// transcript — 1825 segments
0:00 Introduction to the live stream and guest Ray.
0:02 before where I've done a live stream with another person. So, I'm really
0:06 excited for this. I'm like waiting on my other monitor here to make sure that uh
0:10 things are actually starting because I'm not starting the stream from my own
0:14 device. I've never done that before either. Um but yeah. Okay, it's actually
0:19 looking great. Perfect. All right, we're good to go. >> Cool.
0:22 >> All right. >> All right. Hey everyone, welcome to our
0:29 next live stream here. Um, Rey is a I guess I maybe I'm more of the guest on
0:32 his live stream than the other way around, but I've never done a live
0:35 stream before with another content creator before. I'm really really
0:39 looking forward to this. Um, Rey, he does a lot of things around AI coding
0:42 like me experimenting with different workflows. Uh, he worked for Apple for
0:47 what was it 12 years, Rey? Right. And I'll I'll let you introduce yourself
0:51 more in a little bit as well. But uh yeah, I'm just looking forward to this
0:54 because we're going to be synthesizing ideas together around AI coding
0:58 workflows for this stream. And specifically what we're going to be
0:59 Explaining the main topic: Anthropic's harness for long-running agents.
1:02 focusing on is um the anthropics new harness for longunning agents. So it's
1:06 something that I created content for on my channel last week where I showed what
1:11 it looked like to have Claude through the Claude agent SDK running for 24
1:17 hours straight coding an application. And I've taken it even further. So, I
1:21 built my own kind of flavor on top of it, inspired by some things that Ray's
1:25 been cooking up with his workflows as well. So, we're going to show that
1:29 today. And um the main thing that like I want to show you guys in the stream,
1:33 what Ray and I want to show you is just how you can think about these
1:37 principles. Like, sure, Anthropic built this really cool harness for
1:40 long-running agents that you can go and use yourself right now. But more
1:44 importantly, I want you to like learn with us from this harness. How can we
1:49 take these ideas and apply them to our own workflows and systems for AI coding
1:54 to make them more powerful and able to do things for a long extended period of
1:59 time? Not that you always want to let your coding assistant rip through
2:03 something for 24 hours, but it's cool that we can do that. And I think that
2:07 more and more as we build our systems and have more and more powerful LLMs and
2:11 tools, we're going to get to the point where we'll be comfortable even if it's
2:14 like creating a proof of concept, but letting it do that for like 24 hours
2:18 straight. And so I really do see this as the future of Agentic Engineering. And
2:22 so that's what we get to explore together today. And um the application
2:28 in particular that I have integrated on top of this harness that we're going to
2:33 be focusing a lot today is linear. So, for those of you who don't know Linear,
2:37 it is a task management platform kind of similar to like ASA or Jira. Um, and
2:43 it's something that Rey has used a lot in the past. Personally, I've not used
2:47 it before, but I wanted to add it in for this live stream because I think it's
2:51 just like a good mesh of what Rey and I are working on um in tandem and like now
2:55 bringing it together to you. And so, yeah, it's going to be really cool. Uh,
2:59 I got some things to like showcase ahead of time because I already tested some
3:02 stuff before this stream. So, we'll get into that and then we'll do a live
3:07 coding session where we'll kick off the assistant. We'll have it like create all
3:11 these tasks in a linear and start knocking them off one by one and I'll
3:15 explain the principles behind it as well, just so that you get like a really
3:18 clear grasp of what's going on. And there's just a lot of really good
3:21 takeaways from this harness that Anthropic has created. And so, yeah,
3:25 that's enough of me talking here. I want to hand it over to Ray here to introduce
3:29 himself and uh and share more about like what he's got in store for us for the
3:32 Ray introduces himself and his AI coding background.
3:32 live stream as well. >> Yeah, I'm super excited to be here. This
3:35 is our first time kind of doing a live stream totally on Riverside and stuff.
3:38 So, uh if you guys have any technical difficulties or anything, just drop the
3:41 comments right now letting us know uh how the sound is doing and everything
3:44 like that. We're just kind of making sure um that everything's going and
3:47 streaming to you guys, okay? I would imagine everything is going pretty good.
3:51 And so yeah, for those who don't know, I used to be an engineer at Apple for um
3:55 like over 12 years and stuff. And now I've been playing around with AI agents.
3:58 So for those who aren't familiar, I do a lot of live streaming where I'm building
4:01 live. So I'd like to build my apps live, learn on the fly with you. So usually
4:05 when the latest models come out, I immediately start using them to ship
4:09 features. And doing that live is how I've learned to do prompting and AI
4:12 coding. And so we feel like this is a great space for everyone to kind of hang
4:16 out and ask some questions. And this is probably a really good time if you've
4:19 been kind of thinking about, okay, should I start trying this stuff out?
4:24 There's something that um is happening right now in AI coding where people are
4:28 able to run these things for longer and longer periods of time. Uh and
4:32 Anthropics blog basically is showcasing that it's going to be able to run it
4:35 for, you know, 24 hours or something like that. So, can we harness that power
4:40 and do that? And I'm I'm excited to talk to Cole because he's pretty much at the
4:44 forefront of doing this type of stuff in his communities and there's so many
4:47 people who have hit me up saying, "Yo, you got to check out what he's doing,
4:50 man. This guy's like really going all in." So, this is going to be nice cuz as
4:54 an engineer, I like want to ask more questions. I'm super curious and I want
4:58 to kind of go down this rabbit hole with you, Cole. And yeah, I want to see if we
5:01 can kind of kick stuff off. I know there's a little bit of sync issues with
5:05 my audio. People have actually commented that I look like AI. Uh that's something
5:08 that I'm trying to resolve right now with Riverside. Uh because of the frames
5:13 per second, my my output is 60 frames per second on my camera and the uh
5:18 Riverside I think cuts it down to 24 if I'm not premium. So I need to figure out
5:22 what that issue is. But aside from that, I think everything should be okay. Uh so
5:25 Cole, I want to go ahead and hand it off to you. See if you can kind of share the
5:28 screen and um take us through this type of thing. What are we going to be
5:30 building today? What's it what's it kind of look like so far? Uh and then kind of
5:33 work ourselves backwards. Also, Tech Friend, huge shout out to Tech Friend
5:36 kind of coming on the stream here. Hopefully, I can I can show the
5:39 comments. Yes. Okay, cool. Yay. It's so fun to actually show the comments. Uh,
5:43 so this is kind of answering that question about the audio sync. So, if
5:45 you have any comments, let them know. We're going to go ahead and just drop
5:48 them in here. And um, excited to be building with you, man. This is so cool.
5:51 >> Sweet. Yeah, I have been looking forward to this all week. So, um, yeah, Ray and
5:55 I have been talking about doing this for I think like a couple of weeks now. And,
5:59 uh, shout out to Scott. He's the one who introduced us initially. Um, so yeah, he
6:04 follows both of our channels a lot and he's here in the Dynamus community um as
6:08 well. And so yeah, he's like, you know what, these guys should do a live stream
6:13 together. And so he CCed both of us in an email and now we're off to the races
6:17 and and here we are today. So I've been looking forward to this. Um, and even to
6:22 the point where like I created something specifically because of Ry and that
6:25 that's the linear integration I was talking about. So yeah, I'll go ahead
6:29 and share my screen and then we can dive into things here. Um, someone in my
6:32 chat, Ray, said that my volume is a little low, which here, I'll just kind
6:36 of move my get my microphone a little bit closer to my face here. Okay. Um, so
6:39 yeah, when you guys see me looking over at my other monitor, it's because I'm
6:42 looking at the chat quick, which I try not to let that distract me, but also I
6:46 do want to make sure that I answer questions as they come up. So, yeah. All
6:51 right, let me go ahead and share. I moved my microphone as close to my face
6:56 as I can. Um, so yeah, I think that should be good. Okay, >> so share my screen.
7:02 All right, cool. How does that look for you, Ray? >> Yeah, let's see if I can Yeah, here we
7:06 go. Perfect. >> All right, sweet. So, yeah, the article
7:12 from Anthropic that inspired everything today is this one right here. Um, so I'm
7:17 going to go ahead and put this in the chat for my stream here. Also, if you
7:21 just Google Anthropic effective harness for longrunning agents, you'll find it
7:25 as the the first result in Google or whatever. Um but yeah, I would highly
7:30 recommend reading through this article if you are really really curious um want
7:36 to dive into things because obviously that's what I did right when I first
7:40 read this article. Um typically I I skim through things initially because there's
7:44 just so much information these days especially in the AI space and so I
7:48 skimmed through this um but yeah it caught my attention. So I I sat down I
7:52 read the entire thing. It comes with this whole quick start that they have as
7:57 an open source GitHub repo. And um yeah, it inspired me to run this thing myself
8:02 and then make some additions on top of it that I'm going to be showcasing uh
8:07 with Ray here in our stream. And so what they've built in this article here, like
8:12 what they're showcasing, I want to explain how it works today. So, I've got
8:13 Explaining the core problem: agents and long context windows.
8:17 this Excal diagram that I also showed in my last YouTube video um just to help
8:21 you guys like understand the principles behind this because like I said, taking
8:26 away principles more than just like here's how this specific harness works,
8:31 I think is the most powerful takeaway. So, we'll go over that and then I want
8:36 to um give you guys a little bit of a teaser of what it built when I tested
8:40 this before the stream. So, that's what we're looking at here. Um, and then
8:45 finally we'll run this live and so we'll we'll watch as it creates all of the
8:50 tasks in linear and it kicks off the different agents to go and knock things
8:54 out one by one. It's really really cool to watch it live. And this is the kind
8:59 of thing that we could run for 24 hours straight. Um, now we are not going to
9:03 stream. Ray and I are not going to stream for 24 hours. I'm sorry.
9:06 >> Come on, man. Let's do it. Mr. Beast style. >> We're going to stream and every hour
9:09 we're going to give away [laughter] >> Right. Yeah. Exactly. Yeah. $10,000
9:13 giveaway every hour. No, every >> Yeah, maybe someday. I think that
9:17 actually would be cool. I have thought about doing like a live stream marathon
9:21 before just to like see how long I I can go before I like collapse at my desk,
9:25 but that is not for today. And so, we won't see the full application get built
9:30 out, but you guys will see everything you need to see to watch it all happen,
9:35 you know, in real time where it'll create all the tasks in linear. we'll
9:39 see the first few coding agents get kicked off and assigned to some of these
9:44 tasks. So we'll see all that happen. So all the principles we will cover and
9:47 that's like I said the more important takeaway versus you seeing the full end
9:52 product and that's also why I want to show this here um because then like yeah
9:57 this isn't from our run today but it is what it created when I ran the same
10:01 thing. So yeah, and speaking of running the same thing before, um that that is
10:05 my the latest video on my YouTube channel which I'm hovering my mouse over
10:10 right here. So you can see the little preview like even shows the the Excal
10:14 diagram that I'll cover today. So a little bit of this is review if you have
10:18 seen this video already. Um but what I didn't do in this video is the linear
10:22 integration. And so before it was managing all the tasks and things just
10:29 in a JSON file um locally and so um like this this file right here. So I'll kind
10:30 Why Cole integrated Linear for task management and observability.
10:32 of talk about this >> in a bit as well and like how that has
10:36 evolved into linear and then like what I changed to make it work with linear.
10:40 >> Um and the biggest reason I'm excited about this besides the fact that Ry just
10:44 really loves linear is I think it's really powerful having our agents um do
10:50 their work where we also do our work. And so if we're going to have these
10:53 agents run for a really really long time, um, we need to have observability
10:59 and being able to like add human in the loop, but like we can't interrupt it
11:03 because it's set up as a process to just run in the background. So we need some
11:07 way to communicate to it. And we can do that by managing the tasks in linear.
11:11 Like I can click into this task or maybe this is a bad example because it's
11:14 already done. I can click to one that isn't done yet and I can edit the
11:18 description. If I want to change what it builds for this task, I can do that. So
11:22 then when the next agent comes and picks it up, then it takes in my feedback. And
11:25 like this is the kind of thing that I think we really need if we are going to
11:29 get to the point where we actually trust our agents to run for, you know, 24
11:34 hours straight or beyond. And so like right now, I would call what we're doing
11:38 here experimental, but it's going to get to the point over the next year or two
11:41 where this just becomes less and less experimental and more like how people
11:45 actually work with their coding agents. So we're on the forefront here.
11:49 >> Yeah. I think another thing why I started to check out Linear was there's
11:52 somebody in my Discord, his name is Cory Alice who kind of brought that up to me
11:56 because he is a chief technical officer and does a lot of work like this and
12:01 he's working on massive massive projects and so he's the one that kind of put me
12:05 on to this whole workflow. I'm like wow it's he says it's very token efficient
12:09 because you're not having to read and write out all these different things.
12:13 You just hand that off to the MCP. Um, but there's like because he's an
12:16 experienced technical person, also project management background, uh, it
12:20 kind of makes sense if you're actually going to track it in your head. But also
12:24 other agents because agents are trained from human beings and human work, they
12:29 naturally see a system like this and can already know, you know, what logical
12:32 pieces they need to grab. Uh, and then because this is already a pretty
12:35 established framework and and professional companies from all of your
12:40 Fortune 500 people are using linear, this is not an ad. It feels like an ad.
12:43 They probably should be sponsor us, please. [laughter] >> You're not sponsoring.
12:48 >> Yeah, I know, right? Um, it's really interesting for
12:53 monitoring the work. And you can do this on your phone. You can do this in
12:56 various places. Uh, other agents can even spawn out and then check the work
13:00 and then grab pieces of work. And, you know, you don't have to make up your own
13:04 system for this. It's kind of already pretty well established. So that's kind
13:08 of another advantage of why uh you know like this insight from Coreis was
13:11 extremely you know useful in my workflow. I was like okay this is really
13:14 interesting here uh and why we want to do this and it's so cool to see you kind
13:16 of adopt it here. This is really awesome. >> Yeah I appreciate it Ray. I'm excited to
13:22 get into this. Um yeah and one more thing before I really dive into the
13:26 article and explaining how the harness works. Um someone mentioned in my chat
13:32 can Archon use linear? Ray I don't know if you've heard of Archon before. It's
13:35 an open source project that I started earlier this year that I've been putting
13:39 a lot of effort into. Uh maybe I'll pull this up really quickly because there's a
13:42 Discussing Archon, Cole's open-source AI command center.
13:43 lot of ties between Archon and what we're doing here. >> Oh.
13:46 >> Um so this is like my command center for AI coding.
13:48 >> Whoa. >> And there's two main components to this.
13:52 It's a place for us to curate our knowledge for our coding assistants to
13:57 search like libraries that we're using. Um and then it's also the place for task
14:02 management. And so we have a kind of canban board or sorry conbon board that
14:07 um looks very much like this and it has an MCP server just like the linear MCP
14:11 server. It can move around tasks in real time. >> Um and so yeah a lot of like what people
14:17 are seeing here if they've seen Archon before and I know you haven't Ray and a
14:22 lot of your viewers. So um just like kind of giving a quick um or like making
14:26 a quick parallel here. I built something kind of like this as a custom solution
14:30 that's more like AI optimized. Um, and so there definitely are some parallels
14:33 to be drawn. The reason that I want to do linear specifically is because like
14:38 yeah, Archon is really cool, but like I want to integrate it with an application
14:42 that people are already using that has all these features that like is way more
14:46 than Archon because it's like, you know, a billion-dollar company building a task
14:50 management solution that has a lot more. Um, so yeah, I have had definitely ideas
14:56 of how I could like build Archon into different software like Linear or Jira
15:01 or Asauna. Um, and so like maybe that will come in the future, but um, also
15:05 just like a lot of decisions of like where do we want to take Archon next?
