Episode 44: Always-on agents, real-world bottlenecks

[00:00] Hello, can you hear me? Hey Vincent. How was your weekend? Yeah, good, good, uh, and yourself. How you doing? Uh, what did I do? Oh, also, I decided that I did not want to touch the computer over the weekend, or at least, at least not for anything work related and just for whatever I wanted to do. So I ended up on Friday working until, um, whenever, It's, um, you know, working until I reach the milestone that I set out for the week, right? Normally on Friday afternoon, it's easy to say, oh, you know what? I'll just catch up on the weekend. And then you end up, you know, getting distracted and then spending on me. So I sent, I decided to do it, finish it, and then I ended up working until 5 AM, Saturday morning. And then I was used this all weekend. I've embraced the other approach by just letting AI into my life at any point. And uh,

[01:00] with Orca, I'm on the phone, driving my home assistant, Automations and uh, improving them and, I mean, to be honest, Astra 6 has changed my, well, it sounds a bit too crazy to say this, but it's, it's pretty, it's pretty much changed my life in the last, uh, like 10 days. I mean, Astra 6 is, I'm really enjoying OpenAI’s, Astra 6. The results are just amazing to me. I'm just, I'm just in awe. So I'm like, you know, I have like more than a 100 GitHub projects on my, uh, on my GitHub profile. I't know if you've ever looked there. I mean, it's a dark place. It's a lot of... 100 what on your on your... 100 what? There's a 100 like get repositories on my GitHub. Oh, like origin, not fork. Yeah, many

[02:00] of them, many of them suck, but like it's so fun and rewarding to just say, hey, Astrok, how would you improve this project? And it goes, yeah, no, no, no, no, no, no, and then it goes through it. You know, it really gives some great recommendations and I say, yeah, I think you should do this and then you should do that. Um, yeah, so, I'm, I'm, I'm loving Astra 6 is incredible. Okay. interesting. I didn't try Astra 6 Astra 6. Um, I, I'm still, you know, content with Claude, and to be honest, it, it was quite an amazing because my laptop, my MacBooks, they run cafe, Nate, which keeps them awake and they don't sleep. Um, yeah. And so that actually kills, if they are not on power, they die, they drain very quickly. So I brought home another power adaptor

[03:00] today so that both of them can stay online. And what happens is that Claude Code, like I said, I think last time it automatically enabled remote control, but that does mean that whenever I, wherever I am, I get, I get a notification, like, Claude, Claude is stuck. And the only things that I, that stop cloud right now is it'll be as temporary credentials expiring. Yeah, to be honest, I've gone back to access keys for that very reason. And I feel like I feel like someone's kind of... It's a skis with AI. Oh my god. Oh, no, I've got I'm totally red pilled or something like that. Yeah. my personal account anyway. It's just whatever. I should have an agent to watch the spend it every any one time. The, The last podcast, I think I was bemoaning Orca doesn't have a way of me using it

[04:00] from another machine, but it actually does. So I run Orca, with all my stuff on this machine, which is permanently plugged in to power and ethernet. Then I use my other laptop when I'm at home, whatever, and then I just connect to all my sessions through the desktop app. And as well as, as well as the uh, the orca, uh, iOS app. So I have like, I have this. I have this, I have like a really nice, like, I'm organised feeling, because usually I'm working on a client laptop and I have like a 1000000 tabs open and I'm like, and I'm like losing track of my, of my life. Now Orca has completely organised everything and it's just amazing. I love it. Interesting. Orca. But like Orca to me sounds like Kiro. Kind of. You gotta try. I mean, there's a lot, there's a lot, the orca offers, I feel, that,

[05:00] that just, I mean, it's just got a lot of polish. I wouldn't say it's got a lot. It's just like when you use it, it's just streamlined. It's nice. I love it. Okay. Yeah, then the most powerful thing for me was that serendipitously or coincidentally, when I 1st tried Orca, it has like a, like, it opened a Claude tab and then opened up a Codex tab, and it was just like one click, and I, and I did that. I opened up a Claude tab, opened up a Codex tab, and then I just tried the same prompt in both, and I was just, and then when it does that, it opens up a separate isolated work tree. And I was, I was stunned. Absolutely stunned that like that some, in some cases, Codex was doing better than Claude. And in some cases, Claude was doing better than Codex. So then, so now I feel like I got a superpower, because now with Orca, I can say, oh, I want I want these changes, not these changes. I want these changes, not these

[06:00] changes. And, yeah, without getting into details. I'm just putting the models against one another, and that can be, that's just amazing to me. Cool. Um, yeah, I guess, like the reason that I mentioned Kiro and remote control on Claude Code is because, first, yeah, Kiro, I'm not really using it a lot. using it at all, actually, and I feel I'm missing out. I'm still having a lot of Hermes. Like Hermes is doing all the reviews for me. I just added stacked PR support. That means that if PRs get stacked, Hermes is reviewing each PR as part of the stack because otherwise it would review the like the 3rd PR in the stack against the page. What is this? I don't think I've used stack PR. So stack PRs. So basically what happens with AI is that your PRs get really big, right? Um, and and Opus already like months ago was able to slice down the commits really nicely, but ultimately when you're

[07:00] reviewing a PR as a as an open source or a project maintainer, the PR is really just like, you don't usually go commit by commit. I mean, it depends, right? Well, I get my slot centres to carve it up because it can't go over 100 lives. Enough about that. You don't need that anyway. So anyway, what I was saying is that like Opus is slicing up comets pretty nicely. they're reviewable and but I don't think a lot of people review commits like commit by commit. Um, so what GitHub released. is a concept of stacking PRs. That means you can create multiple PRs and one PR stacks on top of the other. And you can get up CLI stack in it. It's an extension, actually, you have to get up extensions, install the stack extension, and it's an API capability. So you can do it on the UI as well, but you can use it with the get up CLI as an extension. You can then you need a stack and then you can stack multiple PRs together. That means the 1st PR, and then

[08:00] you see the PRs as part of a collection. Like if you go into, I can show you... I got deja vu. The problem I have with that is that is that AI tends to make put huge descriptions in. So then aren't you going to have the problem with like, like 10 PR stacked with like, Tolstoy novel inside the descriptions. So let's look here. Here you have a 6 stack PR, which is... Wow, I've never seen this before. This is a 2 stack PR. So this one's maybe a little bit more interesting. This is where after we did a release, there were a couple of errors reported in sentry, and I wanted to fix them. I had related ghetto issues, but I wanted like this to go in, and then the other one kind of dependent on this. So the other one couldn't merge until this one was in. So this one, if you go into the bottom here, you will actually see the stack, right? This is Maine. This is the

[09:00] 1st PR that you're looking at, and this is the 2nd PR that is stacked on this PR. So if you were to go in here, you again, at the top, you see this is 2 of two. You can see the full stack. There's a lot going on here. Vincent. And then you can see at the bottom here, this is the 2nd one as well. So isn't it better? I mean, I'm just I'm just comparing this approach to just breaking up your commits. I mean, Yeah, that's what I started with, right? I said, like, Opus already breaks up your commits, but, um, I think a PR on its own should be able to merge, um, and commit on its own might not. So I think the unit of work is different. Um, it also allows the team to look at like, I don't know if people actually go and look at like each individual commit. Um, it depends. I guess what GitHub clearly had the demand for this. I think the main the main demand

[10:00] here was to allow agentic workflows to create like larger, larger sets of work and then break them down. But then I have this 6 stack here, which honestly, I don't think is much easier to review. Can I assume that all 6 of them need to go in, right? Yeah, so I think I can. I cannot. Yeah, I cannot. I cannot merge an individual commit, it looks like. Okay, PR. I cannot merge on individual PR. So this one has like 5 commits. And all it does is it updates a dependency with a new one, like several major versions ahead. Um, and it's quite complicated and it sets up for the next PR. But what's good is what I find really good is that once this is green, right, once this is approved, you can just ignore because then you don't need to like rebase everything else again. Like this one is done and now you can focus and handle the...

