The Adam Sand Report - Thursday, July 9, 2026

Well, hello, hello, welcome. If you're watching this livestream today, we are going to be testing out GPT-5.6. We have been out of town, uh, meeting partners, investors, teammates, and I flew home today, and so I have been gone, uh, and but I'm excited to try out GPT-5.6. I have not tried it yet. If you're watching these live streams, I fully expect them to suck. They're not expected to be wonderful. The whole idea is for you to have an opportunity to see how we work, what we're working on, and more or less it's just my opportunity to kind of hang out with you guys. So I like to see who pops up in the chat. I like to see what you guys are interested in. I like to see what you guys want to see and what you guys want to learn about, and also learn stuff from you, believe it or not. So I, I love to get the engagement in the comments to know like what you guys see what you guys are thinking about. For those who don't know who I am, my name is Adam Sand. I am building out a platform for roofing companies, uh, an agentic orchestration layer that goes on top of any roofing CRM— JobNimbus, RoofLink, JobProgress as it used to be called, now it's called Leap, AccuLynx, whatever, right? But, uh, yeah, and so I tend to work in, uh, in some pretty rudimentary lip loops where we use linear and whatever coding agent we're working with and grep tile to try and get some good feedback and improve our coding. But we're also going to try some new things today. So I'm just grabbing my mouse because, like I said, uptown, we're going to get to work. And so if you're just popping in, I would love— I see the people that come in, love to hear, hear where you're from, what you're up to, what brought you here, what you're hoping to learn. And, uh, yeah, you guys will just kind of get a chance to see what I'm up to, and I will occasionally get off the screen so that you can't, um— so obviously I'm not sharing credentials and stuff like that. But yeah, excited to try out GPT-5.6, and so with that, we will jump into it. Okay, move you down here. We're going to share screen, share screen up here. All right, and so without further ado, now I heard that the ChatGPT app actually, or the Codex app, got replaced by a ChatGPT app. I heard something about this. But we'll see. By the way, yes, we're working on iOS app. So that's why you'll see that there. Okay. Codex is now the ChatGPT app. Keep coding Codex, pick up where you left off. Work beyond code. All right, I don't know if there's anything new. Okay, I see. So you can load chat down here now. Scheduled tasks are still here. Is anything, anything new up here? We got a little pet up here. It's all my nightly automations. Let's go see what is in the settings, if there's anything new. No. Doesn't look like it. Okay, well, let's go hop in and see what the Automate— what the scheduler said. We'll slide a little guy over here. Welcome in. I just saw someone else joined. If you're here, I'm working with the new ChatGPT app and the new 5.6. We're testing it out tonight. If there's anything you want to see, anything you want to hear about, or if you're just here to say hello, uh, well, say hello and let me know where you're from. And what you might want to see. All right, so what did we get? Where things stand, Linear's foundation platform had no issues since the last round. The active movement is progress. One pilot wedge is now in progress. Folks are removing dashboard mock store bypasses. ADA 338 ETH. Okay, so again, this one open PR. Okay, so this is me. Okay, so go fix the post-merge 101 regressions in Foundation Client or open a small hotfix PR. Open quest. I like— okay, I usually skipped. That's always the thing with the livestream. I never want to like bore you guys with like reading it, but I, I should always read it, right? Like I always like to read these goal prompts that it writes for me. Fixed app UI context search. Configured GUI with source proceeds. Server to never fall back to the mod context fallback tree limited to gateway-dev or transport-server failures only. Fixed app— so the contacts page does not silently truncate tenants above 100 contacts. All API patchations are reported on those covered. Fixed app portal. Hence the collaboration badges progress rendered from the reverse live integration list. Those must span integrations feature space. Yeah. Okay, so I always like to kind of take these. Should the hotfix be tracked under 88340 or a separate linear issue open for PR101 post-merger follow-up? Should the UI eagerly load all pages or tap out of product maximum with visible pagination. Should 500x gateways hold back the mocks in production or only in local dev? So I don't understand, I don't understand the implications of question number 3. Can you walk me through that? We're gonna go to the new 5.6 and we're just going to try, we're going to try— I like to do Fable low, that's like been my like little aura lately. What's in here? That's effort. Okay, so you have effort and speed in here. Okay, well, what happens if we would speed up? Let's do that. So for those of you just jumping in, we are testing out the new GPT-5.6. Uh, what we do is basically run a little loop, um, and so if the Essentially how that works is we have— I'll just type it out. So basically we have a scheduled automation, right, that