15:08 And so anyway, I just wanted to mention that really quick because I know a lot
15:11 of you might be curious if you've seen both before. Um, but right now we're
15:15 working in linear because linear is something that like people already do
15:18 their work in. >> Um, like like Rey, people that he knows,
15:24 people I know as well. Um, and like I already do my work in something. I I use
15:28 a sauna a lot for my task management and so it's very similar to this.
15:33 So yeah, any other thoughts you have Ray before we dive in here? So I'll just
15:36 kind of start by explaining the harness once we're ready. >> I think we're good. Yeah, excited to go
15:40 in and um yeah, just going to get a quick check for the the chat. I know
15:43 we're we're kind of doing this live stream thing. I'm trying Riverside. Uh
15:47 this is not AI Ray, by the way. This is [laughter] me. Uh you know, if you're
15:50 kind of wondering u I'm trying to get my my stuff synced up. This I'm using
15:53 Riverside. So, there's a little bit of delay with my lips uh and and and my
15:56 speaking, but that's okay. We're going to kind of deal with it. Uh Cole's going
15:59 to be basically driving this whole thing. And if you're kind of just tuning
16:02 in on my side of thing, we're going to take a look at this effective hardness
16:05 for long-running agents. Anyone who's just kind of popping in, um this is
16:08 something that I saw McKay Wriggley actually talk about, which was really
16:11 interesting on Twitter. if you don't follow in the last 24 hours, uh it's
16:14 going to get a lot of you're probably going to see really big people start
16:17 commenting on his thing. But basically, he spent a really long time in the last
16:21 two weeks trying to get these things to run autonomously for as long as
16:25 possible. And he said he was going to release a bunch of videos on it. You
16:28 know, Cole has already kind of been down this track. This is something that I've
16:32 been wanting to get into and I'm hoping that like uh this is going to be
16:34 something where all of us are going to be able to participate. So, any
16:37 questions you have, you know, Cole's going to walk us through basically how
16:41 to build this type of system. Uh, but this is like something that you
16:44 shouldn't really be scared to to run into, you know, just because you see us
16:48 doing it. We would highly experiment. Uh, highly encourage you to go to the
16:51 blog post right now, maybe follow along with us during this live, ask some
16:55 questions, get this stuff set up. You're going to probably have way more
16:57 questions and you're going to say, "Okay, maybe take one workflow that
17:01 you're currently doing by hand and see if you can hand it off to an agent like
17:04 this." Right? Uh right now we're going to build an app, but then like maybe is
17:07 there a workflow that you do all the time like by hand? Uh these agents are
17:12 actually quite good at looking at your screen and figuring out what to do next.
17:15 Uh and then maybe controlling some other types of systems. So I want you to kind
17:19 of start, you know, to take this in your mind and start thinking about this and
17:22 maybe see, you know, what what's kind of around the corner here. And uh by us
17:26 walking through it live and you walking through it live with us, uh you know,
17:30 this is going to be a great form for us to interact. So, we'll have a section
17:33 towards the end where you can ask us any questions and we'll be able to answer
17:37 them AMA style and so forth. So, um yeah, that's kind of is what it is and
17:39 appreciate everyone's patience through this and uh excited to see you cook.
17:44 Sweet. Yeah. All right. And like Ray said, like don't be intimidated by this
17:48 because even if it seems like a lot going on here because it is pretty
17:51 involved and I'll try to break things down the best as I can. Even if there's
17:56 just a couple of principles that you get out of this harness where you're like,
17:58 "Oh man, like I want to apply this to my own system." Like that's what we really
18:02 encourage you to do. So again, I've already said this, but it is worth
18:05 repeating it. It's taking the ideas from this that I want as the main takeaway
18:10 for you and and what Ray wants as well instead of like you trying to like fully
18:14 understand everything in this harness and know exactly how to use like this
18:19 specific tool from Anthropic because in the end this is more of a experimental
18:24 teaching tool than anything, right? Like Anthropic is not building this into
18:27 cloud code or anything. It's just some of their u research that they're doing
18:31 to continue to prove that they're on the cutting edge, right? And that's what
18:35 we're doing as well. And so the repository that I have for today is uh
18:40 fully open source. So I'm going to put this as a link in my um stream chat as
18:44 well. And you could do the same if you want, Ry. Um >> yeah, if you pop into studio chat, then
18:49 I'll be able to send it over to the >> Yes. All right. Did that just now.
18:53 Perfect. Cool. So yeah, I took the harness and I kind of just created my
18:58 own fork, right? Like my own version of Anthropics harness where I changed the
19:03 prompts and I added in the linear MCP server. So I made it so that it works in
19:07 linear. And so I want to start by explaining how the base harness works
19:12 and then talk about how I've adjusted things to work with linear because I
19:16 also think this is a really good lesson of just how we can take someone else's
19:20 work and adopt it for our own system. Like I said, the original harness is
19:26 using a JSON file to keep track of all these features that it's mapping out for
19:31 this very longunning task. And so now I'm kind of adopting it for my own
19:34 system. I mean, this might be more Ray system because I don't use linear
19:37 myself, but just showing you guys a demo of how you could do this kind of thing.
19:42 And so the main problem that we're addressing here that Enthropic is
19:47 addressing is agents. They struggle when they work across many different context
19:51 windows. And this is a a big engineering challenge that um it's not new and it's
19:56 something that I have been developing solutions for for months and months at
20:00 this point. And so it's not like anything here is revolutionary, but like
20:04 I said, there are just a lot of really smart ideas to take away from how
20:09 anthropic is tackling this specifically. And so you think about um you know
20:14 strategies like memory compression or how when we start a new conversation
20:19 with an agent we can have it like read certain files in a code base to um
20:23 understand you know like what we built so far and what's coming next. And so
20:27 some people they have kind of like a change log they build up or maybe you
20:32 rely on like the git change log or you have like certain files that you have it
20:36 read as far as like the plans that have been executed at this point. There are a
20:39 lot of different strategies that people have to make it so that like we can um
20:45 kind of pick up work in a new context window very quickly. And so that a lot of the ideas that um
20:51 you all have probably been experimenting with already if you're using AI coding
20:55 assistance, these are just taken to the extreme for this demo here so that we
21:01 can run agents for a very long time. And so the main thing that we're thinking
21:06 about here is how can we create this massive set of requests like we see here
21:11 in linear we have dozens and dozens of tasks that we have to knock out in order
21:15 to build this application. So how do we define these tasks and then and then
21:20 actually like handle them with many different sessions and we need to
21:24 somehow like carry over context between these different sessions. We can't have
21:28 one wrong for too long because large language models always hallucinate when
21:33 their context window gets very filled. And so we have to like define these
21:36 break points for like okay we just build in like one feature or just a couple of
21:41 features and then we cut the agent off we start a new window and then we have
21:45 some kind of way to like carry over all of the most important context to that
21:49 next window. Right? Like that that's essentially what they're doing with this
21:53 harness here. And to like really help you understand how this works, I have
21:55 Architectural deep dive: How the agent harness works.
21:57 this diagram that I teased earlier that I just want to go over really quickly
22:01 here that explains how the entire system works and the whole idea of like running
22:07 many different agents in a loop and like having these kinds of artifacts and
22:11 strategies for sharing context between them. Um, and so I'll kind of zoom in at
22:16 the start here because um, this, by the way, this whole demonstration today is
22:20 more for um, green field development um, as in like starting a project completely
22:24 from scratch. Now, you could definitely adapt this whole harness to work for an
22:29 existing codebase because instead of creating, you know, 200 or 50 whatever
22:34 new test cases or features for a brand new project, it would just be like for
22:37 what you want to build on top of what you already have. So, just keep that in
22:40 mind. I'm going to be focusing more on building something from scratch because
22:45 the the very first file that they have in their harness is the app spec. Um, a
22:51 lot of you might know this as a PRD. It's really just like that high-level
22:56 file that outlines the scope of work for your application. Like once everything
23:00 is built out that you define here, that means your app is done or at least the
23:05 MVP, the minimum viable product of your app is done. And so this is the primary context that
23:12 goes into our very first session. This is our initializer agent. And so we have
23:18 an entire agent dedicated to creating the feature list before we get into any
23:22 of the coding. So you can see that session two, the the second session is
23:27 where we start with the coding. And so that the only job of the initializer
23:33 agent is to create this featureless JSON file. This is the one that I was showing
23:38 earlier that outlines all of the features that we have to build out based
23:42 on that appsp spec. And then it even has the steps to take. And I can zoom in a
23:46 little bit more here. It has the steps to take in order to like validate that
23:52 this feature is working. And so this whole system also has the Puppeteer MCP
23:58 server built in. And so our agent is able to spin up a browser autonomously
24:03 and it can navigate it. So it'll like start the back end, start the front end,
24:08 go visit the website and try sending in requests or like you know whatever
24:11 website we're built or application we're building here. It'll interact as a user
24:15 interacts with the site to verify things. And then only once all these
24:19 steps pass will it mark this feature as done and then move on to the next one.
24:25 And then this is exactly the process that I moved over to linear that I'll
24:29 explain in a bit. But like I said, I want to start by covering how the
24:33 harness works, but without my additions on top. >> So, did you make that feature list by
24:37 hand or you just kind of prompted an agent and like yapped a whole bunch of
24:40 things or how did that kind of come about? >> Yeah, so the feature That's a really
24:44 good question, Ray. Um, yeah, please keep feel free to feel free to keep um
24:48 interjecting with questions to keep me on track here. So, >> no, it's okay. [laughter]
24:52 >> No. Yeah, I appreciate it >> because sometimes I'm like, I'm not
24:54 going to write this by hand for sure. Like, >> Right. Yeah. Yeah. So, I did not write
24:59 this by hand. Definitely not. Um, so it's all in the prompt for the
25:04 initializer agent. And so when the initializer agent begins, it um, here,
25:10 let me open up this prompt in preview so it looks better. This is sent into the
25:15 initializer agent and this dictates everything that it does. And so
25:19 basically what we tell it to do is read our appspec. This is the one thing that
25:24 we would create ourselves or maybe use an AI coding assistant to help us
25:28 create. And then based on the appsp spec, it creates the feature list JSON.
25:34 And we define here the exact format for the JSON. So we we don't ever touch
25:39 anything here. Maybe we validate some of these things, but also because this is
25:44 like 200 plus test cases long. We're probably not going to do much
25:47 validation. This is all experimental anyway. We just want to see like how
25:50 Showing the detailed `app_spec.md` file for building a Claude.ai clone.
25:51 long it can run. So, we pretty much just let it generate this based on our appsp
25:56 spec, which by the way, the appsp spec that they have. This is the default one
26:01 that comes with the anthropic harness just as kind of like their demo for you
26:05 if you want to like just take their prompt and run with their prompt. And
26:09 so, what we're building today and what they have as their basic demo is a
26:15 claw.ai clone. And um that's the little teaser I had for you earlier. Like this
26:19 application right here is running. you can see on localhost like this is what I
26:24 had this harness build in a previous execution and so this entire prompt it's extremely
26:30 detailed because obviously we have to have a lot of context to be able to pull
26:36 out you know 200 plus test cases and so if you want to and you want if you want
26:39 to build something entirely different from the cloud.ai AI clone. All you have
26:45 to do is change this appspec file. So you could build your own website, like a
26:49 portfolio website. You could try building a SAS application. I don't
26:52 care. It doesn't care. You just have to edit this however you want. And then
26:55 within the initializer prompt, we're telling it to like read this and then
26:59 make all the features based on that. So it's entirely flexible. Like you can put
27:02 whatever you want here. >> Wait, hold up. So you're saying that if
27:09 I want to create a new app, the structure that is currently laid out in
27:14 this type of thing here >> Mhm. >> is something that we can give to Claude
27:20 Opus as an example and say here's a spec that I normally like to use. Uh
27:24 >> here's my Yapp idea, you know, one paragraph long whisper flow or something
27:26 like that. >> Yep. >> Um please >> go ahead and see if you can start to
27:31 build the structure for this >> as my prompt. Right. and then ask me
27:35 follow-up questions because there's a lot of stuff that's probably going to be
27:39 missing in my initial request and then maybe do subsequent calls like I have
27:43 two MCPS I really like and one of them is XA code and the other one is Ref
27:47 tools. So then I could just have it then go search for documentation to help fill
27:52 in this spec document which the spec document once filled out is your pretty
27:57 much I think it's the thing that the agent will always kind of anchor to,
28:00 right? So as it goes through for very long sessions and creates new agents um
28:05 as part of the architecture, it's probably going to keep referencing that
28:08 spec and where it's at. And then that's kind of in essence that's probably like
28:11 the map, right? For if you want to travel across America, you you need a
28:15 map. And I think that's that's an interesting concept. Wow, that's amazing
28:18 how long it is. It looks very intimidating at first, but once you
28:22 think about once you have the template, >> then you can kind of move forward from
28:24 there. Right. >> Right. Yeah. And so I really appreciate
28:28 you saying that because I actually did exactly what you described earlier. I
28:35 gave Opus 4.5 this spec as context and I said like this is the format I want to
28:38 use, but then like here's a brain dump of another application that I want to
28:42 create. And so I had it create a version of the app spec but specific to a
28:46 different application. And so everything looks pretty similar here. It has the
28:52 same kind of like XML um prompting but it's for a completely different
28:56 application. And so never feel like you have to write this entire thing out from
29:00 scratch. It can just be like here's the template, here's my brain dump. Now
29:04 please me mesh these two things together and then this is what we send into our
29:09 initializer agent. And so kind of like the map like you said Ry that is the
29:14 feature list. So um surprisingly and like maybe it would actually be better
29:18 to reference the appspec but it actually never goes back to the appspec. The
29:21 initializer agent will create the feature list based on the appsp spec and
29:25 then this is the source of truth for all the agents the coding agents going
29:27 forward. >> I see. Yeah. Because in in my scenario what I was thinking was like for those
29:34 who are just kind of coming along here. This is probably going to this is this
29:38 these parts are probably the most important parts in your app building is
29:41 defining what you're going to build. U because the agents are really good. You
29:45 can almost think of the agents as like having zero memory, right? So they have
29:49 all this space knowledge but they don't really have a lot of configuration
29:53 details and nor do they really need to know a lot of things. We've learned over
29:57 time because their context windows are so small, if you just give them a very
30:01 well- definfined task, they will execute and if they have something to verify
30:06 against that loop of the agent doing that is kind of the secret sauce. And
30:11 that's kind of what we're doing here in essence, you know, with what Cole is
30:14 outlining. It's like this is part of the template. Anthropics models have been
30:19 trained to follow these instructions. That's why you have these weird
30:23 brackets, you know, and their XML structure for this type of thing. And it
30:29 just so happens that if you can also, you know, say explain your idea, give it
30:33 an example, it's going to fill in that type of details, you know, using your
30:37 idea. Um, you'll obviously want to review it, but like make sure that all
30:41 the spec meets your needs. But I think this is the biggest unlock. And if you
30:44 spend the most amount of time in this area, it's going to probably pay off a
30:49 lot more than later on wasting a bunch of tokens. Uh, and I have a couple
30:52 Pro-tip: How to use a Claude Max subscription instead of API credits.
30:52 follow-up questions here. It's like, >> yeah, >> if you're using this um agent harness,
30:58 are you can I use my Claude Max account or do I have to pay API cost to build
31:02 this type of thing? >> Yeah, that is a really good question.