[11:00] Wait a minute, that approval was done by an agent or... Yeah, it doesn't matter. It's clearly not approved by a human yet. So then there's like this one, it's also like, I'm not going to look at this again until, like, I guess the other maintainers are not reviewing it yet because they're waiting for the full stack to go green. So there's a couple of weird stuff, though. Like, this PR sets up some things and then the review says like, okay, this sets up for its load bearing for the next PR. So, like, it might not be important for this PR, but actually, it is very important if you take into context the fact that an XPR depends on some of these capabilities and therefore they should be there. They should be part of this PR. So, yeah. I think that's kind of an interesting way of slicing up your work. I think sex PR. So when I, how do we end up in this? Because you started, you keep talking about Orka. I said, for me, Hermes is still the workhorse that is doing most of my stuff. I just landed PR stacked

[12:00] PRs support. We've talked long enough about stack PRs. What I wanted to get to was that Claude Code, remote code, uh, remote control, kind of has given me, like I, when I'm on my phone, there's 2 things. Either it's Hermes pinging me on discord with a new review of multiple like PRs coming in, or it's Claude Code pinging me in the Claude Code app. And the Claude Code app is amazing, actually. Now, the web app has this, when you look at an agent, you can see the quote diff as well in the browser as well. I had my laptop at home working on something. I was in the office and I was able to see the work going. That was pretty cool. I noticed like that sometimes the code sessions sometimes can show the diffs. But on the topic of discs, I've, I've been in the habit of using this programme called TIG, TIG, so... I switched to Lazygit, though. Ah, it's got too much. Lazy get,

[13:00] I like, I think it's just a pager and lazy get is like a whole ecosystem, which I can't be bothered to learn the key strokes for. But Tig didn't work on certain architecture. Like they didn't build and publish certain architectures, if I remember. Yeah, tickets kind of old. And, and so basically, I use take and take status to basically base up to stage and push my hunks or whatever, and do my commits, everything with this, which is take status. But then when I installed Omarchy, which uses Lazygit, and I didn't want to do additional work of installing Tig, I just aliased the Lazygit status, and I've got pretty much exactly the same thing that I was doing with Tig. Maybe I just need to learn something new. There's another one that I use and take would take blame and then the path, and then I can see what changes were made by who. But that, I mean, VS code gives me that with the, um, if you, if you use the git lens and extension.

[14:00] Oh, I can't stand that extension. It's too much. It's too much for me. You don't need to. Like, the only thing I use it for is whenever I go over a line, it hoovers up, which, which author, who is the last one for the set line, when you hoover it, you can see the PR that touched the line, you can go to the file history and it goes on into the separate extension. I do like that. But I just can't be bothered with extensions anymore. Oh, let's talk about... It bothers with VS code. Anyway. Yeah, so that's another good thing about Orca is that I don't have to jump into VS code. It has like a basic sort of VS code editor differ thing in there and it works for me. I'm happy. I feel like I feel like the web application of Claude Code now kind of starts to be exactly what I used an ID for, like which was to look at the diffs that the agent had been doing so that I can have an overview of what it's been doing. And I mean, I was presently surprised

[15:00] today when I had a laptop at home building something and I was able to just go to the web app or my phone, but my phone is not very nice to look at the get tips. But still, it's quite interesting that they have collapsed everything and you can really dive into, hey, I want to really see what's the difference on this particular file. This experience, yeah. Yeah. So anyway, that was the whole topic of Orca and being remote and using remote. Yeah, using, I think they're called agent development environments, ADEs. I mean, I mean, Claude has has, I mean, Anthropic has their own one. I think I think Codex is pretty much caught up. I was pretty upset with my colleague this morning. I told him this is actually, this turned my laptop into open claw. Cloth code, just behind my back, turned my laptop into open cloth. The whole reason I never install open cloth. on my laptop is because I don't want an agent to be managing my machine. And what happens is now

[16:00] I'm at office and I have an agent telling me, yeah, I ran it on my laptop, it's like, that's not your laptop, that's my laptop. You know, you just, you just completely owned my laptop and I never approved remote control. So I think I think all of the organisations like Anthropic and Codex, they're moving and even Kiro CLI, they're turning these CLIs into basically agents that run on your machine and take ownership of your machine, which I don't like. I mean, I don't like, but I thought you were going to say take ownership of your life because surely when you when you clock off from work, as I was saying last week, when you're brushing your teeth, you get a ping for to do some work thing or, you know, and then you're going, nah, I don't want to be bothered with this sort of stuff while I'm trying to look into my children's eyes. You're brushing your teeth while looking into your children's eyes. That's is that a power move? I'm trying to say that these AI notifications are

[17:00] coming out of our hours and it's a bit insane. But I feel... Okay, let's talk about Omarchy. I think we're done with ADEs. We've brought it up. going into a sensitive area there with too much. Yeah. A colleague shared with me some crazy neo-cities page about how people are anti, um, a march. Is this real because DHH shared it and I thought it was a mockery. I'm not sure if it's is it real? I don't know. I mean, it looks kind of real. It's hard to know nowadays I saw it posted either on, either he posted it on LinkedIn or he posted on the on March's sub on Reddit because I'm, I followed that sub because I'm crazy about what they do. There was also Justin Garrison, who I kind of really, you know, appreciate it. Yeah, he does a lot of like, he has a lot of visualisation, visual explanations on how communities, like pod scheduling works. I think he works on

[18:00] talos. I'm not sure if I'm confusing people now, but I respect him. And he posted a very public message about, well, I'm cancelling DigitalOcean because they decided to sponsor a racist. I'm like, so I had this whole conversation with ChatGPT about where is it, like, acceptable to attribute political opinions to an open source contributor, and where do you separate political, you know, visions? Like, so basically ChatGPT helped me understand that if, so if you supporting that project results in financial gain, for somebody expressing such visions and you do not agree and you do you want to protest, then you should indeed not like avoid using it. But in the case of a mark she, there is no financial gain, right? So that was the, this, that was the hold on. Well, I mean, that was the hold on. Hold on. That

[19:00] was the state 4 months ago, right? So, so DHH, he built a marchy, you know, basically Arch Linux, um, bunch of shell scripts, literally a bunch of shells, uh, and a bunch of Golang and maybe some agents. MD for your agents to help you manage your your operating system. It's not much, right? And a separate maybe. Very, very good, uh, Very good design, if you ask me, very good implementation in a way, like the way. Or much is the thing, right? Is it a distro? I'm not sure. It's kind of like it's, it's, it's a bunch of shell scripts that you installed a couple of, of, of Arch Linux. And it makes, he said that his goal for his company was like, we don't want to pay Apple for overpriced hardware. We can get nice hardware and Linux is ready for the big, like, extreme. And, and, and he basically said, you, oh, it sucks. macOS sucks. Here