runs and it looks at the entire list of all of our linear issues, and then it picks out an issue to tackle. Well, and then it writes a /goal prompt, right? Then we usually do a few turns back and forth just doing discovery on the open questions. And then we fire it off. Now what happens is it does its work, tests itself, then we have a contract completion skill, and then we, you, we go to Claude or another model, and we run the independent verification skill. And then when it's done, it pulls post-PR. If it's go, right, if it's like validated done, it's validated done, then it posts PR, then GrubTile runs its code review, and then it loops and does turns until it gets a 5 out of 5, 5 out of 5 GrubTile score. And then we merge, and then we do a post-merge skill, and then we do it again. So that's what we do. If you guys have questions about that, happy to answer them. Just throw your questions or comments in the chat. If you guys have— if you guys are smarter than me and have, uh, something cool to say or add to this, then that would be super helpful. I don't proclaim to be the best in the world. This is just an opportunity to kind of hang out and then see how we do things. If you're curious what Linear is, it's just a project management tool that I use to keep track of all of our issues. So basically all of our projects break out in here. So then we have projects in here, and then inside of those projects, right, then we have a number of issues. As you can see here, my productivity has been low because I have been away for a couple of days. And so we are just coming back into the world. And so this Linear workspace, you can see Productivity is down, something I love. But Linear holds us to service level agreements, basically keeping track of things, like keeping things on track. And so under here we have basically, you know, where we're at. Multiple team members are in here. Sometimes we snag them, sometimes we don't. And then, you know, every once in a while we say like basically like go rip through Linear and like realign everything. Back to, uh, back to like a, um, how do you say it, uh, you know, get everything realigned again, right? Because we add tasks and stuff like that, and so it kind of turns into a bit of a mess. So then you have to kind of go through and do kind of a project management rip and go clean up all your— clean up all your projects, right? So that's what we do. Would love to see your comments. So what we're doing right now is we got Arlo back. We're just doing a couple turns on questions that I was like, I don't understand the implication of that. So local data, this is useful, right? Because essentially what it's saying is we have— I've created a seed of data from one of our client accounts to say, um, to say, okay, well, I want to, um, basically be able to work with live data, like real actual data from one of our roofing company clients, company doing $10 million a year. You know, this is like real stuff. Deals with sold dates and inspection dates. And I don't have to create all that synthetic data, but what we did was we did strip all the PIs. We changed the phone numbers, we mixed the addresses up— they're real addresses, but we mixed them up— and then we changed all the names to superhero names. But all of the other data points, you know, a couple hundred, uh, properties across— I believe it's about 800 deals— that is all real, but it's stripped of all the PIs. Now, Mervyn, to answer your question, this tool here, this is linear. So this is a very simple, very low-cost project management tool. It's not expensive. It's very agent-friendly, right? So you can, you can basically just add issues here, but you can also add issues with an agent, right? So you can use an agent to do it, and it's a great way to just like keep track of like everything that you plan to do. And so then as you like work through whatever feature sets you're doing, you can be like, okay, that's an issue, add that as another like new company run to a problem. And then you can say add that to an issue, and then eventually you can kind of say, okay, well, I wanna, I wanna go and like clean up my projects into, um, say maybe some initiatives, right? And so it's just a very nice project management tool. I'm still learning it myself, but what I do know, or what, and what I quite enjoy, is this Pulse feature, right? So just keeping us on track here. I do like that. I like that it works very well with agents. So my OpenClaw, my Hermes agent, my Codex, and my Claude code are all connected. But then another one of our developers who prefers, you know, VS Code, right, he's, he's also connected to Linear. And so then we can work together, um, and it kind of keeps GitHub a little bit cleaner. Um, so that's why I hope that answers your question. And then, uh, yeah, Chris, thank you so much for being here, man. Appreciate you. If you guys, um, you know, so I, I just treat these live streams like I don't expect them to be these like hot things that are gonna get thousands of views. Um, I'm not gonna sit here and go, you know, oh, I'm, you know, 5.6 destroys, uh, Claude because I, I ran a benchmark making a 3D video game. Like, I'm not gonna do that. I know, shade towards the guys who do, totally respect those who do, and I watch those guys' videos, to be fair. Um,. But me, I kind of just like