31:07 So, you can use your subscription. You don't have to use API credits. But this
31:11 is actually something I added in myself. So if you look at the repository here,
31:17 not the one that I sent with the linear built on top because I have the
31:21 instructions here for getting using your subscription, but if you look at the
31:25 original one from Enthropic that they linked to in the article, they have you
31:29 use your API key. Uh which I this is like the first gripe that I have when I
31:32 saw this. I'm like, "Okay, man. I have to build something myself because I'm
31:35 not going to use my API key for this. It would be literally hundreds of dollars
31:40 just to have this run for 24 hours." >> Wow. And so yeah, I'm using my Mac
31:43 subscription. Okay. >> And so that's one of the edits that I
31:46 made on top. Um, so it's actually very easy. If you have the Cloud Code CLI
31:50 installed on your machine, you can run the setup token command that'll give you
31:55 a long running OOTH token that lasts for a year. And so now when we call cloud
32:01 code programmatically using the cloud agents SDK, so we're actually like using
32:06 cloud code in our Python code for this harness and but it'll be using our
32:10 subscription, not our API key. >> Oh, that's that's brilliant. Um I just
32:14 want a big shout out to uh this person says pure gold. Uh thanks guys. Yeah,
32:19 and uh I think another person had a comment here says like great topic. Uh
32:22 Cole made a great video on that a few days ago. How many runs has he tried
32:25 since the video? and you tried to use work trees to work on multiple tasks at
32:28 the same time. Just kind of curious about that there. >> Yeah, it's a good question. So, I have
32:35 only done one more run since that video on Wednesday. Um, I was working on some
32:40 other stuff. Um, so I I do both like AI coding and AI agents on my channel and
32:44 so I've been doing some things like actually building some AI agents and
32:47 some rag agents over the past few days. So, I I've kind of like swapped over to
32:51 that. So, I haven't done too much more experimentation with this. I haven't
32:54 done anything with work trees either. Um, however, well, I have been doing a
32:59 lot of stuff with work trees, just not with this harness. Um, I've been doing
33:04 some things with work trees for um, an application in the Dynamis community,
33:08 something that I'm working on for remote agentic coding. So, that's a very
33:12 separate thing. I have not tried work trees for this harness in particular.
33:15 >> That's cool. >> Yeah, I'll have to stay subscribe,
33:19 folks. Make sure you share this with a friend if you if you're watching right
33:22 now. share this with some other friends, people who are currently building with
33:24 us because um that's kind of what we do around here, you know, like to share
33:28 this information before we proceed. Uh big shout outs to get max. This is super
33:31 interesting repo. Thanks for sharing. Um yeah, let's go ahead and kind of dig in
33:33 further. Yeah, >> sweet. Yeah, and someone in my chat asked um what does a harness mean? Is it
33:41 some new IDE? And yeah, very simply, a harness is just think of it like a layer
33:45 on top of your coding assistance that defines the process, right? So we have
33:50 the initializer. It creates and I I kind of want to like finish explaining this
33:54 diagram here. It creates all of the features that we need to build out. It
33:59 also creates this script that our coding agents can run going forward to spin up
34:04 the entire website. So like this script will start the back end, it'll start the
34:09 front end and then it is able to validate things like actually running
34:13 the application going forward. And then the initializer agent will also scaffold
34:17 out the project, you know, kind of creating the boilerplate for the
34:21 application, the front end or whatever it is. And then it'll initialize the git
34:25 repository. And so all these things it's doing is just to get it to the point
34:28 where once we run the first coding agent, it's able to dive immediately
34:33 into understanding the codebase like what we have right now and then building
34:37 that first feature or that next feature for all of the the sessions that we have
34:42 going forward. So this N here, it just means that we don't know how many
34:45 sessions are going to be. It's just we're going to go until the application
34:50 is done and all of the features that we have in the feature list are marked as
34:54 passes or with my integration with linear that I'll talk about, we go until
34:58 we have all of the features in the done column. Um, and then the other thing
35:03 that it creates that is really cool is this clawed progress text file.
35:09 So let me go to that really really quick here. So the clawed progress text file,
35:15 this is where the coding agent will put a summary of what it last did so that
35:20 the next coding agent that runs because remember it starts in a brand new
35:24 conversation window. It's totally fresh. It can read this to quickly catch itself
35:29 up to speed on what the last coding agent did or for the first coding agent.
35:34 It gets to, you know, understand what did the initializer agent do for me? And
35:38 so this is like the the source of truth of what like what just happened because
35:43 oftentimes what just happened is going to guide the next agent for what it
35:46 should do next. So this is a really critical file as well. So I kind of have
35:50 this as like the the core artifacts, right? Like we have the shared state
35:55 that is really like the glue that connects all these agents together. And
35:59 these three files are the big ones that solve the problem of when we have a
36:04 brand new session with an agent, how do we have it quickly understand where it's
36:08 at in the codebase right now? Like what should we build next? What maybe isn't
36:12 working yet? Those kinds of things. And so the agent when it runs, it reads all
36:17 these files. It uh primes itself on the codebase like just getting its bearings.
36:22 It also does regression testing. So we have regression testing built into the
36:26 system where it will kind of spot check by randomly picking a couple of previous
36:30 features that were implemented and making sure that they actually work. And
36:35 so that way if it ever breaks anything in the code as it's adding in new
36:39 features, it will hopefully detect that and then loop back and fix that before
36:45 it moves on to the next new thing. Excuse me. But assuming that it uh
36:49 passes the regression testing, then it'll pick that next feature. So it'll
36:55 go back to the JSON file um and it will find that next one that has passes as
37:01 false, which for if we go down here for this test, I have to go like about at
37:05 the halfway point here. Now we start to see these things where passes is false.
37:09 So it'll pick the next one to knock out and then it will implement and do a lot
37:13 of testing as well. So it uses the Puppeteer MCP server. so that it can
37:18 actually go to the browser like we would as a user and it can send in a message
37:23 like um for example I'm gonna I'm just gonna like copy this one here and send in a new chat like I
37:28 can say create a simple JavaScript function that adds two numbers put it in
37:31 an artifact and like what I'm doing right here to test the application and
37:35 show it to you live it would do the exact same thing because it has control
37:41 over a browser um just like we do and so it's the validation is never perfect but
37:47 like man is it good. And so it's able to catch like basically everything. Even
37:52 even um where like the UI just doesn't look good. It's still able to find that.
37:56 And that's a newer thing. Like we haven't really been able to trust coding
38:01 assistance until recently to find things that don't look right in the user
38:03 interface. Like it's more of the functionality that it could detect that
38:08 isn't working in its own testing. But now it's even the user interface.
38:13 And it'll even go so far as to do like mobile uh style checks. Like I can go to
38:18 an iPhone and like making sure that this actually looks good on an iPhone and um
38:22 yeah, the title here looks bad. Like it's not perfect, but it's still like
38:25 pretty good. >> Wow. >> And and functionality app looks good.
38:29 Dang, man. >> Yeah, it looks great. >> Yeah. Like there there are some oddities
38:36 here. Like if I refresh um and I like send it a new message here, you'll see
38:41 that like it's duplicated here until like start streaming out the response.
38:44 So like there's things where it's not perfect and and that's why I want to
38:48 reiterate that like this is experimental and you still want to like come in and
38:52 validate things yourself and have like follow-up prompts at the end. Um but
38:56 also like this demonstration that I have right here from a previous run, I let it
39:01 go for exactly 24 hours for my YouTube video. And uh if I let it go for it
39:08 completed about like 54% of the tasks. And so if I let it go completely end to
39:12 end, it probably would have addressed these things as well. >> But I'm just saying that like even if
39:15 there's little things like we can just come in and iterate and it's still going
39:20 Showcase: A Claude.ai clone built by the harness in 24 hours.
39:20 to save us hours and hours and hours. Like it built so much here. We have the
39:23 whole settings panel where I can switch between my themes. I I can adjust my
39:29 model that I'm using and the max tokens keyboard. It even built in keyboard
39:32 shortcuts. Like there's so many features here. Like there's a reason why this
39:38 feature list is over 200 long because >> you're blowing my mind right now. So
39:43 autonomously this is just has been running. This is obviously from the blog
39:48 post. But the power like we're getting the template. It's like they're not just
39:52 saying, "Hey, this is a theory. This is a theoretical paper. Our agents can do
39:56 this." They're like, "No, like this is the actual recipe, guys, and you can
40:00 take that and try it yourself." which is kind of what Cole is doing and we're
40:03 walking you through the entire overview here. But like the thing that's kind of
40:09 shocking me and blowing my mind is that testing feedback loop which I normally
40:14 have to be there by hand and I've usually have to prompt puppeteer or
40:17 something else like cursor now has that built into their browser but this is
40:22 done in in essence headless. So you weren't there you weren't in an IDE
40:26 typing this type of saying you know load the web browser and then tell it back
40:31 the response. It's using the MCP tool call underneath the hood adjusting
40:35 itself, adjusting for mobile view, fixing little things, prompting in a
40:39 loop until it completes that task. Mark said it's true. Moves to the next thing.
40:44 Moves to the next thing. And yeah, you just said 24 hours like you as soon as
40:48 you hit go, you said, "Okay, I'll check back in later. Let's see what's going
40:50 on." >> And wow. [sighs] Wow. Wow. >> It's very cool.
40:55 >> Yeah. And there have been a lot of other people that have tried this since I
40:58 shared as well and they're getting really good results too, even for their
41:02 own applications. Um, like one guy I know, um, Sean, he actually has done a
41:07 lot of work on Archon. So, shout out to you, Sean, if you're in the stream here.
41:10 He he took it for brownfield. So, he actually applied it to do like a really
41:15 large, um, refactor on a codebase that was like 12 major versions behind on
41:19 Typescript. So he had to like update TypeScript and do like this refactor
41:23 that would typically be the most um painful thing you could do as a
41:27 developer. Like trust me, I've been a part of like .NET migrations. If anyone
41:32 in the chat here is a developer that's had to do like a .NET 4.8 tot 6
41:38 migration, uh oh that was that sucked. That was so bad. And I did not have a
41:41 coding assistant back then. And so being able to just like have it rip through
41:45 all the refactoring and do the kind of validation to like make sure that
41:48 there's no like deprecated code still is really powerful. Um, webdev Cody, it's
41:53 cool to have you in the stream. He said, "I tried this last night for 24. It's
41:58 pretty amazing. 128 commits in. I love it." Yeah. Yeah. So, it actually makes a
42:01 commit every single time it implements one of these features, which is another
42:04 one of those like little things that I just really appreciate with the harness
42:08 here because it'll basically create a save state at every point so that if
42:12 anything fails, you could revert or it could revert back and then retry that
42:16 feature. Um, or you know, if it like implements one feature, but then it
42:20 actually breaks like 10 previous ones, then it's like, yeah, probably go back
42:25 to a previous commit and then retry. I don't know how much you would do that
42:27 reliably, but it's like that's the kind of thing you can do when you create a
42:31 safe state at every point. >> Somebody was asking us how does this
42:35 compare to like spec kit? Um it's an interesting question here.
42:40 >> Yeah. So spec kit is I would say it's a little bit comparing apples to oranges
42:44 because um when you have like a I like to call these context engineering
42:50 strategies like um spec kit or BMAD or Claude Flow or PRP. Hopefully at least
42:54 like one of those rang a bell for all of you. All of those are more like you're
42:58 involved with the coding assistant and it's your job to like go through the
43:03 planning and um do your validation as well, right? And so this is a very
43:07 different thing where like I mean you could kind of like build spec kit into
43:12 this and have it um go through and like create the constitution and create like
43:17 the I forget all the documents that it creates but like the the it could you
43:21 could have it create the data model and things like that but um this is more of
43:25 a separate thing where it's like if you want to just let it rip for a really
43:29 really long time try this harness if you want to be a part of the process and um
43:34 really code with the coding assistant. That's when you would do something like
43:37 Speck Kit, which when you're working on like a very production grade
43:42 application, I I wouldn't trust this right now. I would have it do something
43:46 fully end to end. Um, but definitely like for just like experimenting with
43:50 what is possible and building proof of concepts, this is really powerful. But
43:53 like spec something where if you want to build something, you know, more
43:55 production ready, it's like that could be a good strategy. >> Man, that's cool. Yeah, thank you for
44:01 kind of answering that. Such great participation. And anyone who's tuning
44:03 in live right now, we're basically going over this um you know, how to basically
44:08 get this thing running for a very long time and u we're kind of just breaking
44:11 it down a little bit step by step and appreciate all the audience interaction
44:15 here. Um you know, this is probably going to be multiple different live
44:19 streams and stuff like that. So uh yeah, that'd be cool. >> Yeah.
44:22 >> Yeah, I'm looking forward to showcasing this more and more. Um maybe not even
44:26 necessarily this harness, but just as I take ideas and apply it to my own
44:30 systems, right, like I told you guys to do. So definitely going to be doing more
44:34 of that. Um and yeah, at this point we've pretty much already covered the
44:38 whole architecture of the harness here. Um the last thing that I'll say is um
44:43 yeah, I kind of like veered off of the diagram once I was talking about this,
44:46 but we already talked about how it makes a commit makes a save state at the end
44:51 of the coding agent and then it's going to just loop and it's going to do this
44:55 over and over and over again until all the tests are passing.
45:01 And um yeah the way that you define when the agent decides like I am done let me
45:06 update the cloud progress and then go to the next agent like that is all up to
45:10 the prompt that you have for the coding agent. And so just like I covered very
45:14 briefly the prompt for the initializer agent. We also have one for the coding
45:20 agent. And so the entire process of like make sure you get your bearings. Here
45:24 are the commands for that like reading through the git log, reading the cloud
45:28 progress so you know what the last agent did. Like all this is just laid out as a
45:32 prompt. Like you could think of this as a slash command for example, but we're
45:36 just like loading this right into the prompt for the agent using the cloud
45:41 agent SDK. Um, and then we have it start up the website using the init script that the
45:47 initializer created. And then this is the verification like the regression
45:50 testing that I mentioned earlier. So it picks one to two features that are
45:54 already marked as true to make sure that they're already working. Then we choose
45:58 one to implement. We implement it. We do the verification with the browser
46:02 automation. So at this point, this is all review, right? Like I already talked
46:05 about it, but I'm just showing it to you in the prompt as well. Um and then we
46:11 update the feature list JSON um to mark it as true, right? Like it's now
46:15 passing. And then we make the commit update the cloud progress and then we
46:19 are done. So it's very much just like boom boom boom 10 steps that's your
46:25 process. And then if you want to in here like this is where you can define the
46:28 behavior of the agent like after you knock out one thing in the feature list
46:34 that is when you should uh call it and then pass it on to the next agent. or
46:38 maybe you want to try to move a little bit quicker and you're like, "Okay, I
46:42 trust you to do five of these and then you should pass it on to the next
46:45 agent." So, however much you want to customize things, it's totally up to you
46:50 to edit the prompts. So, like just like you can change the app spec to adjust
46:54 what application is being built, you can also change the initializer prompt or
46:57 Detailing the prompt changes for the Linear integration.
46:58 the coding prompt. And um this is a nice segue into the linear integration
47:03 because that's exactly what I did to make it so that this application uses
47:07 linear to track all the features instead of this featureless JSON file which I
47:11 think just gives us a lot more control. And then the agents working where we
47:16 work like I was saying earlier. And so I have a second version of the codebase
47:19 that'll pop up here. This is the one that I'm going to be using going
47:24 forward. And so um the appsp spec I have as the same thing like I'm going to just
47:28 build the same application here for the sake of the demo. But then the only
47:33 other things that I really updated in this codebase besides making it so you
47:38 can use your cloud subscription is I just changed the prompts. So like
47:43 looking at the initializer here you can see that I'm having it use the linear
47:49 MCP server to create a project. So like one of the things that the um agent
47:53 does, the initializer agent does as well as setting up the scaffolding and the
47:57 git repo and everything, it now also creates the linear project because we
48:00 just want to like make it so that everything is set up initially from the
48:04 first session. And so you just see like a lot of edits that I made to mention
48:09 linear throughout this here. And another thing that it does instead of that
48:13 claude progress file, it also has a meta issue in linear. So I'll show you that
48:17 really quick. So it also manages like the context to hand things off between
48:21 agents also in linear. >> Wait, that's crazy. So you don't have to
48:26 prompt it in the middle. You literally have thought like through every single
48:29 step. >> Yeah. I wanted everything to be linear. Like no more local files cuz that the
48:35 other thing the other problem that I have with local files is um you can't
48:39 really run the harness in the cloud. Like what if I want to deploy it as
48:42 something that like I can kick off remotely? I don't want to like if I want
48:45 to be able to validate the progress or validate the feature list, I can't do
48:50 that unless I go and like SSH into that machine in the cloud and that's just
48:53 annoying. So, I want everything to like be immediately accessible to me. Um, so
48:56 that's another reason that I just like the idea of having all of the context
49:02 Ray's 'AGI' moment seeing the fully integrated system.