[20:00] you go. You know, you use, what is it? Wayland, window manager. Well, I just, I just love how, like, Codex is so much faster on my Arachi laptop. Another thing, I think there's a lot of like benchmarks on file systems or performance on Arch Linux compared to macOS that actually the performance on macOS file system is abysmal. I think it's again... It's terrible. I mean, it's the same story with Windows, actually. Okay, so we agree that Omarchy is kind of a good thing in a way that it really promotes Linux on the laptop and, you know, it kind of, you know, for people that are maybe not not, it's an entry, um, it's a gateway. It's a gateway for people that might not be inclined to adopt arch because they don't want to tell everyone they use arch, by the way. Because they spent several 4 weekend, 4 weekends to make it work. So now they have to make it make it visible. So,

[21:00] so, Omarchy kind of, like, burns an, an ISO, you pop in an USB stick, and before you know it, you, you, you have a useable machine, and all of your macOS capabilities are kind of at your fingertips. You didn't have to do it. Yeah, yeah, the way I understand it is DHHH, uh, doesn't want to be a slave to Apple, and he has, he runs a company, a successful company. He has employees, and if he can run his own OS, which he and get everyone to use it, then he's winning. And I think that's a, that's like, why doesn't other people do this? Yeah, yeah. And I mean, and and kind of what makes me interested in it is that he stepped forward and says, look, every single person in my organisation is using this. So it is for wording. Okay? So to me, it's like great. I want the simple Linux setup because I don't want to run a whole virtual machine just to have a Docker demon. you know? Yeah, I'm sure. I'm sure things will change quite quickly, like, because like

[22:00] device management, MDM is required by any bigger organisation to be compliant. Yeah, so back to the, because you, you, you talk about on, on Archie, you want to talk about the Archie? I step back, where I say the political vision, as long as it is not financial gaining to someone, you can use the project without guilt. And I was happy with ChatGPT, I was like, great, I can use Comachi, right? Not that I would care too much. anyway. So now this whole Omacom thing happened, right? called Omacom, right? Because I have Oma, Omacom, uh, it's it's like a community fund. What's it called? Yeah, I haven't... I wrap my head around the the terminology, yeah. But there's some backing. Yeah, some money. So Omacom is a foundation. I take a non-profit foundation that sponsors on Omarchy. Um, because they have Omar

[23:00] Cube, which was the Ubuntu version, and then you have Oma Omar Chi, which was the Arclinus version, then that they have Omacom, which is the foundation that, you know, manages the communities of the different Oma, um, flavours. And now they created a foundation that has several companies backing it. People pledging. I think six. I think initially when they announced they had 8 or six, um, like a six, each one a million, like DHH within a 1000000, Spotify puts in a million. I think another one. don't know. And then DigitalOcean stepped up and also says we're committing like, I don't know how much to this foundation. Wow. And this is where Justin Garrison posted on on LinkedIn. I guess DigitalOcean is fine with sponsoring racism. Yeah, this... And then a picture, a screenshot of him cancelling the account. And then I went back to my ChatGPT, you know? I was like, how does this, like, how does, how, why, why is this person that I respect and follow? Um, so tightly combining, like coupling

[24:00] whatever DHA says, which I don't agree with, um, then does not want to even use DigitalOcean because they are backing a nonprofit foundation. And ChatGPT was very quick to tell me, but now he's financially gaining from it. Right now they are having employees and now they are paying people. So therefore, we have crossed the threshold. And I was like, oh, man, but... Yeah, but the but that seems to be... What do you call it, perpendicular or adjacent to people calling DHS DHH some sort of right wing, um, you know, like. You know, like, you know, the blog post he wrote is, you know, is, I mean, if you go back to Nazi, uh, Nazi Germany, and you look at the propaganda that the Nazis put out against the Jews, it's the same verban.

[25:00] It's the same words. Come on. I mean... What is it? Wolves among sheep. It's like... I think a lot of, I think a lot of text, even, like, I find it weird. Like, there's a lot of political discourse nowadays, like, like, you deserve this. And I never forget when I visited a concentration camp, which was a harrowing experience, um, there was a German thing like uh, Arbeit Macht Frei. Well, no, about my fry? I don't know, I don't know. I don't know what the Germans were, but in English it was something like, you get what you deserve or something like that. Hmm. And so basically, I feel like, yeah, let's not go, but the thing that shocks me with the marchy is that, people just are just, very negative against it. And, and, uh, they might, they might say the right, it might, they might say, oh, he's got right wing thing. Uh, they might say, like, what was the other thing that uh,

[26:00] someone was complaining about, okay, the money thing. I mean, that's that's a difficult thing. I think there was a colleague of mine who was like, oh, I don't like DHH anymore because he's anti-tester and development. And I'm like, um, like, you know what I mean? The problem I have is that, like, people, To be honest, if I don't agree with somebody. I would say like, okay, cool, you're crazy. You know, that's probably the worst thing. You told me that. You said, you said, I was crazy. That means you didn't agree with me. It means I don't agree with it. No, I don't agree with you on that every point. That's not very respectfully calling Creek. Yeah, well, okay, it's not very respectful, but that's the worst thing I can say that you're mad, okay? But like, I wouldn't, I wouldn't say that like, hey, I don't like you. I just think it's so immature. And that's, yeah, it's like, So,

[27:00] so I was about to comment on Justin's Garrison post something like, um, I think we should be able to decouple the political, uh, vision or, or, like, political statements from an open source contributor, honestly, um, who has created many open source projects that a lot of people uh, benefit from. A lot of people get into programming through Ruby on Rails, before AI, people created their businesses on Ruby on Rails. It was a nice, very quick way to start. Exactly. It was a gateway for programming. It was a lot of people. And now Omarchy is a gateway for people that want to get rid of windows and of macOS and just get like an operating system. DHH is doing God's work in a way. Wow. In one slice, right? Wow. Wow. Okay, sorry. He's not mad is what I wanted to say. He's very insane. Okay. I'm not gonna say anything. I have a lot of opinions about

[28:00] what you just said, but okay. Um, so... And I was about to comment. So I went through all the comments in it. And then I read, like, people were confused, like, what are you talking about? And then I, I, some people were like, I think it's related to this and that and then, and then I saw a comment, which was like, to be honest, and I kind of agree with that. And I saw this a lot from other people talking about the Marchi. It's just a bunch of shell scripts, not even good ones at that. That's what they were saying. It's a fact. I mean, honestly, some I think he's made a product out of a whole bunch of packages. And the product is impressive. I was, I'm impressed. It's fast to install, everything works. Yeah, agree. agree. And like having in front behind it because he also put his money and and created... Yeah, I mean, what's it called? I haven't wrapped my head around like the plug-in ecosystem, which seems to be a marketplace in itself. And fine, you know, fine, fine, fine. Uh, though,