to work on what I'm working on. I'm working late anyway, right? I just like to work what I'm working on and then enjoy that time with other people who, you know, might have some kind of feedback to give. Sometimes people learn from it and they get better from it. And so the hope is just I just find like-minded people who are working late. Um, and if you have suggestions to make the live stream better, please do give them to me because it would be cool., to have more people come. So if you do the whole like and share and comment stuff, it helps me, right? Um, the whole subscribe thing, it helps me. And then otherwise, I mean, I just like engaging with you guys while I'm working. Uh, so what we're working here is just saying, okay, what was it asked in this goal prompt that it wrote at the start of this loop, which I started at 9 o'clock when I start the stream? You know, what is this one question that it needs clarity on before it starts, right? And so basically I'm just asking what it means. So the proposed policy is local development, no gateway configured, allow clearly scoped production gateway returns 401k, return the denial to the user, never substitute box, production gateway returns 500, yeah, prefer explicit unavailable error state, only fallback if there's intentional non-sensitive. So, so the key choice is whether production should ever silently show simulated data during a real backend outage. Okay, I agree, no. Now, did I understand the fourth question? Should the PR also address the unpublished local ADA ETL enum unwrap or keep that a separate PR? I agree, no. To question 4, um, keep that as separate. And then one 1, keep it under ADA— what was that, 301? ADA 340. And then, uh, for 2, um, cap it at a product route maximum with page pagination. I'm going to say, please— I should just for clarity, it'll probably figure it out— please rewrite the goal prompt for me. Okay, so we'll let that go. So as you'll see here, I'm using Terra Light So, um, I, I love cranking on the biggest model as much as the next guy, but what I found in my Fable 5 experimentation is that the real aura was using Fable 5 on low and then kind of, and then switching between low, high, and like UltraCode. And it's like UltraCode is for like the big project, but then high was like, because if you went to extra or max, it like overthought things. I mean, you just burn tokens for it to like second-guess itself to only end up in the same place. Um, and then Lo was like super fast and super smart, so I really like that. So I'm kind of playing with Terra right now. I'm probably gonna bang off all these models tonight just to kind of see what happens. So then when I go to the— when I go to actually implement this goal, right, um, then I will probably, uh, I'm going to ask it how complicated this is, and then I want to kind of see like what model do I want to try next, right? So I want to say, based on this, how complex is this task? If you were picking a model, would you say I need to pick the highest UltraCode type model that has a ton of reasoning, or is this, you know, straightforward, basically just a knit? Like on that scale, where would you rate it? All right, so it's a 4 out of 10 implementation, 7 out of 10 in consequence if handled carelessly. Okay, so we will try to go in the lens of like what we were doing in Fable. This is probably a— I would call this like a Fable high, right? But well, maybe we're gonna go a little bit more. We'll go like a— I don't know, you guys put in the chat what you guys want me to do. Should I go SOL crazy on a 4 out of 10 implementation complexity and 7 out of 10 consequence. What do you guys think I should use? I'm— you not need the high software using model. Solid coding model should handle it. I would preserve for review. Okay, I'm gonna go with its suggestion unless you guys throw some crazy in the comments. If you guys want to see me burn money, you can say SOL, but we're going to go with SOL high. And yeah, we can keep it on fast tonight. Okay, so we're going to snag this goal prompt, hop over to our actual folder. Okay, we're gonna go say maximum would be this. Should export apply only to the current filter page, which become asynchronous, all resource export later. Export should apply to the current filtered page. It could. What UX should production gateway outlet use? Inline error, full page unavailable state, or retry button? Um, so a production UX for production gateway image, use retry banner. Should the contacts API expose disable cursor contact now or retain offset page number pagination for the pilot? I'm just gonna say, please make best decision for whether— no, why don't I do that? My soul. Hi. There we go. Alright. Let's see what happens. So while we're doing this, I could show you guys something else kinda cool because it wouldn't touch what we're working on. And I'll drag this up right here, and I'll show you guys a little bit show you guys a little bit more front-end stuff, so a little more fun and flash. Okay, while we're, uh, while we're letting that thing work, so we'll go to design systems. So for those of you hopping in, um, we're gonna go and show you what this is, what we're working on for the mobile app. So the concept of the mobile app is that it has 4 modes. So it has a decision inbox, right? So if I go to here, I'll show you the decision