49:02 managed in a place that's easy for me to view, right? >> My guy, everyone in the chat spam AGI
49:08 cuz Cole Medin has literally unlocked AGI right now. Let's go. Yeah.
49:11 [laughter] Oh, I appreciate that. >> Yeah. Well, yeah. Thank you, Ray. You're
49:20 a good hype, man. [laughter] >> I'm Okay. I love this stuff just like
49:24 everyone else watching this stream. I'm like geeking out live like live. This is
49:29 my reaction. [laughter] >> I love it. And it's like every single
49:34 step along the way, like you've you've locked in super hard and you've taken
49:38 these different bits and pieces and now you're like, "Okay, I like this linear
49:41 thing. This is really interesting." And then you discovered like, "Wait, I can
49:44 run this on the cloud." And it's like all these revelations, right? And you're
49:47 just like, "Oh yeah, that's just like a feature. This is kind of what I do." I'm
49:51 like, "No, bro." Like you don't realize. And I think I think a lot of people are
49:54 having this like aha moment right now. It's like, "Oh, I think this is it."
49:58 Like this is the sauce. This is the the blueprint. It's literally right here for
50:02 us. At the end of 2025, now coming into 2026, it's about to be crazy. Holy
50:05 smokes. >> That's right. Yeah. Yeah. And and even if you guys don't use linear, just think
50:12 you could do this in any application that has any kind of similar user
50:16 interface. This could be notion, it could be your own application, it can be
50:21 a sauna, it could be Obsidian. Like, man, it doesn't take much to apply
50:26 because all of this is just like, okay, here's the thing that people don't
50:29 realize is that like everything that seems fancy with AI coding assistance is
50:33 just a bunch of prompts. Like that's all it is under the hood. Like everything
50:37 that I'm doing here is just like prompting it how to use the linear MCP
50:42 and all of the context in here is just a bunch of text and I prompted it to tell
50:46 it how to do this kind of thing. like it's um once you like understand that
50:50 it's just really easy to take these ideas and apply it wherever you want in
50:54 your own applications with your own process. Like one thing that that this
50:58 Explaining the Test-Driven Development (TDD) approach used.
50:59 whole harness does is the idea of testdriven development. And for those of
51:03 you more experienced engineers, maybe you kind of already picked that out. Um,
51:07 but all test driven test-driven development is is we define the criteria
51:11 like these are the things that have to pass and it'll even like write out the
51:15 tests that will fail initially and then we like build out the features and make
51:19 the tests pass. And so like slowly we have a larger and larger percentage of
51:23 our tests that are passing and that's exactly what we're doing here. Maybe you
51:26 don't actually like that though. Personally, I don't typically do
51:29 test-driven development. But it doesn't mean that this harness isn't useful for
51:32 me to extract certain ideas. like maybe the whole idea of the clawed progress. I
51:37 just love that it's so easy to have agents share context and then going into
51:41 linear and building it in through this meta task like I did. Like I did that
51:45 for a reason because I really like that part of the harness that Anthropic has
51:49 given us here. And it's cool because we can see that um and this kind of veered
51:54 off a little bit from my process. I kind of wanted it to update this um project
51:59 tracker in place, but what it did instead, which is still kind of cool, is
52:03 each of the agents just like wrote a comment as like a reply to this issue
52:09 here where it it did an overview of what it it did like and even calling out the
52:12 specific tasks that it marked as complete. And so for the sake of
52:16 brevity, I just had it run like three sessions here. Um, so this full demo is
52:22 from my YouTube video on Wednesday where I did not use linear. And then this it
52:26 was just like I'm not even showing you what it built because it was really
52:29 small at this point. Um, since I just wanted to do like a couple of things.
52:32 But yeah, that's what we're looking at right here. And so like all these tasks
52:36 that it marked off as complete was just with the first three coding agents. And
52:41 so I gave it a bit more um flexibility like you don't have to just do one and
52:44 then pass off to the next coding agent because I want it to go a little bit
52:48 faster here. Um so just showing you an example of how like I've customized
52:52 things however I want it to work. And it's cool that it even thinks through
52:55 like little details like it marks this as urgent so that like it always knows
52:59 like okay this is what I have to look at at the start of my session here. And
53:02 obviously we have that as a part of the prompt to the coding agent as well. So
53:06 just showing you that really quick. If I go to the coding agent prompt, it's
53:11 going to look pretty similar. Like we have the same 10 steps overall. I guess
53:16 there's one more step here or a couple more steps because it's a bit more
53:20 involved with linear, but overall it's the same thing where it's like get your
53:24 bearings. Here are a couple of um commands to do that and then check the
53:29 linear status. So using the linear MCP server to do that. Start up the
53:32 application. I mean this is like exactly the same as what we had in the harness
53:36 already. So I'm keeping as many ideas as possible, just adjusting things where I
53:40 need to take advantage of linear. Like now it's going to use linear to go and
53:44 find one or two completed features to make sure that that the regression
53:48 testing is looking good. U but it like still is using puppeteer. Like if I go
53:52 to the agent definition here that I have for the cloud agent SDK, the two MCP
53:57 servers that we have connected are Puppeteer and Linear. And in the read me here uh for this repo
54:05 that I've already linked you to I I have like all the setup instructions. So you
54:08 just have to get your subscription token, get your linear token, and you
54:13 set those as your environment variables, and the linear [clears throat] MCP will
54:17 automatically be connected to your projects and tasks. That's incredible.
54:23 That is absolutely incredible. It's just kind of the whole thing thought out
54:26 here. This is really cool. I think if anyone's trying to check out the repo,
54:30 um, we had linked it earlier a little bit above. And so if you're currently in
54:33 the live chat, I think you also have it in your stream as well. And we're going
54:36 to also try to pop it into the description. So if you're re-watching
54:40 this like in the video on demand also for those who are watching, maybe you
54:43 came in a little bit later after this video is done. I'm going to run it
54:46 through my transcription and stuff and then have some timestamps also in the
54:49 video. So that way you can go back to this live stream and you can kind of go
54:53 to those certain points to come back and reference again. This is uh really
54:56 really amazing stuff. Wow. Wow. >> Cool. Yeah. Thank you, Ray.
55:02 >> Yeah. And um I would say that like I probably want to keep iterating on this
55:07 more. Like for example, I typically like to manage my environment variables in
55:10 AENV. So I don't think I've set up everything perfectly for this yet, but
55:13 um I'm still kind of just like experimenting with how I want to set
55:18 this up overall. And then also there's a separate application that I've been
55:22 working on called uh the remote agent coding system. That's what I covered in
55:26 my last live stream on my channel. And so what I have there is like the ability
55:30 for you to interact with your agents in um Slack, Discord, Telegram, um GitHub,
55:36 and there's a lot of different platforms that I'm building support for. And so
55:40 like maybe adding this kind of harness into that application as well. So I have
55:44 I just have like a million ideas running through my head right now of how I could
55:49 take this a lot further for all of you. Um, and like maybe even having linear as
55:52 like an integration directly within the remote coding agent. Um, yeah, and like
55:56 having it so like this harness is something that you have as like an
55:59 option to kick off if you want a long running thing or you can just like be a
56:04 part of the the loop if you want to and then like use something like the spec
56:07 kit um to that person's question earlier. So yeah, I I want to like think
56:11 about how I could design this kind of thing into a larger project that has the
56:16 flexibility of like you can use this if you want or if you want something a bit
56:20 less experimental and more like you're a part of the process then you can. So
56:25 >> I I love this comment right here. Uh they're saying it's it's not Claude
56:30 code, it's Claude Cole. [laughter] >> Yeah, I appreciate that. Yeah, it's a
56:35 mix. It's like cloud code is the beast that powers everything and then I'm just
56:38 working hard on building a layer on top of it. And and like I said earlier,
56:42 that's exactly what a harness is. Is you just take something amazing like a
56:46 really powerful LLM or a really powerful coding assistant and you build the
56:50 context layer on top of it. Here are my rules. Here are my prompts. Here's the
56:55 process that I'm laying out through all of that. Um and and really all this
57:00 harness is is is two prompts. like it's actually so basic because it's like
57:03 here's the initializer, here's the coding prompt. I I guess that's a bit of
57:07 a simplification because we also have the the Python code where we're using
57:13 the cloud code or the cloud Asian SDK, but like overall it's these two prompts
57:17 that dictate the entire thing. And so that's like if there's any little golden
57:21 nugget takeaway from this stream, it's like just don't over complicate it. You
57:25 can have something really elaborate like this with just a couple of prompts and a
57:28 little bit of code. Um, you probably could even just use the the Cloud Code
57:33 CLI directly, not even have to like have any Python code. I think that the Cloud
57:37 Asian SDK is just really powerful because it allows you to define your own
57:40 workflows and code to do this kind of thing more easily. But still, like it's
57:45 it's not too much overall. Let's go ahead and zoom back out for anyone who
57:49 could have been just recently joining in and kind of just go over the quick
57:52 highlevel overview of what we just covered because we went into a lot of
57:55 detail in a lot of these different sections. Uh, so maybe just kind of give
57:59 them a quick rundown like maybe how they can just quickly get started. Um, and
58:03 then the resources and stuff. Uh, so that way they can say, okay, you know,
58:06 step one is this and then step two is this and this is kind of what we're
58:10 going to be looking at. Um, kind of like Yeah. Yeah, >> that sounds good. I think that'll be a
58:15 good like preface to go into the demo as well >> because that's pretty much where I'm at
58:19 right now where I'm ready to get like actually kick this off so we can see it
58:22 do things in linear and and get the application started. So yeah, th this is
58:27 uh I'll put two links in the chat right now. I'll put the link in again to the
58:30 article from Enthropic that inspired everything and then also a link to the
58:33 GitHub repo that I'm going to be demonstrating um in a little bit here for my additions
58:40 that I built on top of this. So, quick recap or for those of you just joining,
58:45 um, the main problem that we have with agents right now is they don't do that
58:50 well when their context gets overwhelmed. Like 100,000 to 200,000
58:55 tokens, like once you start to get to that range where there's just so much
58:58 information that it is looking at at the same time, it doesn't perform as well.
59:04 And so we need a way to um kick off or sort of like reset the context window as
59:09 we're building out a larger application but still have it so that the agent has
59:15 all the context it needs. So when one agent does a bunch of work and then it
59:19 cuts itself off because the context is getting that too long and then we go to
59:23 the next agent, how do we have it quickly understand the code base and
59:27 what to build next? And that's what this harness is solving for us. And a harness
59:33 all it is a harness is a layer on top of our coding assistant that's just
59:37 prompts. It's just prompts that tell it how we want to work with this system.
59:42 And so what we have here is our initializer agent and its sole job is to
59:46 get the project started for us and define all the features that we want to
59:50 run for or that we want to implement for this very long running tasks. And so our
59:53 goal here is to create a very large number of features so that we can have
59:58 these agents run for a very long time before the application is actually
60:02 complete. And so it creates these core artifacts. These are the files that the
60:07 agents will read every single time we start a new agent so that it can quickly
60:12 catch itself up to speed on what our project is all about, what has already
60:16 been implemented and then what we need to build next. And so we have core files
60:20 like the cloud progress for example where the previous agent has just given
60:24 a summary of what it's done. And so that way we can start the context fresh every
60:28 time but it's not like the agent is flying in blind for that next feature.
60:33 it has a really rich set of context of what has already been done and and so
60:36 like that's like the main problem for us to solve with AI coding in general um
60:40 because we want our coding assistants to help us with these longunning tasks but
60:44 we don't want them to get overwhelmed at the same time because most of the
60:47 problems that you have when you're using an AI coding assistant is just caused by
60:52 the fact that it's overwhelmed or your maybe your description of what you want
60:55 to do is kind of vague but that's kind of a separate problem from what we're
60:58 focusing on here. So, we're kind of assuming with this demonstration that
61:03 you are very specific with your request and like you know what you're looking
61:08 for and like that's why our agent or our app spec file is very detailed is
61:12 because like I said, we're we're expecting that you know ex what to
61:17 build, right? But it's solving this other problem of how can we have it go
61:20 through all of this, right? Like if you were to give this appspec file, like if
61:25 I were just to like go into claude right now and um paste in this file
61:32 and be like go and build this right now, like this is going to fail so miserably. There are
61:41 so many problems with this. We're not giving any rules. We're not outlining
61:44 any kind of process for any kind of validation. And it's way too much that
61:48 we're asking it to build at the same time. And like that's all the problems
61:50 we have to solve and make it so that we're actually managing our context
61:54 effectively. And that's what we're doing with this harness here.
61:58 And so yeah, I think that's like kind of a all I wanted as like a little overview
62:01 here unless you think there's anything else that we should explain, Ray, before
62:03 we >> No, that's perfect. I'm excited to kick this process off. And if you've been
62:07 following along and you've been loving this, I love everyone who's been able to
62:11 kind of follow along and and uh show show their regards. You know, obviously
62:15 thumbs up goes very well for the AL algo. sharing with other friends right
62:19 now. Just letting them know like, "Yo, Cole, Claude Cole [laughter]
62:25 is cooking." And this next part is about to be fire. So, make sure you stay
62:28 locked in for this because we're going to see this in action right now. Yes,
62:32 that's right. And we're going to be using linear. I guess that's the one
62:35 thing I didn't review yet is that I've updated this harness myself so that it
62:39 uses linear to track everything because what Anthropic has shared in their quick
62:44 start is just using local files. And so I have specifically called this project
62:49 old don't use because when it uses the linear MCP to list my projects I want to
62:52 make sure it doesn't think that it's going to continue this project in
62:57 linear. So we wanted to start completely from scratch here. And uh surprisingly
63:01 getting this entire demonstration up and running is really easy. And so the link
63:06 that I already put in the chat to the repo the readme here has the
63:10 instructions to get everything up and running yourself. And then once you go
63:14 through all the steps, all you have to do to kick off this entire process is
63:19 just run a single Python script because this Python script is responsible for
63:25 using the cloud agent SDK to load in the initializer prompt and set in send it
63:29 into the initializer agent and then do the same thing with the coding agent for
63:33 all those iterations going forward. And it's even smart enough to understand if
63:38 the project directory that you reference here as the parameter is a project
63:42 that's already underway. So it'll know like okay this is loop number 10. I'm
63:46 going to go to the 11th coding agent. Like it won't try to restart with the
63:50 initializer agent. And so that's also built into the harness here. And then if
63:54 you want to limit the number of iterations it does um so you don't like
63:57 get rate limited with your max subscription you can do that as well. Um
64:02 but obviously that's not as fun. So, we're going to run it right now uh
64:05 without any kind of limit to its iteration. So, let me go back over to my
64:10 terminal here. And then I'm just going to like call this um I don't know,
64:15 claude AI clone. This is the same folder name that I used in my video on
1:04:18 Kicking off the live demo.