[29:00] yeah, I think, yeah, I think we've talked enough about it. It's a great gateway and I just don't understand. Yes, some people's negativity. So the 1st part was it's a bunch of shell, and most people that install it end up moving on, saying like, hey, that was great, but you know what? I'm actually using I think I might be better off using art directly. I don't like all of the pre-bundled stuff that comes with it. And I think that's a great way on its own, right? I mean, now you've got a working machine and you start exploring further. Yeah exactly. Maybe we should be using PewDiePie. version of it then. It's pretty, but I may have OS now. Yeah, PewDiePie built the whole AI machine like you need to own your own inference. He says we shouldn't be slaves to, but he had his, he started with Linux moding, right? He started with going hard into Linux. After being famous for games. He went hard into Linux, and then he went hard... I like PDP a lot too. Yeah, I mean, I'm, it's, it's something I'm,

[30:00] I'm struggling. I tried to, you know, it, in the intro of my YouTube channel. I think I try to make a video, but I wasn't I didn't really like the video, so I I removed it or something. But like, my whole philosophy, with, with my computing has to be, has been, uh, you know, trying to be independent and trying to, um, foster... Well, I don't, I just, I just want to be independent. Like, I don't want to be a slave to Microsoft. I don't want to be a slave. Good luck with AI. Apple. But at the same time, I'm a bit of a heretic because I guess I'm just looking around. I have my Dell, I have my Dell monitors, but my phone and all the laptops inside are on Mac. Of course, out of sight, I have my thing pad and my raspberry pie and all that sort of stuff. Maybe I'm just, what do you call it? an absolute hypocrite, but I'm a big, Believer having

[31:00] your own infrastructure, and knowing how, you know, if you're on a desert island, like how, how to build it from scratch, is the way I think of it. Like, what do you need? Like, you need this, you need that. You need your terraform pipeline. No, you don't need your terrible pipeline. Or a different island. Stay away from terrifying. Of course, like, what I'm saying here in the light of the AI infrastructure podcast probably doesn't make a lot of sense to people because I'm pretty much addicted to frontier models by anthropic and OpenAI at this point. Now how does that make you independent? No, Kai. You're sucking on the teat of big corporate. So the... Yeah, no, I, I, um, when you were saying that, like being independent and and I was at the computer store this weekend to

[32:00] pick up a laptop that we dropped off a while ago to get fixed, like a motor port issue, then a battery issue, many issues with never buying. I mean, I've had my share of Lenovo problems. Oh, Anyway, um, I don't know for Legion. And basically on the desk, at the maintenance desk, there was this machine, which was a tower, but without a tower. It was just a motorboard with like the liquid cooling on one side. The GPU on the other side, and then the coolers on the GPU had these band displays. So you know how these, like, 0 my god, iPhone, I don't want to talk about iPhone 2. But you have these band displays now at all the rage, right? And it was a band display. Oh, yeah, yeah, yeah. Sorry. No, no, what do you mean? Bent displays. Okay, sorry, I was... Yeah, display that's not flat, but like... I'm pretty you were talking about cooling or something, but yeah, I don't get it. Oh, yeah, so on top of the cooling, they had this display. There was like

[33:00] playing videos of of like the Nvidia logo and other things. And the machine was, I don't know, it was 360000000 Vietnam dong. How much is that in US dollars? Give me that ding dong. Jesus, Kai was... Oh, that's close to 14,000 US dollars, 14. Okay. And I wanted to ask them, like, how much of that? total is the GPU card? Just the GTX, right? RTX. And, um, my son asked them in Vietnamese and they explain, they told him it's like 30%. So 30% of that of the cost is the GPU. I guess the other, you know, 10% will be the memory. And then the rest is the motorboard, the CPU who cares about that and power supply. about the CV. Anyway, so so, and then what I thought was, and

[34:00] even then, for that amount of money, what do you get? I think you get maybe 22 per 2nd or what? No, you should definitely. I mean, I get 50 toes per 2nd on this thing. Of course, I'm not, I'm running Qwen. Yeah, well, apparently 50 tokens per 2nd for one. LLM call, right? For one inference provider, whereas if you are running Cloud, you're running 90 agents in parallel. I mean, like, I guess when it comes to freedom and independence and all that stuff that I'm aiming for. At the same time, I'm, I guess I'm a hypocrite because like, of course, I'm not going to buy 14 K machine and host at my home. Of course, I wouldn't mind, but like, The reality is, is that, you know, when I'm not using that machine, other people need to use it. The only, the only way, uh, this technology is gonna scale so that everyone can

[35:00] use it. Like this, I was, I was thinking of a guy I know in Singapore. I hope he doesn't mind me saying his name, Chian. He did, he's done, he's done hacker web. He done rail router, trainmap.sg. He's done all these, he's done like the trees.SG, where you can see all the trees in Singapore. He's like one of the most talented, Developers I've ever come across in my, in my career. He, his UX and his, his, his product, Ownership levels are next level. So I don't know the reason why exactly, but he's been unemployed for a while. He's kind of like been living off, like, I don't know how he's been living, to be honest, but he's office savings and GitHub, sponsorships and things like this. And, and like, I was just thinking, I mean, I'm probably, I shouldn't mention him by name, but like, I was just thinking to myself,

[36:00] like, he's probably not using AI as much as me because he doesn't have the resources, resources that I have using, you know, I basically have a, a Claude Max, and I have an OpenAI pro or whatever it's called. If he has open source project, you who have a sponsorship, you don't know that. But why what makes you say that? Like, do you think there's, like, he would, you would see more, more, um, code gen? Yeah, like this. I mean, I'm probably, I'm probably making the wrong... A lot of assumptions, yeah. I'm definitely making the wrong conclusions. But like, I do... But I think the thought is still valid, that like there's some people who are not using AI as much as me, because they just don't have the subscription. Mm-hmm. And that's, that sucks. It's why open quote says that they're most popular in like the other count, the um, East Asian countries because of uh, cheap, cheap

[37:00] models like Deep Sea Penguin, right? I think we mentioned that before. Oh, is it easy to set up at the box? So what would you have to set up an open router? OpenCode is their most... I guess because they are independent of OpenAI and an entropic. They are also a gateway for engineers in Southeast Asia to use Quinn and DeepSeek, and they had another model access to the one. So doesn't code do the LLM provision too? How does it all work? I mean, some money... They have, that's demonetisation, I guess, right? Their their their agent, their their harness is open source and free. They have their, um, they have their service. Yeah, when you when you when you start OpenCode, you can just go and put some like money on your wallet and then use the subscriptions via their routers. You were impressed from light. was it light LLM that got acquired for, like,

[38:00] 1000000000s of dollars and how easy... Open Ridge is the one I enjoyed. I mean, I was impressed by. Yeah, I get that. So OpenCode, what, one thing that, I've been pushed in, uh, maybe it's because I'm an ex-user, is uh, that that Groc build product, and it's like really expensive. It's like, I think it's like, it starts at $200 a month. And I'm wondering to myself, Yeah, a lot of people said croc is extremely expensive. But you know, when I'm on Instagram or Facebook or anywhere, I get bombarded with Groc advertisement. They have these social influences, usually pretty girls talking about how they build this or that. And that's not my feet, right? That's my son used my Instagram. Yeah, yeah, of course. So, so the, you know, talking about how they build this or that product, um, using Robbot. Often