inbox. It would let me log in. So, oh, this is, this is an old version. This is not the right version. So the decision inbox is— imagine this, your roofing company essentially runs on decisions, right? So you're going to have— rather than having to go do action, agents can do actions. Like, think if you were standing over a person at the computer and you didn't know how to use the software. Like, just think about that. But this is JobNimbus, AccuLynx, HubSpot, Zooper, whatever, right? You don't know how to run the software, but someone just comes to you and says, hey— and I'll pick on somebody here, I'll pick up— I'll pick on Chris. But Chris, what— there's a storm moving in, heavy rain, looks like it's coming in 2 hours. The radar bubble shows it's going to be over top of these 3 jobs. What do you want to do? Do you want to ignore the problem? Do you want to ping the crews? Or you could chat about it, right? And obviously this will fit in the mobile app, but you could chat about the issue and then make a decision. So you could just say, yep, ping the crews, let them know. Bang. Now the agent's going to come back a little bit later and say the guys think they'll be done in 2 hours. So the agent went and took care of like contacting the 3 crews., right? And it's like, hey, you got 3 certificates of insurance coming up soon, you know. Do you want to send reminders to the crews to update their certificate of completion? Yes, we do. No, we don't. So it's whatever— you just make the decision, the agents take the action, right? That's the, um, that's the idea. And so same thing goes for sales, right? It's like, hey, there's a hot lead 4.4 miles away, do you want to add it to your route? Yes, no. Now it could go to somebody else. You could chat about it, learn a little bit more about the lead. Let's just say yes, right? Here's a proposal going cold, right? Well, then you can chat about it and just be like, remind her that the supplier is raising prices. And so then it will be able to add that to the follow-up, and then you can drag it over here or just send follow-up. So the idea is you just make decisions. So that's the decision inbox. Now, there's 4 different modes for how you interact with it. Now, when you want to switch between the modes— when you want to switch between the modes, you would hold this. And so the mode you just saw was decisions, right? So you would just see you're just in decision mode. That's how you do it. You just unclick it and you go to decision mode., but you can also go to classic mode, which is going to feel like every other, like every other kind of CRM. So this is all this design is, is just all of the— like, this is the light mode, dark mode. This is just the entire design system for the mode switcher. But now, and again, how far— how's this thing doing? Still on step 1 of 6. Okay, so then if we go back to our project files Now you also have, um, should I go into classic mode? Not yet. So classic mode is going to feel like your classic Roofing CRM, right? So again, the whole kind of design system is laid out here, so you have like all the different screens and how you get to them. Right, and again, this is just interacting with your CRM data, but agents are actually working here. So you— so the mode switcher button is always in the center. You can go back to home, right? So home will look like this, or like— sorry, yeah, home will look like this, right? And it's different based on the role, right, for who it is. But you kind of have all of your work surfaced to you, and then you'll have your schedule, you'll be able to search, right? And then If you want to, I guess I would say go over here. Uh, what do I show you next? Capture, I guess, is really what I want to get to, is capture. The whole idea is that at any time it's easy to capture pictures from in here to show you what houses are nearby. Bang, you can start capturing pictures. Those will go back to CompanyCam, right? And so Classic Mode is completely laid out like that. We use— again, always have your mode switcher here. And if you want to add jobs, it's just— you just use the little plus button that you saw right here, right? If you go right here, right, that plus mode, that's how you add things. And so that is the classic mode. And again, all the components are kind of laid out here so you can kind of see the design. Now if we go back to— now again, if you use the mode switcher, right? Then you can get to Agent Feed, which is the one that I'm the most excited about, because the Agent Feed— before I did this, I gotta quickly scroll— before I did this, I thought, what, what, um, what tool or what app, right, what app experience seems to cater to high-achieving, you know, people who like to do hard things and creates camaraderie amongst people who are even strangers? What app does that? Put it in the chat if you can think of what app does that. An app that caters to high achievers, right? People who want to do hard things, who can be competitive but also build camaraderie and community among strangers. What app would you guess that it's? Put in the chat. So it's a, it's a 6-letter, it's a 6-letter app. So put your guesses in the chat. What app gives you a community of people who are overachievers, who try hard? You can practically figure it out if you're looking at the screen, because what it was