64:18 Wednesday. So, I'll go ahead and kick this off. And then this is going to take
64:23 a good amount of time to get to the point where like we can really see
64:26 something ourself because the initializer agent, it takes a while to
64:31 run. Usually, it's not 10 to 20 minutes like they're warning us here. Um, but it
64:35 still will take a while. And so, this is a good chance for for Ray and I to like
64:39 answer questions that you guys have as well. And so, yeah, probably like the
64:44 second half of our live stream here will be us more just um answering you guys'
64:48 questions and then like watching this go and then um seeing some of like the
64:51 interesting things when they come up like when it first spins up the browser
64:55 and we get to like watch it navigate the browser for example. But we should get
65:00 to that point here because it it um will maybe be like 20 minutes about until we
65:05 get to that point. But just calling this out here. So it notices that this is a
65:08 fresh start. >> So it's using using the initializer agent. And I have it set up so that it's
65:15 going to create 50 tasks in linear. >> And so my Excal diagram earlier, if you
65:21 remember, I had 200 listed out here because that's the default for
65:24 anthropic. But that's another thing that I adjusted in the prompts just to make
65:28 it so it won't take as long. >> Wow. Wow. That's really cool.
65:32 >> And what about rate limits for uh somebody was asking about the max sub.
65:37 Um have they relapsed them recently or what is your take on this?
65:41 >> Um yeah, so I don't really know what's going on with my account. I have cranked
65:46 the heck out of it for months and I have not hit a rate limit. And so this entire
65:51 time I had this running for 24 hours and mind you for 12 of those hours I also
65:55 had this running twice and I did not hit a rate limit and I was using Opus 4.5.
65:59 This is using Opus. It's not using Sonnet. Um yeah because you can see here
66:05 the model is Claude Opus 4.5. Now I don't really know why I haven't
66:08 hit a rate limit if I'm going to be totally honest because there are other
66:11 people telling me that they have hit rate limits very quickly especially with
66:14 Opus. >> I don't know what to tell you. It seems like you have $200 max plan, right?
66:20 >> That That's right. I do have other Yeah, even other people that have the $200
66:24 plan, they're still hitting rate limits um quite often. >> Okay.
66:26 >> So, I don't really know. Like there there's something really weird going on
66:30 where different accounts have different limits. It might be a regional kind of
66:34 thing. A lot of people in the Dynamist community have been suspecting that.
66:40 It's hard to tell. Um, there also is a theory that I'm never going to be able
66:44 to prove, but seems more and more true that like for content creators,
66:48 Anthropic has specifically singled out our accounts and made it so that we have
66:52 a very generous limit so that we don't talk bad about Anthropic. And um, that
66:57 is I'm kind of that's kind of like throwing a dagger at them. Um, so please
67:01 [laughter] take that take that with a >> Now you're going to get very limited,
67:02 bro. >> Right. Yeah. Yeah. Take that with a massive massive grain of salt. Like I I
67:07 don't think that's actually happening, but I'm just like saying like I'm trying
67:11 to theorize like why I've never hit a rate limit in the past. Like it's been
67:14 like four months since I haven't hit a rate limit since July with cloud code.
67:18 So yeah. >> Oh, that's interesting. That is that is fascinating. I I you know that would
67:26 kind of make sense because we test them for a long time and I mean so for me I I
67:31 use a lot of different agents and stuff. So that's the other part of it. It's
67:35 like I go between cursor, droid, cloud code, code codex, like by the time I do
67:38 my full loop, you know, the month is over and I everything kind of resets.
67:43 >> Yeah, that makes sense. That's smart. That's actually what someone else in my
67:46 chat said just now. Um, no rate limits for me, but I regularly toggle between
67:50 OpenAI, Cloud Code, and Gemini. So, same as you, Ray. That's smart. I mean, like
67:53 it's good to try different coding assistants, and I still consider Cloud
67:57 Code to be the best out of all of them. Uh, but that gap is diminishing very
68:01 quickly. [clears throat] Excuse me. So like back in like especially May and June, it was
68:06 like nothing was even close and it was kind of pathetic like how much of a
68:09 monopoly cloud code had. >> Yeah. >> But now with like codeex and um Gemini 3
68:15 and like the IDE anti-gravity uh we're seeing things catch up. Open code is
68:20 fantastic. Factory is fantastic which um I also wanted to so I have the factory
68:24 logo in the thumbnail for the live stream today because it's like your
68:27 thing Ray and I I want to talk about that as well. I know that we're not
68:30 really like using it for the demo here. Um, but also like Factory is um doing
68:35 something very different than this harness to solve the context issue
68:38 because they're focusing on memory compression. And so even though it's not
68:42 in the demo here, I really want to talk about that with you as well. Um, so
68:46 we'll get into that like maybe once I'm done explaining this here and we go
68:48 through some of the questions we have in the chat. >> Cool. Yeah. Somebody was asking if
68:52 there's an easy way to change providers like Wama or Grock for you know this
68:55 harness. >> Yeah. Um, and that's the question I had in my chat as well. So yeah, happy to
69:02 answer that. Um, maybe I want to like finish because I'll do that next. Let me
69:05 let me finish explaining what's going on here and then I'll get on to that
69:08 question because that is something else I really want to talk about is how like
69:12 this is not limited to cloud code. Um, because yeah, like I said, claw code is
69:16 my favorite. Um, but like there's a lot of people that have gripes with it
69:19 because of cost and rate limiting and I mean it has outages sometimes too. And
1:09:21 Live demo: The initializer agent creates tasks in Linear.
69:23 so I definitely want to talk about how like this is agnostic to the coding
69:27 assistant. Um, but really quick I just wanted to call out that our initializer
69:31 here is doing everything to get the project started. So it reads the
69:36 appspec. This is determines when or what it's going to do when it creates all of
69:39 the issues in linear. And you can see that it's making all of these calls to
69:44 the linear MCP server. And it's doing these in parallel by the way. So that's
69:48 why you see like all these done in a row and all these calls in a row. So this is
69:52 happening a lot faster than you'd think because it's just doing it all in
69:56 parallel. And so or like doing it in batches I should say. It's not all
69:59 entirely parallel. And so by the end of this we have everything set up in
70:03 linear. It's going to set up the git repo and everything. And then now we get
70:06 into session two which is the coding agent. And so the coding agent is going
70:12 to get its bearings like we saw. And um I actually have to correct myself. Uh
70:15 Ry, I told you that it doesn't look at the app spec. It only looks at the
70:18 feature list but that's not true. I noticed that in the prompt earlier as
70:21 well. So it does actually read the appspec. So that plus the linear issues
70:26 are as ground truth of what we're building and what's coming next. And so
70:31 you can see that there's also there's one local file that I have here. I have
70:36 like a linear project JSON. The point of this file is to tell it like here is the
70:39 linear project that you should be looking at. So if we go and we look at
70:45 this right here, we have the project ID and the team ID. So these are the two
70:49 parameters that it needs for all of its calls to the linear MCP server. So that
70:54 way it knows where to look. Um, so that way like what I could do is I could
70:58 completely crash this application. Like I could do control C here and end it and
71:02 then restart it and it still would know like okay this is the linear project
71:07 that is associated with this codebase that we're working on autonomously.
71:11 >> That's nice. >> That's so cool. >> And so you can see here that we have a
71:17 third issue now. Cloud AI clone. If we go to the issues, then we've got our
71:22 same meta issue that I showed earlier. And then we have one in progress right
71:25 now. It's initializing the express server for the back end currently.
71:28 >> Let's take a look. I want to see. Yeah. >> Yeah. And each one of these has a
71:31 description. So gorgeous. >> This is where we can edit things if we
71:35 want to add ourselves in the loop, which is another thing that that's um really
71:39 nice with using linear. And then um yeah, I can like change the display to
71:43 be a board here too. I like. [laughter] But yeah, so all these are um marked as
71:48 high priority right now because these are the first things that it's going to
71:52 knock out. I I think what this means um is that like I said, I haven't like
71:55 spent a ton of time with this yet. It really is experimental. I think this
71:57 means that these are the things that it's going to build out before it goes
72:01 to the second coding agent session. Right? So it'll knock out all these
72:05 things. It'll move them all to the done column. Then it will make a comment on
72:11 the meta issue and then go to the third session i.e. our second coding agent.
72:16 >> Whoa. Whoa. Wow. >> Yeah. >> And let's see all the tasks are
72:20 generated here. I'm kind of curious. >> Yeah, there's a lot.
72:24 >> Holy smokes. >> Wow. >> And the and it's not like it's trying to
72:29 do way too much. Like it's not overengineering this application. It's
72:34 just that each one of these tasks is very very granular. For example, this
72:38 one is like just for adding in the ability to edit a message that they
72:42 previously sent and so that you can regenerate Claude's response. Wow.
72:46 >> And so it's just getting very very specific which I think that's really
72:50 good. I always recommend people one of the things that I cover as like a core
72:54 part of my agent coding course which I just finished last week by the way in
72:58 the Dynamis community. I I cover the idea of the PIV loop and that's like my
73:03 core mental model for coding. What what I do with my coding assistants is I plan
73:07 a feature. Oh, actually, you know what I could do is I could pull up the diagram
73:11 for this here. Let me do this quick in my Obsidian Vault. >> That's cool.
73:16 >> So, this this is my diagram for the the PIV loop. So, I always start by planning
73:22 my feature and I do what's called vibe planning. And by the way, this is not
73:25 what we're doing today. This is when like I actually want to be in the loop,
73:29 which I recommend for most things. Um, but I start with vibe planning. And I
73:32 call it vibe planning because I'm more just having an unstructured conversation
73:37 where I explore my ideas with the coding assistant. I think about the
73:40 architecture concepts. I'm asking it questions if there's anything I don't
73:45 understand. And so I get to the point where I'm just on the same page with the
73:50 coding assistant for what I want to build. That's the vibe planning stage.
73:54 And then you take that conversation and you use it to create a structured plan.
73:59 So you go from unstructured to structured, which also like takes a lot
74:02 of mental burden off of you because when you first start a new feature, you don't
74:06 have to worry about like what are my goals? What are my success criteria?
74:09 What documentation do I want to reference? Because at first you just
74:12 want to like get your bearings with the coding assistant like here's what we're
74:15 going to build next. Then you move into having it help you build this kind of
74:19 structured plan that feeds in as like all the context the coding assistant
74:24 needs when it gets into implementing. And it even has like the validation
74:27 strategy for example like we're seeing today so that the coding assistant can
74:31 check its own work with the unit test integration test. This is actually
74:34 perfect timing. You see the browser pop up here. I didn't do that myself. That
1:14:35 Live demo: Agent uses Puppeteer to validate the app.
74:38 is the cloudation SDK using the puppeteer MCP. And so it's currently
74:43 because we're just starting the project. It's literally just calling the health
74:46 endpoint of the API to make sure that the express server that it's working on
74:52 is working. Um so yeah, going back here it moved it to done. Like we're watching
74:55 this all happen in real time right now because it initialized the express
74:59 server and then it used Puppeteer to make sure that it's working, moved it to
75:03 done and now we're moving on to the next thing setting up SQLite. Um, so yeah,
75:05 this >> this is crazy. It's happening real time. >> I know. Yeah. Like I I literally went to
75:11 Puppeteer and then like in the background I like saw it like move this
75:15 to in progress. And so with linear it's like all real time. Like I don't have to
75:20 refresh my page to watch the new things coming in from the coding assistant. So
75:23 I can just like sit here and watch it move everything which which that's also something that I
75:29 thought about doing for a YouTube video is like a really like sped up time lapse
75:34 of it like of like having this up so like we're watching it test its own
75:38 stuff and then like having linear so it' be kind of like a like a split screen
75:41 thing right >> you know how they had play Pokemon for whatever like it was one of the Claude
75:47 code things you should do a 24-hour live stream not you but Claude Cole
75:51 >> Yeah. Yeah, >> Claude Cole as the autonomous agent will
75:58 actually just run in the live stream and then you pop in every so often and check
76:00 up on it and just start talking to it and just say and you just show like
76:04 these uh you maybe use Riverside or something like just let it go like
76:09 >> oh that's I actually love that idea so much. Yeah, cuz that cuz Yeah, I mean
76:14 both would be good like really sped up time lapse on YouTube and then like also
76:17 just having a place for people to hang out and watch for like an actual 24-hour
76:21 live stream. That would be very very cool. >> And then like you can make beats or
76:25 something for it with uh Suno like [laughter] >> Right. Yeah. Oh man, I like that a lot.
76:32 You know what do >> Right. Yeah. Yeah. That'd be really cool
76:37 to do like right before Christmas, like Christmas Eve or like or the day before
76:40 Christmas Eve like 23rd maybe like just do like a holiday like more like chill
76:44 live stream where we're just kind of watching it go and I don't have to lose
76:48 my voice for streaming for multiple hours like I always do, >> right? Yeah. You'll just pop in at
76:53 certain times like top of the hour or something and or whatever like you know.
76:56 >> Yeah, [laughter] I like that idea. Okay, that's cool,
77:01 man. I love that. Thanks, Ray. Yeah, actually I might take this harness and I
77:06 might take that idea too and we we can both do it at the same time or do it
77:08 different times. >> Um, you know, I was thinking about maybe
77:11 making like a little video game, like a little sprite video game.
77:14 >> Um, you know, or make have it see if it can do like a lightweight Pokemon clone.
77:18 >> Um, because that could be really fun. And then it it will build itself and
77:22 test itself live. [laughter] >> Yep. Yep. >> Oh yeah.
77:26 >> Wow. Wow. >> This this is crazy. It's like so
77:31 somebody had asked earlier um about the JSON versus linear. So this
77:36 didn't create any JSON files right from the get or it didn't start from JSON
77:39 files for the tasks. It generated them all straight into linear from the get-
77:41 go. >> That's right. Yep. So if I go back to the codebase here, um it created the one
77:49 JSON file it created was this one >> that I mentioned earlier. So it has the
77:54 context of what project am I working in linear. But you can see in this this
77:58 codebase there is no featureless JSON. So it is generating everything directly
78:04 into linear. So our appspec here as well as the tasks in linear that is the the
78:07 source of truth. >> Wow. >> And so that's very different than like
78:13 if I go to this is the old version of the codebase like directly from
78:17 Enthropic. All of my generations they have this featureless JSON file like
78:20 this one. I have this one from a previous test. >> Um what else do I have? this old one.
78:26 Yeah. Like they all have the JSON file. So that's not in the new one at all.
78:30 >> Ah, so the new one doesn't have that cuz it doesn't need it because it's going to
78:33 put it into a linear. >> Yep. Linear. >> God, this is gorgeous. This is so cool.
78:38 >> It's pretty fantastic. Yeah. And you can see here like we can watch the logs in
78:42 real time. Like it created a comment on I I assume this is a comment on the
78:48 cloud progress or like the meta issue or maybe it comment or might have just
78:50 commented on an issue that it marked as done. But yeah, like calls out like,
78:54 okay, these are complete. Let's go ahead and comment on them, whatever, and then
78:58 mark them as done. So if I go back to linear now, sure enough, the ones that I
79:02 was working on before are now marked as done. So we have three in total, and now
79:07 it's working on getting the cloud API set up with streaming. So it's going to
79:10 do some backend stuff before we get any kind of really cool um look at the what
79:14 it's doing at Puppeteer. So hopefully it'll get to that point, but it just has
79:17 to get the server working. Then it'll start working on the user interface. So
79:20 this application takes care of everything and um this is >> yeah it's so cool. There's one other
79:27 thing I want to call out here. Um yeah so Scott mentioned it again in the chat.
