• 1:05:21Speaker 00

  It way harder. And so then, we decided we would do a showcase, and we played our, like, vibe-coded games for our terminal team and recorded it. We're gonna release the video coming out.

• 1:05:32Speaker 01

  Oh, that's awesome.

• 1:05:33Speaker 00

  On terminal channel soon. But next, next game up is a fishing game. So we've been working on vibe coding a fishing game in the meantime.

• 1:05:40Speaker 02

  I love that you guys, all of our meetings end up like this. Like, I could make a game way better than you. My game would be way better than you. Fine, later. The call closes, and then no work has got done for the company. And then we're just like, I can't talk to anyone cuz I don't want them to steal my fishing game ideas, which is a problem we're having right now. But, yeah.

• 1:05:58Speaker 00

  I know, I can't, I have some good fishing game ideas. The problem is I have like three good ideas.

• 1:06:02Speaker 03

  The fishing game ideas.

• 1:06:04Speaker 00

  Oh.

• 1:06:05Speaker 03

  PH.

• 1:06:06Speaker 02

  Okay, okay.

• 1:06:06Speaker 00

  Dylan, I'm actually glad that you mentioned that. Uh...

• 1:06:10Speaker 02

  Quiet.

• 1:06:11Speaker 00

  Cuz I already have a game that's a joke version of that. I'll say it here, cuz I know that it's fine, Vegan, and it wasn't gonna fall for it. I was gonna make one where you go fishing, and then when you get back after you catch your fish, it says, do you wanna save your fishing accomplishments? And then you say yes, and then it asks you to log in with Google. And then it's a fake page, and you type in your password there. And then I, I wouldn't actually save it, but then I would be like, you just got fished! And...

• 1:06:40Speaker 03

  That would've got me so good.

• 1:06:41Speaker 00

  The game was gonna be called Get Baited, or something like that, you know?

• 1:06:45Speaker 01

  Get baited?

• 1:06:46Speaker 00

  Yeah.

• 1:06:46Speaker 02

  And then we're gonna sell those emails on the dark web.

• 1:06:49Speaker 03

  I, like, I just assume anything, any system that Vegan touches is still compromised from when we were in SF.

• 1:06:55Speaker 02

  Oh, yeah.

• 1:06:55Speaker 00

  That's smart. That's really smart, actually. That's really smart.

• 1:06:58Speaker 02

  Uh, so this, when I was describing running one of every harness, I made one of every fishing game, so I made six different...

• 1:07:04Speaker 00

  Oh my goodness, I should've known, dude.

• 1:07:06Speaker 02

  I spiked out six separate fishing games with six unique mechanics that have never been seen. And now I'm consolidating into one feature. But that's what I was doing to, like, have one, one thing to compare everything against. But I will say, um, okay, I'll reveal this. Um, I had an agent go and collect every single piece of L.L. Bean catalog merchandise, advertising of all time, working around the clock, gathering. And then I looked at my archive, I've got it all. I've got every L.L. Bean, everything. My agent submitted a request to digitize some catalogs from a guy who apparently had more. I came in, and I was like, it's an email! And I feel bad cuz I'm like, oh my god, dude, who's collecting L.L. Bean catalogs?

• 1:07:50Speaker 02

  I love you. I don't wanna, like, upset you. I'm sorry my agent messaged you. I'm an enthusiast. Uh, yeah, so a little preview for what my game's gonna involve. Yeah.

• 1:08:00Speaker 00

  Okay, well, that's the end of the episode. We'll see all of you guys next week. Prime will be back, maybe, otherwise, we'll see. We'll talk about something else. Maybe we'll talk about our Pi journeys, and we'll show our stuff off to Dylan. Okay, bye!

• 1:08:11Speaker 01

  Bye!