is Strava. Right? I thought, you know, this app called Strava is this app for people who do hard things, who overachieve, who work hard, who do really cool stuff. And so I thought, how do we create for the 3 primary events in a roofing company— the 'I just finished an inspection,' 'I just sold a job,' right? Could be in the— in at the kitchen table, or it could be after the fact, right? And and I just, um, finished installing a roof, right? And the idea is that if you think about what agents are, right, like, what are agents in this relationship? If you think about Strava for a second, right, when you're looking at Strava, this app, right, takes your watch, the sensor, right, and it is like a full Olympic coaching team in the '80s. Right? Because if you're a Russian Olympian athlete, right, they were monitoring your heart rate, they were monitoring your VO2 max, there was somebody watching your running gait, there was somebody like looking at like taking care of like your hydration levels and monitoring your sleep and all this kind of stuff. And they were trying to optimize human performance. Well, what were they doing? They were, they were, they were taking actions and delivering you data and giving, giving better information to your training, right? That's what that team was. Well, in my example, that is how I view agents in the relationship. So the agents do actions. So the whole idea is that it's like, okay, what do we have here, right? We have like the person, right? What crew did it? And then there's some kind of— the cool thing about Strava is you get to add these little one-liners, right? So if you think about one-liners, it's like they said yes before I hit the truck, beat the rain by 2 hours, or you know rebuild. Like, they put something quippy in there, but then the agents are also doing work. So this is like a news feed of everything going on in the company that keeps everybody excited, right, keeps everybody engaged. But when you— every time you complete a major action, right, you get to put your little— like, you get to put your little one-liner on it, and it makes it so that work can be fun, right? Like, that's the anticipated goal is that like you have this fun that's happening, but then you also see like what agents participated in the deal. So the scheduler agent worked, the nurture agent worked, and then you can give the kudos, the fist bumps. And then you also get kind of like this leaderboard of like what are like some cool accomplishments that people are having, because in Strava you get these like accomplishments and these like badges and stuff like that. So then how do we create, right, this kind of overachiever, high output, do hard things environment that builds camaraderie amongst the team. So that is what this is. Now the final mode— so again, if we go back to the mode selector and kind of revisit that for a second, always pops up this part. Uh, if we go back here, I'm gonna click both. So you've seen Classic,, right? You've seen the agent stream, you've seen the decisions, then there's chat, right? And so that's the last thing is your ability to chat with teammates and the ability to chat with agents. And so we go to chat mode. Nothing revolutionary here if we're being honest, right? It's just essentially Slack, right? It's going to feel very much like Slack, but you're going to have your mentions, your jobs, company channels, direct messages with teammates, and then the ability to actually chat with the agents so you can get work done. And you— so it's kind of like you can bring decision inboxes to chat. That's the whole idea there. But then at any time, you can also use the mode switcher, right, to get back to the other sections of the app. So this is our concept for how we're doing it. So hopefully in that time 5.6 retrying with a faster model. Hang tight, retry with a faster model. Interesting. So this is our first GPT-5.6 prompt, and we're getting told to be retrying with a faster model. So let's see what it says. I wish my little pet told me about that. So y'all are pretty quiet in the chat. We've got a lot of lurkers in the room today. Um, would love to see, uh, hear or see where you guys are from and what you guys are up to and what you guys are interested in. So definitely hop into the chat, let us know where you're from, what you're— what you're interested in. And maybe if you've been using 5.6 since it came out, since I unfortunately flew out of town, um, and haven't had time to play with it till today, if you know something I don't know, share it. Pay it forward, give me the advice, save me the time. I would love to, uh, I'd love— I always love to learn from you guys as well. Interesting. GitHub might be a mess. No, no, it's clean. A lot of work from our friend Gotham. Now interesting, it says it paused the goal. What's this? I learned today claws can go up in the cloud and come down on your other computers. Yes. Yeah, you can. Yeah, you can use cloud, like you don't have to work in local. Right? You can, you can do that right here. Sorry, right here. And you can go in the cloud here, but you can— and then you, with this, you can do it from any other computer. So you can have like a dedicated dev machine, right, that you, that you work in, um, and then you can basically run that one dev machine from anything. And Claude has it as well. And