79:31 Yeah one thing I want to call out is that I have made it so that this
79:36 application uses your claude subscription for the actual coding but if you are
79:44 cloning claude specifically you still need the anthropic API key. So when it
79:50 goes into the testing here and uh well here I'll show like the final product
79:53 like when I'm interacting with claude here this is still using the anthropic
79:57 key. So you need it for the application itself but you're using your max plan or
80:04 whatever for the coding part right um however you don't have to create a
80:07 claude clone you can make any kind of application if you don't want to use the
80:11 anthropic API key as well so this isn't like limiting you in any way it's just
80:14 if you want to clone claude and do the exact same thing as me then you will
80:19 also need your anthropic key but if you want to just build some other kind of
80:22 application or you could just change the appspec so that it's using your openi
80:27 key or it's using O Lama. If you wanted to have this whole Chan application to
80:30 work with local AI, you could do that. So, it's just it's only for the app that
80:34 it's using anthropic key. I just want to be like really clear that you can use
80:38 your Mac subscription for all of the coding that's like actually going to
80:41 take a lot of tokens. Like everything we're seeing in the terminal here is
80:45 using our MAC subscription. So, I hope that makes sense, Scott.
80:49 And you can change this however you want. So, um, yes, this now gets into
80:53 the other question that we had about, um, using this whole thing with a
80:57 different provider. And so, let me talk about that really quickly. So, I'll go to the codebase
81:06 Um, Scott said the logs just you just showed was asking for the key.
81:11 Um, it'll it'll find the key, Scott, because I have that I have the key in
81:17 the temp file right here. So, it it'll find it. It just has to iterate a little
81:21 bit to find it. But it's been able to do that for me without a problem.
81:26 >> It should be able to >> finding. >> Yeah, I think we're good. I'll move on.
81:32 We'll see. We'll see if it if it actually does trip up, then I'll just
81:38 add it myself into the temp folder right here. We should be good.
81:42 Okay. Yeah, you bet, Scott. Cool. All right. So, uh, the next big question
81:47 here, um, and then I'm curious if you got any other ones, Ray, but I want to
81:50 talk about making this work with other coding assistants. So, this whole thing is using the Claude
81:58 agents SDK, which is a fantastic piece of technology because it allows us to
82:03 use Claude code in our programs, in our scripts with either Python or TypeScript
82:08 code. And it gives us a lot of flexibility because we can define our
82:13 model. We can change the system prompt for as well like I'm telling it that
82:17 you're a full stack developer building production quality web applications
82:21 using linear for project management. I can change the allowed tools that it has
82:25 access to. So setting more granular permissions for security purposes which
1:22:30 How to adapt the harness for other models (Codex, OpenCode).
82:30 also this repository from Enthropic is a gold mine of just like understanding how
82:35 we can build security into our coding agents because they do like they block
82:39 certain kinds of commands so the coding assistant can't remove folders for
82:43 example. It's really really cool. If you want to look into the bash security hook
82:47 if you care about the security definitely take a look. I'm not going to
82:49 cover that right now because it's more technical. Um, but I do call this out in
82:53 my YouTube video as well because I really appreciate what they've done
82:58 here. Like they they allow all bash commands by default. Um, but they they then have like a a
83:07 hook that runs before it it ever makes a tool call and that checks to make sure
83:12 that like it's not trying to um kill itself or delete a folder. Like those
83:16 kinds of things that you would always want to block. So it's really really
83:20 robust. like they they just put so much thought into this hardness. Um, and
83:23 that's why I wanted to build on top of it and put a lot of thought into it
83:27 myself because there's just so much potential with this kind of thing. Um,
83:31 but anyway, I was getting a little bit ahead of myself here because I want to
83:33 talk about using other coding assistants. Sorry, I just I had to veer
83:39 off for a sec. Um, so yeah, Claude Asian SDK. There are a lot of other coding
83:44 assistants that have their own SDK. the the few that I want to call out really
83:50 quickly. We have um the CodeEx SDK, we have the Open Code SDK, and then we have
83:56 the AMP SDK. So, let me click into all these really quick. Um, so what you could do, and you
84:03 could probably ask a coding assistant to do this in like a single PIV loop, like
84:06 just one iteration. It' probably be able to change the entire thing to use like
84:10 the codeex SDK instead of the Cloud Agent SDK. So you could use, you know,
84:17 GPT5 codecs for your for the LLM that's powering everything here instead of um
84:22 Claude 4.5 Opus. >> Oh. >> And so it operates very similarly. The
84:26 one thing that's a little unfortunate is that um all of these SDKs are only
84:30 TypeScript, not Python. So you would have to completely change the language
84:35 of the harness here. Um, but the harness is so simple, like I said, it's mostly
84:38 just prompts that you probably would be able to like build this entire thing in
84:44 one go with a coding assistant. Um, and so like all the different ways that you
84:49 can use codecs in the CLI or open code in the CLI or CL like you can do all the
84:54 same thing with the SDK, but now you get to control >> cursor has an SDK too. So you can also
84:58 kick off like a Yeah. >> Really? Really? >> I didn't know that. I thought they only
85:03 had a headless CLI, which is a little bit different. It's kind of well I think
85:06 was it uh oh I don't know if it's private or invite only or >> okay
85:13 >> because you can basically run a version of their agent um which is insane.
85:17 >> Okay. >> Uh >> yeah I mean that's really cool if they
85:21 have it but I haven't seen it. So they have a headless CLI which is a little
85:24 bit different because it's not like directly in your Python or TypeScript
85:27 code. >> Yeah I think it was like an agent API. I don't think it's public. I'll find out
85:34 some more information because I I know that some people were getting invites
85:37 for it because they were making their own like SSH >> uh tunneling systems
85:44 >> and they were Yeah, there was like like and I I think basically you can just run
85:48 their agent instead of running it into their app, you can just run it in your
85:51 own app. In essence, you can basically do that here. >> Okay, that's good. Cool. Yeah, Rasma
85:57 said they have an an API but not an SDK. >> Yeah, >> which you could use that as well. It
86:01 would just the code would look a little bit different than using an SDK. So in
86:04 that case, I don't know if you could like change this harness to use cursor
86:09 in one go. It might be a bit more of an effort. >> But all you probably write your own
86:12 framework for it, something. Yeah. >> Right. Yeah. Exactly. You could kind of
86:17 make an SDK, use an API, definitely a lot of SDKs really are just a wrapper on
86:20 top of an API. >> Uh but yeah, you could use AMP, you could use Open Code, and through Open
86:26 Code, you could use Gemini 3. So the Gemini CLI does not have an SDK
86:29 unfortunately. But like if you want to use Gemini 3 because I know like that's
86:33 one of the newer and cool models as well then you can just do it right through
86:36 open code. Um so they have that listed as one of the supported models. Gemini 3
86:39 Pro. >> Oh that is brilliant. >> Kimmy K2 or GL. We could use Quen 3
86:46 coder. Um so I don't really know how we would be able to use like Olama if you
86:50 wanted to use like fully local for this. um because there's not like a coding
86:54 assistant with an SDK I know of >> that like if Klein if Klein had an SDK
87:00 >> that would be cool but they do not they have they also have a headless mode for
87:04 running the CLI but again that's not >> I wonder if open code lets you configure
87:08 like custom URLs so >> they might actually >> that would probably make sense you know
87:13 >> um but then also like >> you have to keep in mind for the
87:18 specific model that you're using is it good with tool use um and and is it
87:21 trained for these types of things? So, I know Kim K2 um came out with a version
87:26 where they're like they can run like a 100 to two like several hundred tool
87:29 calls >> and and it kind of doesn't really lose memory and so forth like that. So,
87:34 >> and where other agents aren't really trained to do a lot of backto-back
1:27:36 Live demo: The agent builds the initial UI layout.
87:37 agents. So, this is just running right now. It just >> Yeah, there it doesn't look good yet,
87:43 but we shouldn't expect it to at first. So, it kind of has just a boilerplate
87:46 for the application running right now. But there we go. We're seeing it work
87:50 with it live and the the resolution's kind of off because the browser didn't
87:55 start full screen. But if I were to like actually go in my browser to local host
88:00 port 5175, we can see that it looks better there. But yeah, so we got light
88:04 and dark theme. It kind of just getting things started here. We don't have like
88:08 a logo yet and there's no conversation. But again, each one of these tasks that
88:12 we have in linear is very granular. So we're just focusing on the main three
88:16 column layout structure right now. So, that's what we're about to see. Um, and
88:19 then it'll start building and like actually talking to the LLM and stuff.
88:23 >> That's amazing. >> Very cool. >> Wow. >> Yeah. And sorry, I have to correct
88:30 myself here because we can use O Lama through open code. You're totally right,
88:34 Ray. And actually, Rasma said in in the chat that he just tested it himself.
88:37 >> Whoa. >> Um, so yeah. >> Wow, that's amazing. Yeah, there's other
88:42 additional considerations if you're already going down this route that
88:45 you'll probably start to discover of like which which open models and if if
88:49 actually anyone has more experience with this, please share it in the chat
88:51 because it's going to help everyone else in this community. Um, you know, have
88:54 you found an open model that you use Olama locally um that can that can work
88:59 for some of these agentic tool tasks? Uh, we're probably just going to start
89:02 finding out in the coming week because this is still so new for all of us. So
89:06 any information you find out, please share it back because that's going to
89:09 help everyone else save that time that you probably spent for half the day
89:13 figuring that out and and you know, we're happy to tell more people about
89:16 it. And um you know, this is super amazing and there's no way that we can
89:19 check all these different permutations. So community is always really great to
89:23 be a part of. Yeah. >> Yeah. I love that. And yeah, it's very
89:28 true. Like I've built applications before where I've tried to support like
89:31 a ton of different large language models and it's always tough because I can't
89:35 test all of them. So people always end up having issues with specific ones and
89:39 it's like yeah I'm going to be honest I never tried that model because I'm not
89:42 like people are like oh this doesn't work with fireworks with llama 3. I'm
89:45 like well yeah I've never tried fireworks with llama 3. It's just not a
89:50 tool that I use that often or if ever. Like I've used fireworks like twice
89:53 before. A lot of people probably don't even know what that is. It's just like
89:57 another LLM provider like open router for example. But yeah.
90:01 >> Yeah. Also another thing to note if you are diving down into some of these types
90:06 of agentic tools even though you have a provider like fireworks that's hosting
90:11 it or Grock or Cerebras there's also a variance in temperatures. So if you
90:15 change the temperature between the different providers they can actually
90:18 give you different results because there's a lot of different
90:20 configurations on the end. So when you're hitting the API, if you just
90:24 leave the default temperature, it may work, it may not work. There may be
90:27 something either in the the model card specifically for that model that would
90:31 say we suggest you use this temperature and it'll follow it or if you hit the
90:36 API, they may have that set already. So those little details actually do matter
90:39 in in terms of quality of output that you get. So that's just something that
90:42 for anyone who's working with these like open source models, just beware. uh
90:46 you're going to go through these additional little quirks that you know
90:50 Claude or some of these bigger uh Frontier labs have already kind of
90:53 configured by default. So that way you don't have to spend that time doing
90:57 that. So just FYI. Yeah. >> Yeah. That is really good to keep in
91:01 mind and it's better. It's different temperature is better for different use
91:05 cases as well. So that's good to know like for more creative work you want a
91:09 higher temperature. For coding you want a lower temperature generally unless
91:12 it's building a front end right. Like it just totally depends. Um yeah, and we're
91:16 making good progress here by the way. Like we have the artifact column and uh
91:20 we have the conversation column on the lefth hand side with the settings that u
91:23 I click this button and nothing happens yet. But again, not surprised because
91:28 there is probably if we go look in linear here, there's probably a task
91:35 specifically to implement the settings. Um like if I just search for settings
91:38 here, yeah, front end settings with tabbed interface, right? like that is an
91:42 upcoming task and so it's of course that button's not going to do anything yet.
91:46 So very very cool. >> So I think you can also monitor this on
91:49 your phone, right? Like if you're on your phone you can look at all the
91:53 >> Yep, that's right. Exactly. Yeah. Remote agent coding. That's another thing that
91:57 I've been very focused on recently. Um actually building out an entire
92:00 application to be able to kick off assistance from where you work. And so
92:04 the goal of that is very similar to this where I want to be able to watch things
92:09 and manage things, [clears throat] excuse me, directly where I work, which for this harness,
92:15 we're not actually kicking off the agent from linear. Um, but with my
92:19 application, we could. And so maybe even combining some [snorts] of the ideas
92:22 that we have here. >> I was watching this this really funny.
92:25 Um, it's not funny, it's actually kind of interesting. These guys who are
92:29 launching a product live and they're doing a vlog series. So, they're based
92:33 out of New York and they're I forget the name of the product, but they're working
92:36 with a pretty big influencer. And he has like almost like his notion but for
92:39 writing type of thing where it collects all of his ideas. And the guys described
92:45 like being so locked in during launch cuz they're just got like thousands of
92:49 users uploading I think a million files within 24 hours to their service. So the
92:54 guy's taking a shower with his laptop like he's trying to hold it and is like
92:58 getting wet and he's >> oh my goodness >> making sure those servers don't go down
93:02 and they're scaling and you know he's basically doing this thing. It's like if
93:06 he had his phone like with linear up or something you know he could just be
93:09 yapping in the shower like hitting a button or [laughter] >> Yeah.
93:14 sounds like they're successful but uh not leveraging tools as much as they
93:17 could be even. >> Yeah. you know, everyone's moving fast
93:21 right now, so it's kind of really cool to see different perspectives and like
93:26 actually seeing you take this idea with linear and like, let me just see if I
93:29 can add this in as the next step, which is what you did right before this live
93:32 stream and this is different from what you did in the video >> cuz you added this additional step in so
93:37 that we're actually seeing this. >> Uh, wow, man. This is this is fantastic
93:42 cuz I can't wait. Well, the other thing too is like I think in linear you can
93:46 share maybe readonly views or something to other people who are not in your
93:50 team. Um or whatever. >> I I don't know. That'd be cool if I
93:52 could. >> Yeah. I I I don't really I don't I haven't really gone too too far on that
93:57 thing, but >> it's interesting because I was thinking about how fun would it be like you
94:03 actually kind of inspired me. It's like I make a 24-hour stream where this is
94:09 literally live and like I'm making a video game and everyone can see the
94:11 progress of the video game being built streamed live and it's just kind of
94:15 running. [laughter] >> Oh, I love that idea. >> Oh man.
94:19 >> Yeah. >> Yeah. Ray, you've definitely inspired me
94:22 to want to do more live streaming. Not even even more than just like these
94:25 ideas you have, but just for how much you do it. And it seems seems like a
94:29 blast. Um and whenever I do a live stream, like I have a lot of fun with it
94:32 for sure. I just typically do it like once a month. But yeah, man, there a lot
94:36 of there's so many cool things that you can do with a stream.
94:39 >> Yeah, we we got a good I I was looking at your chat. A lot of people
94:43 participating. Uh let us know like if you want to see Cole do some more live
94:46 streaming in the in the in the comments below. [laughter] Um you know, I think everyone's kind of
1:34:50 A fun detour into cooking and favorite pizza.
94:50 building and doing community together. It's just really fun to kind of hang out
94:53 and and kind of watch these things work in real time. Uh a lot of us have
94:59 questions. I I I take this opportunity. It's almost like a live meeting. This
95:02 would almost be like what we would chat about privately. It's like, hey, what
95:05 are you working on? You know, what's the latest thing that you're seeing in the
95:08 AI in this SDK? Because I haven't had a lot of time to play around with it. And
95:11 you've like really gone super deep in this. And not only that, but you're
95:16 sharing back the repo. It's like, yes, >> right. >> This is cool. This is cool.