[39:00] it's very simplistic. Like, oh, I just had this problem and I have growth pot, you know, do the research and build it for me and blah, blah, blah. And I get it, come bombard it, where I'm like, 0 my god, you know the reason? No, like, I hate it because I, I, I, I run agents myself and, and I guess I've never seen advertisement for Kiro Crew or, um, Hermes, well, yeah, they're open source. you know, like for people to get like the gateway. I guess it's a smart move by by X or by, you know, by Musk. So basically, hey, you know what? Most people don't know how to run a bot or an agent. Right? I actually, I give you an ex- when I build the Kero Crew OpenAI, um, ACP bridge, I send you a propo, um, like a project plan of like, hey, if I provide this as a service where people just click a button and very quickly, you know, you sign in with your ChatGPT subscription, which a lot of people have, then

[40:00] you get like, here you go. There's a little node running. And if you want to run it on your own infra, you click a button and they cut sports... This this definitely provide us last time I looked when open claw was a thing. Yeah, open. Yeah, there's plenty for open claw to do that. I just liked Kiro Crew because of the UI and everything. really nice. And it's also... So, you know, that's probably something that you and I or you or I or both, we should, we should probably be building it because like, But no, Musk did it, right? And then he's blasting it everywhere. Instagram, Facebook, X. And then, you know, I think it's amazing. I think it, I think this is the, this is, I said this to, to, to my team because we're building like agent workflows, workflows where you can invoke agent tasks, and they're triggered by by events. Um, but I think the real, um, future of agentic, the agentic future is every employee has a couple agents and the agents

[41:00] are acting on their own identity, but on behalf of the user. So they have their own identity and they are identified as bots, but they belong to this user. And I see this with Kiro Crew, and I think that's what EDBS says. That's how they, like, they saw Kiro Crew adoption. Like after somebody built it internally, as they called it, kiru claw. It was an open clock. I mean, there's so much momentum. It's like, I don't even feel the need to build it because it's coming. I mean, I kind of have it already and it's just the way the whole industry is moving in a way. Yeah, but I think that's how enterprise adoption is going to go as well. Like every employee within an organisation will have a couple of agents that run tirelessly, even when the user is not on their laptop. It runs on its own infrastructure. It has like its own node that it has full control over. It has an identity that it acts on as, you know, a contributor and get, a participator in Slack. There's some projects like beehive or something. There's some slack alternatives

[42:00] where, you know, humans and boats are both... I'm not actually, I was, I think I would have agreed with your, the identity thing. Now I'm thinking the identity must be attached to a person so that someone owning it and being successful, but yeah. That's what I said. I said there will be a couple of agents that have their own identity, but there are, you know, there's a human responsible for them. Exactly, yeah, yeah, yeah. Okay, yeah. So when you look at Kiro Crew, if you go to the repository, if you create a poor request, there will be a couple of both responses, and often these PR comments will show, I am a bot, and my responsible human is this person. So each one of them, like they have a few, and they're acting on behalf of the humans, they have their loops, they have all kinds of, you know, I mean, I'm willing to accept this future. Though just to rewind it a little bit, I mean, I

[43:00] feel like this is going to be out of reach for a lot of people. Because it's an expensive amount of infrastructure, really. And the fact that, uh, I mean, I'm, we're whining that Grock Builders join a, it starts a join a bucks a month. It's probably just going to get more and more expensive, the time goes on really. So, it's a bit, it's a bit weird, isn't it? Like, like, to get started with computing when I was a kid, was expensive, I think, when I was a kid, my dad, like, spent a truckload of money getting me my 1st PC, like more than £1000 and that must be like £10,000 and in today's money. Yeah, it's... Yeah, it's insane. And but like, but like, there's so many factors now, like, Like, one reason why I had a computer back in South Africa is my dad used to go buy fabric in Singapore, he used to go to Sim Lim, buy me

[44:00] hardware, because South Africa had sanctions and bring them back to South Africa. So what I'm trying to say here is like the future is looking pretty wild because like, you know, China is supposed to not have the, you know, the anthropic and OpenAI models. You're supposed to have a credit card, I suppose. I mean, how does a young kid have a credit card and is able to pay 200 bucks a month or something? That's impossible. There's all these like exclusionary, uh, or there's all these barriers to entry when it comes to becoming an AI engineer, right? Anyway, that's just mine. But, but, so I hadn't talk. Well, my kids have access to my accounts, my subscriptions. Oh, your kids have like a, what I, I had a discussion with someone over the weekend talking about, because my son studied uh, algorithms and and data structures and, and, you know, studied

[45:00] a bunch of, my, I, I got in brilliant.org, so he understands like how an index in the database works. Like, Brilliant Org has a really nice course explaining parabolism. what, sorry? Brilliant.org has a nice, has a nice set of courses around, like engineering fundamentals. talking about the difference between true put and parallelism, the difference between like how would a librarian organise books and like how do you look up the books? Like if you sort them by this by author, but then you want to get them by category, you must have some type of look up mechanism to then go to the right shelf. So basically explaining how indexes work in databases and stuff. It's really nice. really experiences me well. I never quite understood the Dewey system, my local library when I was a kid. Yeah. So I kind of pushed my son hard into engineering. into software engineering. My wife is now pushing him hard to do something more like going into NBA or logistics

[46:00] or something else. Are you for real? Uh, yeah, so I was talking with someone who isn't a software, like, engineering team director. I think CTO were the company here in Vietnam. And we were talking about where do you see the future, like of engineering. And I think I read quite a few articles about companies that laid off engineers, believing that AI can build the code, have largely gone, like, ended up with a bunch of non-architected solutions that are hard to maintain and upgrade and operate, and there will be resurgence in hiring engineers to do, like, a proper architectural review and clean up all of the things that has been delivered. what I've been doing. Yeah. And that's what a friend of mine also said, like, very hands-on. Um, they acquired the company. They're like a young company and they acquired another startup and he said that was basically AI

[47:00] AI slop and the performance issues and bugs and things like that were insane. So he spent weeks and weeks of, you know, refactoring and re-engineering the whole solution. You can do that a lot faster now with AI as well, right? You can you can set up basically black box testing to achieve, you know, to keep the characteristics that you like. Yeah, right. or whatever. Yeah, I like that. So I think that's a future. I think, um, I think that that puts, that puts engineers back on the map and I was telling, I was thinking I need to, you know, see if my son really likes software engineering and if he really likes it, he should go further. You can wear a line, like... Being an AI engineer, we need more of them. There's a future for AI engineers. I mean, there's a fantastic future for AI engineers. I think it's, it's, this is the hard bit. It requires a high level architectural thinking. I don't think, like, AI engineer is just a hype term. I think it's

[48:00] an architectural, like an engineer that knows and understands architecture, um, of software engineering, designing for what change, and the ability to, um, consider operational costs. I think that was always the case with like, you know, SRE and DevOps. They always say that those are people that's supposed to have years and years of experience of operational hand on, that are not just only looking at delivering a feature to production, but actually knows on the operational aspect. Like, you don't just throw the code over the wall. You're also responsible for how it operates. I agree. I mean, the... I think it was said by Hashimoto a week or 2 ago, and it's something that gels well with me. is like, if someone has autonomy, has has the autonomy to build the code, run the code, evolve the code, do all those things with AI, and it's, it's, uh, it's actually quite rare. I find.