so, Mark, what have you been— what have you been working on with AIestimate.org? Fully autonomous. Very cool. So going fully autonomous, and then you're Using API to pull based on trigger. API to pull what? We know, we know have Windows the moving app. We now have, we now have Windows the moving app. What do you mean? I understand what you mean, Mark. We— I'm assuming you're saying we now have Windows the moving app. Oh yeah, API to pull measurements. So you're connecting to EagleView or NearMap, or what are you connecting to? Have you seen Aether? Or Artemis. Aether will get like pretty good measurements pretty cheap. I can't talk to text and my, my thumbs are too big. That's totally fair, man, I hear you. Um, yeah, it is definitely harder to voice text when you're watching it, watching the live stream. I appreciate you being here anyway. Um, yeah, Aether is basically a, like, very, very, uh, pretty good roof measurements that you can get. And we're talking, like, very fast, right? Um, and then there's another one, Artemis. Basically, the, uh Artemis is another one. These ones— this is, uh, who Roofle is using now, um, but you're getting like at like roof measurements in seconds. So you can buy measurements here for— so yeah, you want to, you want to take a look if you're buying measurements, um you're looking— take a look at Artemis. And so, uh, Mark, would you say that it's like kind of like XBuild? This is an XBuild. Is it kind of the same premise as this? Um, the Artemis is like $6.75 faster. When was the last time you tried this? I guess it says yeah. Professional estimate in minutes, so you're saying seconds. That reminds me of that movie where it's like, it's like, I saw this guy and he was making 8-minute abs. And you know what? I found a way to make millions of dollars. Just think of it, 7-minute abs. We'll be rich. Who wants to do 7? Who wants to do 8 when you could do 7? He's like, who can beat that? And it's like, well, I guess until somebody makes 6-minute abs. It's like, who's gonna make— but how are you gonna get abs in 6 minutes? And I just like— they called you last week, X-Build did, or Aether or Artemis? What's going on over here? This thing done yet? Okay, it's now clean. Okay, you have old cursor. XBuild. Yeah, I mean, they're, they're market— like, they're A16-backed, so they definitely got the dough. You could probably— if you have something they don't, you could probably sell for a quick penny because they are backed by a16z. And the, uh, the founder of XBuild, um, so this is Marc Andreessen, basically the largest venture capital firm to ever exist, um, probably one of the best at predicting the future. But the founder of XBuild used to He had like SimpleCar or something like that, and then prior to— and then he ended up going to work at Uber. So super smart dude. So if you're doing something that they can't, yeah, if you're, if you're doing something they can't, you could probably sell your thing for a quick milli. So what's interesting here is that this 5.6, it it paused the goal, right? So it worked for 10 minutes on the goal and then it paused, and then it went and jumped in, and now it's working on its own, right? So where did it— what did it do with the goal? Is it going to resume the goal? This is like behavior that I've never seen before. Why'd you turn down 3 if you're— if you're getting a solid— like, you turned down 3, or you turned down 3 million? If you got an offer for $3 million and you've— and you— or you've turned down 3 people who wanted to buy $3 million, I don't know why you'd turn away $3 million. That's pretty good return on your time. If you had an offer for $3 million, you should take it. Like, then, like, because then what you would probably get to have be a part of that other organization? What are you going to learn? Like, you can, you can always start another startup. You can always have a new baby bird. But if you get offered $3 million for something you've been working on for a couple of months, you know what I mean? Like, and then you get to be a part of that much— like that, like that venture-backed firm, you're, you're nuts. I'd take— I, I'd say take it, because if you think about it If you think about it, like, I know a roofing software that sold for $20 million, right? But they were in business for 6 years and had tens of thousands of users. So if you build something in a couple of months and you get an offer for $3 million, right, like, that's— well, it's— yeah, I mean, you I get, I get it that you wouldn't want, you wouldn't want a job with them. Like, you could maybe walk away for sure, right? And maybe you could. If you could get $3 million and walk away, take the $3 million and walk away, right? Definitely nothing wrong with that. But the whole idea is that if you can, if you can then take that and then put it like, where do you go next when you've got the $3 million, right? Like the cash return on time and effort is incredible, right? But a lot of founders— and like, I just know this from, like, not from personal experience, I just know this from, you know, watching a lot about what happens with tech. And you think of like Cal.com, right? They sold for a lot of money. But you look at, like— and they sold early. And if you look at Instagram, a lot of people say they would have sold for 2x, but look at Snapchat, they