95:19 >> Yeah. >> Yeah. I I always try to be very open
95:23 about the things that I'm sharing. So there are definitely a lot of big things
95:26 I'm working on that are exclusive to the Dynamis community like the remote agent
95:31 coding system and just a lot of like the systems I have around um evolving your
95:37 AI coding workflow. Uh, but for a lot of what I do, like I'm sharing it
95:40 completely open. Um, like what I'm doing here, the live stream with you. Um, all
95:44 of the AI agents that I build on my channel, like I've got a really cool
95:49 hybrid search rag agent coming out on Wednesday that is just going to be
95:51 another one of those templates that's just like fully open sourced. And it's
95:55 really cool because it's using MongoDB, which I've never made content on MongoDB
96:00 before. Um, so just continuing to explore a lot of these other platforms
96:03 that are doing a lot of amazing things for AI specifically. And we see this all
96:08 the time where these platforms that like you never thought like are going to be
96:12 AI native, they've become that. Like Notion is literally advertising
96:15 themselves as an AI platform now. Um, Century is building in like AI
96:18 observability and they used to just be like an application monitoring. Like all
96:21 these things are just leaning fully in on AI. MongoDB is adding stuff in for
96:25 Rag specifically, which is why I'm leaning into that. Like it's just cool
96:30 to like try these platforms and and um share it with you guys because you never
96:33 know like what ends up being like the core part of your stack because all
96:36 these platforms I'm mentioning are also like they're not experimental. They're
96:40 like the real deal like Notion, MongoDB, Sentry like these are the real deal. So
96:45 it's cool to to play around with them, but it's they're they're not just like
96:49 play tools either, right? Like it's just me experimenting and like sharing that
96:52 with you. >> I just saw a bunch of tasks get moved to done. Holy smokes.
96:57 >> No, no. So that that was uh just because the in progress column went away. So it
97:01 moved one thing to done, right? And now it picks up the next thing. Yeah. Yeah.
97:03 Yep. >> Wow. >> So we can technically spawn like sub
97:10 aents and other things too. If it was part of the instruction set, if it
97:14 determined certain tasks can be done in parallel, right? And then we would see
97:17 all of those in progress, right? And then work trees like someone was
97:20 mentioning and and so forth. And like you can take this concept and people
97:24 just go crazy, >> right? Yeah, that's right. Oh, man. Yeah, there's so many things you can
97:30 build on top of this. So, I'm going to keep dreaming and I think like it'll be
97:35 cool to do a lot more streams and videos on this in general. Um, yeah, and like I
97:39 said, maybe combining with a remote aentic coding system, but I got to think
97:43 about like how much I want to do there. We will see. But yeah, it's it's pretty
97:48 freaking cool. Um, and then something Rasma said in the chat, tested the
97:53 remote um agent workflow manager harness today. built a circle clone in a few
97:58 hours in Django. That's very cool. >> I haven't heard I haven't heard Django
98:02 in a long time. [laughter] Wow. Have you Have you used Django
98:05 before, Ray? Have you heard? >> A long time ago. I think I some of my
98:09 internal tools at Apple that I built were based off of that. [laughter]
98:14 >> Okay. Okay, that's cool. Yeah, for those of you who don't know, it's a Python web
98:17 framework. >> And um this is actually the very first framework I ever used to build a
98:24 website. Um, and so I I it's kind of nostalgic like going back to Django.
98:29 Their green documentation kind of looks like Superbase, but I didn't even know
98:32 what Superbase was back then. >> Dang. >> Um, yeah. So, a lot of love for Django,
98:37 but I Yeah, I moved to React and then I did not look back, but I still think
98:41 it's a great framework. >> Yeah. >> Yeah. to get started. It taught me a lot
98:47 about like early like web development and um like I mean because I'm not a web
98:53 developer by trade, you know, most of my my work was like in the system level
98:57 things, you know, low-level stuff to just overall how to build a massive
99:02 operating system and like that type of stuff. And so, uh the web tooling stuff
99:06 was more like, you know, internal tools, dashboards, other things. And it taught
99:09 me a little bit about, you know, how can I pass around data objects between
99:13 different things in a like programmatic way. Um, and then, you know, it wasn't
99:18 PHP, so I was like, okay, this is cool, [laughter] right? As long as it's not
99:21 PHP. >> Yeah. >> Yeah. Oh, man. >> Yeah. And and Django implements um the
99:27 MBC architecture. And I assume that's kind of what you were talking about. It
99:31 helps you just understand how pieces are connected together because like
99:34 everything's in one place. And so you see like oh okay this is how my
99:37 components are injected into the front end that I put it from like this API
99:41 endpoint where I render it or whatever like that kind of thing or like here's
99:44 how I fetch records from the model that I have defined and then pass that to the
99:48 front end and like loop over the things and display it in a table like that kind
99:51 of stuff was yeah really really good learning experience. >> Yeah that's true. Yeah, and I think for
99:56 those who are just joining in, if you're kind of new to this AI coding because I
99:59 know there's a lot of people who maybe haven't done this stuff before,
100:03 >> you know, pick pick a framework like everyone's using, you know, the default
100:06 Nex.js, React, or whatever that's just going to help you get off the ground and
100:10 you're going to put in these reps. You're going to run into the problems
100:12 that we've run into maybe over a decade ago in our coding thing. And that's
100:16 actually just part of the process. That's okay. And, you know, just keep
100:21 learning and keep chugging along. I think the intent, so the fact that
100:24 you're here, it's like you want to build stuff. You want to get your ideas come
100:27 to life. And then the the cool thing is actually seeing it, you know, actually
100:31 work like the way we're seeing it right now. >> That's right.
100:34 >> It feels like you can do way more than you could back in the day.
100:38 >> And um this is just going to be part of the process. And I Yeah, like other
100:43 people who are currently in the chat, I mean, everyone's really good people, a
100:46 community, everyone's just really about helping each other out. So, I definitely
100:49 appreciate everyone who's been kind of stopping in the stream and kind of
100:52 checking things out and saying hello. Yeah. >> Yeah. I appreciate a lot as well. Um,
100:57 and there's actually a really interesting comment in the chat that I
101:01 want to uh bring up and hear your thoughts on as well. Um, Rey Jeff said,
101:05 "Just had a thought. No one is ever going to invent new programming
1:41:09 Will AI stop the creation of new programming languages?
101:09 languages again like they used to, which I think is a little bit of a hot take."
101:13 I think I agree with that. I would have to think about that a lot more though,
101:17 but I think that's true. True. And then Jeff also said Cole stuff is becoming
101:20 the new programming language, which I appreciate that a lot, Jeff. But but
101:23 yeah, Ray, I'm curious about your thoughts. Like do you think that because
101:26 it's interesting like AI coding assistants are so trained on React code
101:31 or or I should say like TypeScript or JavaScript or Python code. So it's like
101:36 if we build a new language, then AI coding assistants are just not going to
101:40 do nearly as well to help us. And if we become more and more focused on
101:44 leveraging AI coding assistance, that means we're less and less inclined over
101:47 time to ever care about any new language. I I assume that's what you're
101:51 getting at, Jeff. I'm curious what you think about that, right?
101:55 >> Um, I kind of have a more optimistic take to be honest.
101:56 >> Okay. >> And I think it just goes back to the
102:01 origins of humans. So like as human beings there was probably one language a
102:05 long time ago and then as people spread out to different tribes. Uh those people
102:09 in those tribes like started to have their own dialects and then change up
102:13 the wording and then as people kept spreading to different regions they kind
102:17 of started to change their tones and and and whoever the tribe leader is, they
102:21 kind of established the language for their use cases. And then uh as people
102:25 went into like you know really native like island stuff a lot of communication
102:29 in those languages were based off of uh their farming and agriculture. So stuff
102:34 dealing with water, fishing, land and like uh language kind of evolves from
102:39 that. If humans continue to do human things, which is what we do, and we're
102:44 interfacing with these AI systems, it may seem obvious that like of course AI
102:48 is going to do its thing, and it's just a reflection of what it's been able to
102:52 take so far, and AI continue continue to be this like general language model, but
102:56 there's going to be other evolutions of what happens from here. Um, and so I
103:00 actually think there's going to be an explosion of languages because there's
103:05 going to be um, in essence interfaces and problems to solve in each of our
103:09 communities that are very specific to our own tasks that are going to be very
103:13 efficient for those tasks too. So right like think about the world uh, you know,
103:17 in in farming cultures and all the types of things that deal with that. They had
103:21 to evolve their language and all of their ways of getting food and and
103:24 survival around that too. So like in essence it's sort of based off survival,
103:28 right? Uh and so even why that's even why there's different tribes cla Open AI like all of
103:35 these people have not necessarily converge into one mega conglomeration.
103:41 and it's specific to different tasks, you know, enterprise level to consumer
103:46 level to, you know, uh, SSH server. Uh, and all these are just like different
103:50 types of interfaces. And at at the end of the day, especially as engineering,
103:53 you always want to find the most efficient language to get the most
103:57 efficient thing across. I, you know, so I I I think it's it's kind of there is
104:02 some narrow scope for maybe the truth of what this person is also seeing. Uh but
104:06 in in the broader sense, I actually think there's going to be an explosion
104:09 of languages. [laughter] >> Yeah. >> Kind of quite the contrary. Um because,
104:14 you know, we're going to want something that's more efficient. Like you're
104:16 you're looking at something right now with linear. Um it's possible that
104:20 that's more token efficient than using JSON, right? Uh until someone comes up
104:25 with a new new language model or something or LLM that can do uh more
104:29 efficiencies around that. It's just kind of my theory uh and just observations
104:34 for things. I don't know. Yeah. >> Yeah. That's interesting. You know, I I
104:38 can't get behind what you're saying because I also think that there's going
104:42 to be an explosion of just the number of large language models that we have that
104:45 are all very useful because they're specialized like you said with
104:48 programming languages becoming specialized. So, I suppose that that um
104:51 really like just drawing a parallel there that makes sense. So, that is
104:55 interesting. Scott said maybe every developer will have their own language
104:58 at some point like however we want to work with our coding assistant
105:02 [clears throat] as long as there's a way that the coding assistant can actually
105:06 understand that new language as it as well as it understands like Python for
105:09 example. But >> yeah, it's pretty funny how people say,
105:14 "Oh, that app is just a, you know, API wrapper around XYZ." And then like
105:18 you've seen those memes where they go all the way back. Well, like
105:20 everything's just a wrapper around the C, you know, actual C programming
105:24 language, you know, like Java and everything else is sitting on top of C.
105:28 So it's like Python, you know, and you know, it keeps going back down to these
105:31 like machine level languages. And you know, I get it. There's like an origin
105:36 and everything kind of evolves on top of that. But I mean, there is an evolution
105:39 for a reason, right? It's like >> right. >> Um, and I think this person is saying,
105:43 "Okay, what if you just cut out all the middlemen and just go from YAP to
105:47 binary?" And is that is that the most effective too in >> Yap to binary?
105:53 >> Nice. [laughter] I love that. Yeah. >> Yeah. The meme that I've seen a lot is
105:57 Salesforce is just a SQL rapper. I've heard that time and time again. But
106:01 yeah, it's very very true. Like all everything is just a wrapper. And so
106:03 >> yeah. >> Yeah. Well, having worked on operating
106:06 systems, I could definitely say there's tooling involved in every language to
106:11 help the language compile. And so whenever you're dealing with systems,
106:15 especially like your, for example, like the phone to the Apple Watch, right?
106:17 There's actually different chip architectures in which the way the
106:21 binary bits get laid down onto the device and there's timing and there's
106:24 all these different components. So that's just for one chip, right? and
106:27 then you start to combine them into a system like Bluetooth and Wi-Fi and then
106:32 you have a display driver and then they all have to like operate at a certain
106:36 you know in a little run loop and so you would be thinking to assume it's the
106:39 same run loop that would run on your phone but you know as you start to get
106:43 into the details it's actually not and then uh you also want power efficiencies
106:47 too so that's the reason why your Mac can go for so long your phone can go for
106:50 so long and your Apple Watch can go for at least 18 hours of battery life right
106:54 uh if it threw everything in from the Mac and wasn't efficient for the the
106:57 watch, then you wouldn't really have that long of a battery life and a Casio
107:03 size thing. So, uh you know, while it says like it's still binary at the end
107:07 of the day, these are specific languages for very specific, you know,
107:10 infrastructure for all the metal that they're sitting on too. Uh and the
107:15 language model uh is going to be most efficient and as we see the context
107:19 window efficiency too. You don't want to rewrite like a compiler to transpose the
107:25 code to work for this thing and then do that logic thing. There's stuff that's
107:29 already written and already works really beautifully uh to interface for one
107:33 thing. And so what I'm saying also in the explosion of languages is you're
107:38 gonna get someone who gets so obsessed about some really specific
107:42 microcontroller and because they have AI now they're going to write the
107:46 additional things that no one else wanted to write cuz they're exhausted
107:49 writing it because it's just too much work. But because they know that
107:52 microcontroller now they can go even deeper than they have done before. So
107:56 now you're going to have even more intelligence from existing things that
108:00 people didn't know you can do. Right? There's a there's a cool video of like
108:04 original iPhone running the tokenizer again and it's like damn you know
108:08 imagine if you had chat GPT running the you know so that is the level of
108:12 coolness that you'll probably start seeing that explosion of people going
108:17 even deeper on things that maybe they overlooked before because they thought
108:19 it was too hard to write by hand. But now with an assistance of a language
108:22 model >> you can expose more hooks to the upper levels. Yeah. Yeah.
108:28 >> I love that a lot. Yeah. Yeah. It's funny. I I stumbled across this YouTube
108:34 live stream yesterday um that only had like five viewers. It was the most
108:37 random thing to get recommended to me, but it was this dude that was doing like
108:41 GPU programming >> like really really like low-level C
108:45 code. Um and he was not using a coding assistant. like this mad lad was just go
108:49 and but like to your point Ry like that's the kind of thing where like we
108:53 can use coding assistance to build this kind of thing like really really complex
108:58 super niche GPU programming like most people would never even dare touch like
109:01 this live streamer and so we can go a lot deeper and I find myself doing that
109:05 a lot as well even where there'll be old applications I'll go to where it's like
109:08 oh yeah I remember this thing that I cut a corner on because it's like h I don't
109:12 want to deal with that and it's like really not that important right like
109:15 making your application just a couple percent better. But it's like, well, if
109:19 I can spend 10 minutes prompting a coding assistant to do that, I might as
109:22 well. And then you do that like a few times in a few of these places and all
109:25 of a sudden your application is, you know, 10 20% more efficient or reliable,
109:30 whatever metric it is. And I think that's that's really really that's a big
109:33 deal. And like even thinking about like testing, right? Like people say like,
109:37 oh, coding assistants are going to make your codebases less reliable because
109:41 it's not the human touching everything. And I think that like there's some truth
109:44 to that, but like one of the things that coding assistants really help you with
109:48 is getting over that mental burden of like testing your code, like writing the
109:51 unit integration tests. And a lot of people just default to not doing it
109:55 because it's so time consuming and the games are pretty incremental. Um, as
109:59 long as you're doing a lot of like manual like user validation and so but
110:02 like using a coding assistant to do that in 10 minutes, um, it makes it so that
110:06 everyone's actually writing their unit tests and like that's built into all the
1:50:08 The importance of automated testing in AI workflows.
110:10 workflows now. But I like this verification loop that you're actually
110:13 showing us here is that it'll also pop in like connected to an MCP. MCP can
110:17 launch a browser. >> MCP now takes instructions from the
110:20 language model and tests itself like it's doing right now live. This was not
110:23 planned, right? [laughter] >> Yep. Yep. >> That's crazy. I love seeing that
110:32 automation. That's beautiful. Wow. Wow. >> Yeah, that is Yeah, it's pretty awesome.
110:37 >> Whoa. And so like this this would be kind of like a part of the integration
110:39 testing that I was talking about. But it's yeah it's like the kind of thing
110:42 like even like browser automation. It's clicking through all these different
110:46 user journeys that you're just not going to test yourself when you're doing all
110:50 these little feature implementations. You know >> how soon until we get this
110:58 harness running in a McDonald's for like 24 hours? Like think about this, right?
111:04 We have Gemini vision and it will take a look every 10 frames of like something
111:08 that's cooking or something like that or some type of quality control metric.