[49:00] Super red to someone who has the determination to just drive through problems. Yeah, I think it's a combination. It needs to be, you have the idea of building a product and delivering that product. So for my son who has some coding experience, some AI agent experience. I think for him most importantly is to own a product. and to be able to identify what are the features that I need to build and basically learn how to deliver product features in an, in an, in an, you know. my next question is, surely, hopefully he has a product. Has he built something? I give him assignments. I, we have, I, we have now an Lenovo laptop, which is no longer his because he he got a new one, and I told him, we have 2 controllers. We have a sofa and I couldn't find proper co-op sofa games. I have 2 controllers and I was like, we can play it. It takes two. We can play cult of the, I always,

[50:00] I always say it wrong on purpose to, to, to piss up my kids. It's like cult of the, It's not cultural that goat, which is a really old hacker group. It's it's cult of the that something else. Um, it's called of the goat, or cult of the cow. Anyway, there's this game that's that's like multiplayer. There's, um... There's a few. I don't know entangle, but there's carpet. Anyway, not enough. Not enough. So I told him, I'm going to install Omar on this Linux on this desktop. Sorry, on this laptop. And you're gonna deliver by Christmas to co-op sofa games because he likes playing games and he knows a real engine and how to build games. I must say, OpenAI’s, Astra is amazing for generating 3D stuff. amazing. And my feed on Reddit for the last week since Astra came out is nonstop. Astra built this in a day. Astra Building 14 hours. I've been modelling my new home

[51:00] with it. It's incredible. It's incredible. Yeah, so that's why. So anyway, how is he doing with the 2 co-op games? He hasn't built any yet. So that's why you're asking me, does he have a product to own? Does he have something to deliver? Okay, so that's that's the problem there. Maybe he has December. You also need to scratch your own itch, right? That's that's an important aspect of autonomy. Yeah, he has this idea for a game, but then it never materialises. I send him, there's an AI, game death, subreddit, and they talk about how to generate assets, which models are good with which work... One of the 1st major things I did in my career was make a game, but I'm not sure I can recommend it because I'm just traumatised by making that game. So maybe something easier, like, uh, Oh, I offered them so many options. Like we used to play Valorant together. Like, what if you were in a lobby? Ask him to come with

[52:00] the idea, Vincent, come on. He doesn't come up with anything. So I said to him, it's very simple. You know, scrape this website. I, I, I, I gave him this idea. Like if you're in a lobby waiting for a map, the map loads and you see your lineup. Um, so you have, you know, you have valorances with agents, a bit like um, overwatch. So every agent has to different capab- like abilities. Oh, yeah. So one agent. Each, like they're different roles. So usually you have attackers, I forgot the names of the roles. Anyway, so the point is that when you were in the team, you have certain abilities and certain agents that combine in certain ways on certain maps. So what I wanted is the moment you're in a scrim and you get kind of random, you should be able to very quickly come up with like, oh, this is the map, this is the agents. these are the lineups that you can do with those 2 agents on this part of the map, they can do this lineup together to take over the site.

[53:00] I told do something simple like that. You know, with your phone, take a picture, and boom, immediately you get a couple of lineups. On that note, which is pretty bizarre. But like, in the real world, what agent, like if you were delivering a generic software product, what agents do you think you need? Like, you'd probably want a security reviewer agent, you probably want a, a developer agent, you probably want a reviewer agent. Are we back to these these roles? I actually never use them. And the funny thing is, in Claude Quote today, still, whenever I ask, um, you know, that security review, uh, actions pretty handy. I use it all the time. But then I'm playing, I'm playing the security engineer hat, in my, in my employer. I always just tell Opus, approach this from this angle and then load up your, like, launch an agent with this, with the context tuned for this type of activity

[54:00] and then it just builds it. I don't have dedicated agent profile. agents. uh, oh, I don't know what, this is.agent format, isn't there? I haven't used it, but I know it's there. And I know that that by defaulted chips and explorer and except ships a couple of others. So you don't see the, you don't see the point of being having a sort of a role, role-playing or something like this? I mean, I, I, I, I, I'm on the fence about it myself, really. I can understand that loading a context with a certain set of rules, and I do a three-way review, um, The 1st product that I built was an gentic review with 3 different angles. So it created 3 agents reviewing it from 3 different perspective, from an ops operational. Adversarial and insecurity wise, operational wise, and I forgot the 3rd one, architectural wise, maybe on the code. And those were the 3 reviews that came back. And then the final

[55:00] agent would aggregate the report and post it as a comment on the PR. That was the 1st thing I built for the as part of my probation. But it wasn't with like Asian profiles. It was just like loading up the context. I think the biggest mistake you can make is to tell, to start an agent profile with like, you are an expert, security reviewer. Why do you think that? Because the model we work out of memory, it will assume that it shouldn't ask any questions. It will not verify. It will, it will not go and and say like, Well, get confused what role it is. Yeah, because if you, if you're an expert, you are, you already know everything, you don't have to go and look. So the worst thing you can do is tell an LLM that is an expert. And I told you this, you're an expert. Vincent. I told you this months ago. Yes, I did. I'm pretty sure if you go and search on your on your transcript summary or whatever, even if the full transcript, you will find me saying that this is the worst thing you can do

[56:00] is creating roles. Well, thanks for bringing it up again, because I think your argument is really good, actually. It's not my argument. I read it somewhere. I read it. But what I really wanted to talk about. I think about arguments that I had to pull you up on. It's like, we were talking about rewrites and you said, no, that's one way to kill a product. Why? The, uh... when was that? I think I was just trolling you, I was probably in a bad mood. Okay, I think we talked enough. I need to answer some messages on Slack, to be honest. To be honest, you never talk. I told you I have one topic I want to address. Am I going to help you? I shared you 2 articles. I mean, we talked about one article last time, and it was about the idea of agents running a lot of verifications and self verifications. Therefore, why does the pool request, um, CI, containers integration, need to rerun every command that the agents already

[57:00] ran. Remember that? Sorry, repeat, I got distracted while looking for your message. Ah, there's a ton of WhatsApp, probably people. Let me try and share it. Do you know what? I have I have a notifications disabled everywhere. So every time I'm looking at my WhatsApp was the name like, oh, okay, there's a bunch of things. Oh, not this. So there was this article from... Oh, the one that you shared this morning. Uh, the one that I shared this morning was just a follow-up that I say, hey, this kind of like belonging the same bracket, right? So here we have the original article that we discussed last time, which is CI is now is the bottleneck, and it was posted on the 26th of August by one of the swamp creators, Paul's Tech. And he talks about the concept of having a verification, basically an attestation. I remember, we struggled with the word, say the word,

[58:00] I hate saying that word. Basically, a very, like, a signed a certificate, like, an agent. It can be fully deterministic, right? The agent can only get, um, this this manifest or this certificate that it has ran the test. if you had really ran the test, right? So there's some type of verification that runs. And then you push that up. And basically the CI says, great. The signature valid is valid, I don't need to run a bunch of tests. This is confirmed. The agent runs it. So when I saw this article today on September 14th, not today, today, yesterday, well, today. which is agenda coding is training the CI, which is pretty much the same title here, right? I broke the CI. I was like, great, they're they're going to like... was ahead of the curb by a whole month. Yeah, so I thought they're going to like, you know, explore this and go further.