111:12 >> It'll feed back into the loop and then start kicking off other processes.
111:16 [laughter] >> You introduce robotics into the mix. Now
111:19 you have this like autonomous thing. I mean I want people to start really
111:23 cooking on these ideas. Like it's not only just for software code.
111:27 >> It's inputs and outputs. You can plug in physical things. MCP servers are that
111:31 extra connector that you don't have to write code for. Now you you can even
111:37 write your own MCP to feed back that loop from the camera to other robotic
111:41 systems that can feed back and then you have something like linear that can
111:44 track something. So you can track uh you can organize your linear project to say
111:48 okay uh these are urgent issues these are timely things. These are the the
111:51 back burner. Uh I still have a backlog of other things that it's called out but
111:55 I still need to get to. So now as a human, you're sort of either
111:58 reorganizing tasks or have an agent to hand that off for you. So many
112:04 possibilities. It's so cool. Yes. Yeah. It's very cool. And yeah, what you said
112:07 actually a couple people have told me that already when they saw my video on
112:10 Wednesday. They're like, "Wait a second. This could be way more than just
112:14 coding." Um because yeah, like the context problem is a problem that we
112:18 have for any kind of longunning task. It's just coding happens to be the most
112:23 common use case where you have these very long running tasks. But it can even
112:27 be like deep research for example. You could apply this to deep research. Um
112:32 >> I love this comment from Scott. Ray loves cooking. Yes, I do. [laughter]
112:36 [cough] >> Do you like uh cooking like actual food, Ray, or is it is it just um cooking with
112:42 ideas? I for me I'm like I have celiac disease. So I'm actually starting to explore more
112:49 cooking stuff because I want really good tasting food without like the
112:53 compromise. And so uh you know it's always very risky to go out to a
112:56 restaurant with crosscontamination because a lot of people really aren't
112:59 paid enough to care, you know, like they're just paid enough to just do
113:02 their job. And I get that too and I respect that. And so I just have to
113:05 treat my health, you know, as a top priority. And uh so yeah, I've been kind
113:10 of trying to like learn recipes and to get different cookbooks and you know,
113:15 like just trying to figure out like how do I get flavor? And I think it's become
113:19 a whole like chemistry job. [laughter] >> Nice. >> Yeah.
113:24 >> Yeah, that's cool. What's your what's your favorite meal
113:27 right now? >> Oh, right now. Okay. So, I have a couple. So, I'm obsessed with pizza. And
113:33 so before I found out I had celiac disease, >> I actually went to New York and did a
113:37 whole pizza tour. So >> Oo, >> yeah. If you haven't done a pizza tour,
113:41 there's this uh company, I think it's named Scott's Pizza Tours. He's probably
113:44 somewhat famous now or he's blown up several times because of people
113:47 discovering him. >> And so he takes you to the origins of
113:50 pizza and how it came from like Italy in some way and like what they did in New
113:54 York at this specific place where like in the 20s where it was kind of
113:57 invented. Uh, and then takes you kind of through the process all throughout
114:00 different places like different styles from square style and other influences
114:05 all in New York and Brooklyn. So, you take a little bus, like a yellow bus you
114:09 get into. It's so so cool. I highly recommend people check it out,
114:12 especially if you fly to New York. Um, it's a piece of history and and you
114:15 know, me being obsessed with pizza, it's kind of [laughter] uh really cool thing. So, you get to
114:21 like try different sauces and meats and flavors and even shows you how to like
114:24 actually cut a pizza, which is I thought was really cool. And there's these
114:27 different tools to cut them. >> Uh, you know, there's like there's like
114:30 these little micro techniques, but the that's kind of uh one of them.
114:34 >> Uh, I think pizza is kind of one of my most favorites right now. Yeah.
114:37 >> Nice. Yeah, pizza. You can never go wrong with pizza. And actually, my
114:44 favorite flavor of pizza is pesto like tomatoes, sundried tomatoes, basil,
114:49 pesto, and mozzarella cheese. >> It is so good. >> And um yeah. So, have you guys heard of
114:55 the or Ray, have you heard of the green egg before? It might be more like a
114:58 United States thing for those of you watching around the globe, but
115:02 >> yeah, my um parents make green egg pizza. So, it's like really really
115:07 really thin crust and we do like basil um pesto pizza and like s'more pizza.
115:12 >> I never thought about it, but it makes so much sense, right? Cuz you could
115:16 control it so well. And it's basically this giant eggshaped thing for those who
115:20 are kind of wondering. It's almost looks like >> it's like an outdoor crock-pot.
115:24 >> Yeah. Yeah. >> I don't know. Is crock-pot a US thing,
115:27 too? I >> Yeah. Yeah. >> I I think so. Yeah. Crock-pot is like
115:31 you throw everything into it and like eight hours later you can have a
115:35 beautiful like soup and vegetables and you know all the stuff. It just tastes
115:38 amazing. Uh if you throw meat in there, it just gets so tender. It's so
115:42 ridiculous. So, >> making everyone hungry right now. >> Yeah. Right. I was just going to say we
115:46 got to stop talking about food because I'm getting hungry. Oh, man. Yeah. I
115:50 haven't eaten um for a bit here, so [snorts] >> Okay. >> Yeah.
115:53 >> But uh >> why has this turned into a cooking show?
115:57 Yo, bro, [laughter] >> I thought you knew, bro. This is like
116:00 secretly a cooking show. We we do the AI for the cookings. [laughter]
116:03 >> Yeah. Oh, that'd be that'd be another good live stream. Ray and I just
116:07 randomly show up at a kitchen together cooking and you think that it's going to
116:10 be AI, we we'll click bait you guys. It'll be like some title. It's like
116:14 revealing brand new AI workflow. This is AGI. And then you click in and we're
116:19 just like making our beef. >> [laughter] >> That's awesome. Well, yeah, I think um
116:23 we're coming up at the two-hour mark. I think a lot of people have had a lot of
116:26 questions and we've gotten to basically see the the automated process here for
116:29 those who are wondering if you're kind of just stepping in. Uh we are going to
116:33 have timestamps for this. So, you can actually go back and the agents are
116:36 basically running autonomously. So, this is the effective harness for longunning
116:40 agents uh from the engineering adenthropic blog and that's what Cole
116:44 has basically been able to institute here. He's added some tweaks to it which
116:47 is really really interesting. So he took the repo, forked it, and then added this
1:56:48 Ray provides a complete summary of the harness architecture.
116:51 part where you can do linear. And so with linear, you can track the progress
116:56 of these agents through MCP tool calls and that's part of the instruction set.
116:59 And earlier we covered the architecture for what this looks like for those who
117:03 are wondering and how the agent can go go in this feedback loop of like doing
117:06 this massive feature here. So we have the appsp spec that feeds in to the
117:10 initializer agent and this little loop here of how it creates stuff. And the
117:14 thing that we're solving here is basically the effective context window.
117:17 So, you can't just flood it in with a massive plan and just say go. Uh, in
117:21 order to do this really artfully and actually with a lot of detail is to kind
117:25 of split these off into these little sections. And that's kind of where it's
117:29 interesting as a human, you can watch it it in linear, take that one specific
117:33 task, you know, have the the context window in essence cleared out, work on
117:36 that really effectively, and then move on to the next task and hand that off.
117:40 And it's really cool and it's given us tons of ideas for these agentic
117:44 workflows that we could institute uh from literally you know operating a
117:49 McDonald's for 24 hours to um you know like making a video game for 24 hours
117:55 and uh I have now gotten so inspired so I'm planning to take his repo and in the
117:59 next day or two I'm going to just kick off a live stream. Let's just I'm going
118:03 to do this live myself >> and maybe I'll do something fun. Can it
118:07 make like a video game and then make it interactive with the live stream so that
118:11 people can comment in the live stream and then like the game will play itself
118:14 with the agent like updating and making the game. Bro, >> that'd be so cool.
118:18 >> Yeah, cuz why not, right? I mean, I >> Yeah, >> I could do anything I want.
118:21 >> Let's go. [laughter] >> Yeah. So, I think that's what it is. And
118:25 and we appreciate everyone who's just kind of been able to tune in and hang
118:28 out with us for this, you know, experimental stream here and we're kind
118:32 of checking things out. We are at the forefront of this technology. So ladies
118:36 and gentlemen, don't be afraid uh of if you see this and it's way over your head
118:40 or too complicated. Um please kind of take a second step here and try to maybe
118:44 go with us. I know I will be going live. Cole actually has his own community as
118:48 well. So if you haven't had a chance, um I'll let Cole talk about that. So how
118:51 you can join his community because there's a bunch of folks in there who
118:54 are super dedicated who want to do who are working on this stuff all the time.
118:57 And Cole's just dropping gems left and right all over his YouTube channel. Uh
119:01 so definitely we'll check that out. So cool. I'll let you talk more about your
119:03 community and stuff. >> Yeah, I appreciate it, Ray. And thank
1:59:04 Cole discusses his Dynamous community and coding course.
119:07 you for the awesome recap. You really nailed what we've covered today and the
119:11 key takeaways, too. So, yeah, I'll just call out really quickly the community
119:14 here. I'll put a link in the chat as well >> to to Dynamis. But if you guys are
119:19 interested in these things and you really want to stay on the forefront,
119:22 the cutting edge, that's what the Dynamist community is for. And so, a lot
119:26 of what I do on YouTube, I just go even deeper in the community with weekly
119:31 workshops. I have the new agentic coding course that was just released helping
119:35 you think about how you can build your own systems for reliable and repeatable
119:39 AI coding which is a big thing that a lot of people are looking to solve. And
119:43 so yeah really excited that uh actually Raasmus has been putting a lot of
119:46 messages in the chat. He and I have worked on this course together and we
119:50 finished it uh we did like all the recording and editing and stuff in like
119:53 a month and a half. It was an absolute grind but it turned out really really
119:57 well. So this is where like a lot of the things that I've been explaining in the
120:02 stream like the piv loop and validation and how we can evolve our own system
120:05 like I go into that in a lot of detail and help you apply it to your own
120:08 projects in the course. And so that's one of the core parts of the community.
120:13 Um, so yeah, that's where I'm dedicating a lot of my time. Like this and my
120:17 YouTube channel is where I'm dedicating most of my time these days. And just the
120:22 amount of um gems like yeah, like I give a lot to the community as well, but just
120:25 other people like sharing their own projects and ideas and research. It's
120:30 just insane the amount of ideas that come out of the community like literally
120:33 every single day. So yeah, definitely check out Dynamus if you're interested
120:36 in being a part of that and you like the the energy that we've got here too. Um,
120:40 so yeah, I'm definitely like always experimenting with the latest and
120:44 greatest things like this. That's super amazing, man. Like, yeah, I
120:48 think this is a probably good point to wrap up for those, you know, right at
120:52 the two-hour mark here, top of the hour. This project's probably going to keep
120:55 going and um, you know, Cole has that VIP access. He can slide into the DMs.
121:00 [laughter] Uh, you know, we've got this like long running tasks and we'll see. Maybe he
121:05 gets raid limited. Maybe not. You'll have to stay tuned and find out. Stay
121:08 subscribed. [laughter] >> Right. Yeah. There we go. The next task
121:13 is done. It's so I did 10 out of 37 um in our stream here. And yeah, I think we
121:18 did like maybe an hour and 15 hour and a half that is actually running. And so
121:21 like it'll probably be able to complete this entire thing within 24 hours
121:26 because I did reduce it to 50 or I guess 47 instead of um 200
121:31 >> or 57. So yeah, like it's it's it'll probably get through the whole thing um
121:35 by the end of the day here. That's phenomenal. Yeah, that that's
121:40 cool. Thank you so much for sharing this this workflow. This is man, like I'm
121:44 gonna get some food and I'm just going to like start getting back into this and
121:47 start >> playing around. This is >> way too good. And we'll have some more
121:53 questions. So, I think it'd be fun to kind of do some other live streams again
121:56 and we'll probably cover like some follow-ups to this or different topics
121:59 and things. And let us know what you all want to see cuz I feel like, you know,
122:04 we're kind of at the edge of this thing and there's so much going on. And if you
122:07 have any topics that you would like us to cover or like us to see or learn more
122:12 about, definitely drop it in the comments, drop it in whatever you're
122:14 seeing. So, if you're watching this later and you're watching this as a
122:18 replay, uh we will read through all the comments and and want to, you know, get
122:21 that feedback back for you. So yeah, that's what's up. >> Sweet. Yeah. Yeah, Ray and I always go
122:26 through all of our comments. Um, so yeah, definitely. And yeah, it' be fun
122:29 to do another live stream, like taking this even further or just whatever else
122:33 is is new. And what I know that like you guys should be paying attention to. And
122:37 I don't know if it's it's the same for you, Ray, but like one of the main jobs
122:41 that I see for myself is to cut through the fluff for all of you and present the
122:45 things that are actually worth paying attention to. And I just assume that
122:48 that's kind of how you see it as well, Ray, because I can tell from your
122:51 content. That's what you do for people. And it's a really valuable thing. So, I
122:54 work hard every single day to make sure that I'm researching everything and
122:59 condensing it down like here are the things we should care about like this
123:02 kind of workflow. Like when Anthropic released this article that again this
123:06 drove everything for today, like I just knew like all right, this deserves a lot
123:10 of attention and sharing it with you all. >> Yeah, definitely. And I think what I
123:15 also want to do is I'm gonna drop your YouTube link channel here because
123:19 somebody was asking for it. Uh, and let me put >> I can do the same for you too. So I have
123:23 your I have the link to your YouTube channel in the description of the live
123:26 stream already, but I'll go ahead and put it in the chat as well if you guys
123:29 want to check out Ray. He's he does a lot of cool content around AI coding and
123:33 workflows. A lot of the kind of like same vein um that I explore a lot in my
123:38 content, but it's just also really cool that like it's quite different in the
123:41 end. even the tools that we use like um oh man I wish we had time to talk about
123:44 factory more today. >> Oh yeah. >> Um because factory is the coding agent
123:48 that Ray prefers and I use cloud code. He uses different MCP servers like ref
123:53 and exa. I've got my archon open source project but it's like in the end we do a
123:57 lot of the same stuff just using different tools which I mean that's an
124:00 important thing to kind of end the stream off here is like all of this is
124:04 really agnostic to the specific tool. Even this entire harness like I talked
124:07 about a bit earlier we could use a totally different coding assistant. We
124:10 can change the prompts however we need. We could use the Jira or notion MCP
124:15 server instead of linear. Like we're not tied to any specific tool for any of
124:19 this, any specific LLM. We can take these ideas of prompting strategies and
124:23 hardness strategies and apply it to whatever we want for coding and beyond.
124:28 >> Yeah. And I'm interested in checking this out for like Claude skills and a
124:32 lot of people have talked about the front-end designer that I need to check
124:36 out. I'm still kind of behind on that. Uh, I'm interested in bootstrapping my
124:40 agents to learn skills. So, maybe if they learn a skill on the fly, can they
124:44 document it themselves and kind of build up on like that that type of thing?
124:47 >> Yeah, we have so much more to talk about. So, stay tuned. I think this is
2:04:50 Final thoughts and wrap-up.
124:51 just going to be uh more to come, ladies and gentlemen. [laughter]
124:54 >> That's right. >> Cool. All right. Well, thank you, Ray,
124:57 for doing this. Uh, this was an absolute blast. Look forward to doing it more in
125:02 the future as well. Hope that you guys all enjoyed it. And um yeah, I think
125:05 it'd be a good time to wrap things up if you agree here. >> Cool. Yeah, sounds good. I'm going to go
125:09 ahead and uh end the stream here and drop the comments below. Let us know
125:12 what you want to see next. Make sure you like, subscribe, follow, do all the good
125:16 things, you know, thumbs up for the algo and we'll see you all on the next
125:19 stream, ladies and gentlemen. Peace out.