[59:00] And they talked about how, and I like this diagram, right? It's kind of cool. Um, you know, building was a bottleneck, then the review became the bottleneck, but now you have agantic reviews. So now CI is the bottleneck. And they talk specifically on how many engineers, like they, the code base grew 10 times. The CI jobs increased by 25 times. That's a big number. And so basically how they took a bigger machine, which gave them a bit of a runway of 10 weeks. But then it broke. So then they started charting the jobs so that they are spread across multiple machines. But that only gave them like 29 days instead of 10 weeks. And it's... I like how the progression goes, like, step one increased CPU, step two, shard. So I expected this article to end with like, we don't need CI anymore, right? We have verification by the agent. Yeah, we have exactly what Paul Stec told them to do. Unfortunately, that's not the case. Um,

[1:00:00] they just, I think the end of it is, um, where is it? This is the redesign. The redesign is to basically build a database and memory data store and blah, blah. Honestly, I'm not impressed by these articles by entropic, to be honest. The technical details and the like last time when they they shared about the AI is DLC, I was very disappointed in that article. I think there's a lot more interesting articles out there. We talked about the... On anthropic and his article. Like, uh, inside my, my employer, there's a PDF called uh, 2026 state of AI agents report. And inside that PDF, which you have to like, unfortunately, use a form to download, you know what I mean? Like, you have to give your email address away. The number one bottleneck there is integrations.

[1:01:00] But I find like, like Claude Anthropic is putting out a lot of, like, what do you call it? Slop. Yeah, slop for want of a better word. Like, it's putting out, like, a fair bit of messaging, like, where you're, what you need to optimise for what you need to think about, but I do find it strange, like, in one, in one report, it's like, integrations is the bottleneck. And then another report, no, CI is your ball neck. And it's like, okay, bottleneck, bottleneck, bottleneck. You're selling AI. Exactly. You need more agents to solve the problem. Add agents. So, so I'm disappointed by his article. I send it to you and I thought this would, you know, correlate with the ideas of CI having a different, you know, job. Um, but they're just talking about building a proper, like, CI being a product with um,

[1:02:00] I guess. I guess... The same with, like, organisations like, um, who was acquired, again, windsurf, no, um, by Devin. Devin. Cursor. Maybe cursor. There was this company that is a terraform consultancy company that talks about how they, Yeah, so they came into an organisation like Devon or Cursor to optimise their infrastructure as code. And they described like teralytic terraform configurations on terraform cloud and were costing 1000s of dollars for all the resources that were in them and basically ran hours and were like very brittle and and very painful and so on and so on. All the basic stuff that, you know, if you do infrastructure is code, if you know terraform, um, and then they they highlighted this. I think these articles go onto spacelift because they migrated

[1:03:00] cursor onto space. This is all this talk stuff, yeah. Yeah. I think when I started this podcast. We're talking about talk. Tacos. Tacos, tacos, automation, and collaboration software. But, um, yeah, I mean, maybe it's because we are in this, well, we're doing DevOps, we're doing CICD, we're building mono repos, we're building uh, task runners, we're building caches. We understand the infrastructure's code. We're building, we're optimising for delivering change quickly to production. And then reading organisations that are pure agentic, that are building an IDE or building an inference product that have a lot of researchers, building AI, um, and then, you know, even reading the OpenAI hack, of code artefact being such a, um, not code effect artifactory, having so many 0 days that were able to to being exploited, and then basically no proper uh, sandboxing on the Cubinitis cluster.

[1:04:00] I felt like these organisations that are at the frontier of AI have like hired the best people. They hire the best infro people, the best security people, the best dev-ops people. They would have a proper CICD system. They would have a proper sandboxing mechanisms. They would have air gapped agent training facilities. But, well, that comment is maybe wrong because, I mean, I agree that if an agent has to solve a task we should have access to download packages. So anyway, I just I'm just surprised to read that these these labs. Yeah, it's very much on our turf. I mean, I'm, I do get like uh, this feeling of FOMO that I should be working on uh, bigger, bigger challenges in a way. Because, I mean, I'm just, I'm working in an enterprise with, with quite familiar, Challenges, I would say. I mean, I I would like to work on the frontier, that's for sure. I think it would

[1:05:00] be... If you compare this, okay, let's compare this to an article from Tinder on how DNS broke their Cubinities cluster, or if you compare this from the tech team of at Reddis, sorry, Reddit on how they had to migrate their communities and configuration, these tech block articles are amazing. These infrastructure and technical engineering blog articles are really detailed. Okay, I need you need to maybe share them with me, but like, if I see a headline like how I broke my Kubinichi's cluster. I mean, I'm rolling my eyes because it's easy. That's easy. I've done that. You can you can find a lot of these here. I think, um, cute. I mean, to be honest, I think I'm so, I'm so, like, bored of Cubanese in a way, I just, I find it oppressive. Oh, it wasn't Tinder, it was okay, Cupid, I think. Okay, Cupid. Cupid? Dating websites have

[1:06:00] better tag teams than OpenAI. Wow, what a statement. Yeah, I didn't go there yet. But like if you see thumbnail systems on, on, on porn websites and, and the forefront, you know, they're ahead of YouTube and a lot of these things. There's some crazy technology. I know some sides use, which I... Did you watch the Spotify miniseries on Netflix? Uh, I think maybe a few minutes. like one of the Spotify guys. I actually know personally. So I had a bit of an issue watching it because the character was so, what do you call it, typecast or so weird. It's not like the real person. So I'm I'm not gonna. So there's a Wikipedia article about the network protocol that Spotify developed

[1:07:00] for delivering music as quick as it can. So that I think this episode three, I think it's 4 episodes. One is about the impact Spotify has for artists. First is about the founder on, like, where he come from and how... It goes through each person. Yeah, from like different tech guy. The tech guy. The tech guy episode is really interesting because I watched it and I was like, I didn't want to say it, but I actually know that I know the tech guy. Which, that guy, because they're like four. Uh, I know Andreas N. Um, but is it's, there's somebody right now that is being very, um, like, criticised because he does a lot of articles and posts about their agentic adoption and their delivery of product features with agents and so on. They also build it backstage. I think if you adopted backstage, now that gives you a big,

[1:08:00] big leg up, right? Um, because of um, basically the context engineering part for agents. I think that that really allows you. Really? Everybody hates backstage because it's, it's a mess. Everybody hates it, but I think today it deserves another look because as far as I understand, it has rebuilt itself around MCP and it has rebuilt itself around exactly what agents need, which is like what developers needed from backstage, a front page on where are the services who are the owners, where is my data dog metrics? How can you say this? reports. What is your latest? What is your latest, what do you call it? experience with the backstage? Does your employer run it? No, um, how can you, how can you say this, Vincent, you're shocking me to my core. Next thing you're going to start recommending, I'm out of Microsoft. Wow. That's insulting. I actually shared this with you. Maybe maybe back in

[1:09:00] July or maybe May? Um, basically, Backstage. It's supposed to be this internal developer platform that gives you a view across all of your organisations, all of the microservices and products that are there, where does the code live, who are the owners? What are the services and production, where is the dashboards, all the operational data, exactly what an agent needs, right? So a lot of it, they have rebuilt as MCP service. So when you have backstage already, and then you create MCP servers, and then you have, like, basically it becomes a backend for all of your MCP data, then your agents can fetch everything they need whenever they need to work on something. If they need to build code for a service, they can find all the consumer services, they can find information about the operationals. So I, I, I was looking into backstage as a context provider for agents months ago. Mostly because somebody pointed it out at me

[1:10:00] that they had adopted backstage after I left and MCPs were like a massive enabler on top of vaccine. Oh yeah, for sure. So that's where... So when we talk about Spotify. I know that there's somebody that's very publicly talking about how they are delivering with agents, and I think it's a name that you mentioned, and then you have people like... No, not very active nowadays. So Hashimoto, he mentioned that Spotify is very publicly talking about their identic adoption, but that Spotify is probably the least preferred service that he's using. Like there's tweets out there like that. Like people are complaining. They're talking about how they are delivering quickly with agents, but people are saying the quality and the errors in the bugs. Honestly, I'm using Spotify. I don't have any issues with it. But anyway, the whole point I feel as a product hasn't really wowed me in the last... The reason we talk about Spotify is because there's this interesting Netflix series, they have an actual protocol, because we were talking

[1:11:00] about these deep technical engineering blocks that are interesting, and there's, of course, this Cubnitis failure stories, which is a very interesting collection of those, but as you can see, it's kind of outdated because the last one is 2023. But like, nobody cares about pre-AI. B-A before AI and after AI. That's that's the new timeline, right? It's kind of sad. It's kind of sad. It's kind of sad that big frontier eye laps are putting out engineering articles that are not really interesting to read compared to pre-AI articles. it. Well, it's, I mean, it occupied us for a good half hour. I mean, it is interesting in a way what they're spewing out. It, it, it, the bottle. It occupied us to rent for 30 minutes about how low quality those articles are. For the

[1:12:00] last 30 minutes. Yeah, I agree. I was saying that there's better articles from the community, and I mean, I am not a fan of swim. I do not understand whites need it, but this is actually a very interesting read. I think anyone that basically puts that right, like a caveman, some right, their own words from their own human mind onto the screen. It's probably doing a better job. Someone who's being like, who's owning and being accountable for the stuff that they're writing. Wait a minute. Oh, no, don't quote me now. Wait a minute. Where's your enterprise struggle article? The struggle is real. You're going to point out some AIisms. I'm not pointing all the items. You said that it was the Iisms. I mean,

[1:13:00] this is actually, I'm proud of this document. Yeah. Um, I remember DHC saying something interesting. Oh, yeah, I saw this article. I think I saw his his his podcast before this article. Um, Or maybe I did see it from here. Let's give you credit. I probably saw this from yours. Oh, thanks, Vincent. Wow. Um, but I immediately sent, So remember how I feel spectriven development was not built for, um, like team collaboration, that I was like, multiplayer? Yeah. So D Chase is the opposite. I don't know. I'm going to disagree with the HH here. Well, the the, the multiplayer aspect is still not sold. There was, I, um, do you use Claude artifacts because I only just realised, I don't know who told me. Was it you? I only just realised the other day with Claude artefacts,

[1:14:00] you can, you can comment on them. Collaboratively. And then you can send the comments essentially back to the Claude instance, and then you can iterate like that. So, for example, earlier this week, I was working on a mermaid diagram, and then I got some comments that were shared on artefact. Yeah, you see that, just to the left of share. You can highlight different objects and then comment on them. Hmm. And then you can send, and then you can send it to Claude. So you can get multiplayer experience this way. It's a bit low-fi, but it works. And I was like, yes, this is cool. This is cooler. If you can do this. Oh. No, that's not how it works. Oh, okay. should explore. No, I showed you that you could do this on on Kiro artefacts. that you can you can leave. But

[1:15:00] I didn't know that you could do it. Kira takes some setting up, especially if you want to use OpenAI with your own key. This is out of the box, Claude code, you know, you don't even have to like... Well, it's my argument against Orca. Of course, your argument for Orca is I can use Codex and I can use... I can use cloud or codex. But my argument for against Orca is that Claude Code is giving me all of this out of the box already. I mean, you might be happy living in your anthropic world, but I see the value of using both. Or more than one ecosystem. I'm a heterogeneous kind of guy. I'm sad. I'm sad that it's one by one. Can you send them all at the same time? It's not great. It's very low-fi, as I mentioned, but it's... So this is something that... Oh my god, this is this is frustrating. This is something that Kiro

[1:16:00] does, right? Kiro allows you to put a bunch of comments, queue them up and send them all to Kiro all at once. Instead of like one by one. by the way. Yeah. Well, it's getting there. I guess I guess Claude is working towards... I mean, and like, it's it's so simple. Don't know why it takes them so long. A web URL, but a colleague and asks for some feedback. I think this is probably something that's gonna be more common. I mean, I hope it's going to be more common. I mean, I think I think most better. It's really difficult to get feedback from people in some ways. So display not depth did this like back in March, right? You publish your article. You produce your artefact, and then you get, you can, you put it behind SSO, so you can share it, and then you can sign in, and then you can aggregate the comments. Like as you can see, multiplayer comments, and then you can send it across. So they did this back in March. Well, I'm

[1:17:00] afraid to say that Anthropic is caught up with this. Yeah. Because entropic, I always felt entropic was going to do this. as an enterprise offering, right? Artefacts always felt like you should be able to do this. And initially, when you did share, it was not available. You couldn't do it on team plan. You could only do it if you were in an organisation. Oh, no, I can. Oh, no, I can share it with people in my in the team as well. Yeah, I'm not sure about that. Anyway, let's end the podcast. I really need to get on back onto some things. Well, we cover what I wanted to cover. Um, hey, have you, I know, I think there was a few comments on the last podcast I uploaded. I don't know if you responded to them, but I would appreciate if you could respond to the comments. And thank anyone who got this far. I thank you and I and please be in touch. We'd love to hear from you I respond right now. Vincent can respond to your

[1:18:00] questions. There is, there was one comment you linked to me. There was one comment you linked to. Ah, I remember there was a comment asking why don't we just use something else? Yeah. I didn't haven't used something else. And I was like, but that wasn't the point because Hiro AC like Kiro ACP allowed me to use Codex and I guess it only makes sense if I published it. Unfortunately, I think I missed the window to get my contribution to Kiro Crew merged as I... What does ACP stand for again? Agent client protocol. So the client is a dashboard. The agent is the is the agent. Oh yeah, yeah, yeah, yeah. It's a standard by, yeah. I forgot about that. Oh man. So what's happening. Okay, cool. I'm gonna publish this podcast. What number would it be? You figure it out later. You comment below and think of

[1:19:00] it. Give me an idea what the title of this one should be. What did we talk about? Um, renting. God, okay. I'll ask AI to come up with the title. Okay. See you. Thanks, Vincent. Bye.