turned down crazy money and they're like dwindling, dwindling away, right? So it's like sometimes if you go there and you take that job, right, then you're going to learn like how does a venture-backed startup work, how does the, how does the, um, like How do the politics work? How does the software development lifecycle work when you've got 10 developers in the same code base, right? What does customer success look like at 10,000 users? What does— what kind of, um, what kind of telemetry is needed to help the sales team continue to push, continue to push, right, and grow? Like, these are life experiences that, um, that could be huge. Plus you're sitting with $3 million in the bank, like you're done, you're good. Like you get a 20-year-old shitbox and a 50-year roof and you got— and you got enough money to say F you to anybody anytime. But that is— that is the risk. I just, uh— oh, interesting, what's this? What is this thing? This is new. So I usually give it a contract completion skill, but it is now running a standards review and a spec review. And so basically, as I understand it, it's spinning up little sub-agents just like how kind of Fable does it. But I've never seen this before in, in Codex. Anybody seen this? You know what this is? Put it in the comments if you, uh, if you've seen it. Okay, the review surfaced two contract issues worth knowing for handoff. The search summary was missing the paid matches aggregate type total, and integration outage banner lacked retry control. I'm tightening the contract now. Okay, very interesting. So for anybody hanging out in the room, if you guys have seen, um, 5.6 and you guys have any tips, um, because it's interesting that it went down to Seoul, like, because I started this goal at, uh, SOL. And then it gave me this interesting error. It was like, it told me that it's been thinking for a long time and suggested going to this faster model. It paused my goal, so I don't know. I, I feel a little like it kind of like went on autopilot without me knowing, or without really like telling me what it was going to do. And it went down to Luna High and I already had fast turned on, we burned up 59% of our context window on a 4 out of 10 technical task and a 7 out of 10 risk task. And so we're just going to see like what actually came out of this, what actually comes. But then it popped in this with the spec review and the standards review. Which is interesting because whenever we do a PR, if you look in here, I go look at a closed PR, I'll go down to one that I did, right? So then we let Greptile review it, right, until it gets a 5 out of 5. So if it doesn't get a 5 out of 5 on the on the first try, then basically the agent listens and loops back, right? And then comes back and like keeps fixing it until it's done, right? So this grep tile is pretty cool. It has this— has the ability to basically run it, run your app locally on its machine and like test it. But then basically when it's done, you get this, you get this report. But it's nice because then you can kind of have a built-in loop So I go here, and one of these you'll see, and it went through multiple rounds. Yeah, this is a big one. So you'll see here, Cryptile, it did a review and then it had fixes, right? So then it gives you a prompt to fix it, right? Or you can go fixing Claude, fixing Codex, and it goes through and did another review. You can see Right here, right? You know, the review, you know, the review, and eventually it got to the point where it was done. So it's nice that it basically runs this, and I probably would have squash merged that to keep it a little cleaner, but definitely handy. I can go to another one that's not so big. Let's do— I want to find you one that doesn't— didn't do T-Rex. All of these T-Rex. Yes, this is confidence score. Yeah, and then sometimes it gives you a really cool design, like a, like kind of like a— you see, it'll give you like a kind of like a, like a data flow model, like a mermaid diagram. It's kind of neat that it does that. I was trying to find one where I could show you that. All right, implement on fresh from Automate. Okay, so Gateway on Synapse, see, now done. Draft PR. Okay, so reptile checks. So this thing just like jumped right forward. It went past contract completion and went right to the push of the PR. So it went right past this goal prompt. So I don't— I guess I can just toss this goal prompt. At this point. And so if we go over here, you guys will see Graphtile live. So now Graphtile is doing its review. And so if we go into Graphtile, I can show you what that looks like. So you just— it takes you right to here. You can see the past scores that it got. And so Codex has an agent in waiting. And then we'll get to see what did it say. So one thing we could do, say, while you're waiting for the grep top to review, can you load the app up in your in-app browser? And before you do so, define your expected function. And then you can also then then run your action and then give me the actual function to see if what you expect to happen in the real world works and functions the way you expect it to. And so it'll probably jump in here. Oh, this is new too. Import, bring over your passwords and cookies in the built-in browser. That's pretty cool. Didn't have that before, but that's kind of neat. I'll probably do that off screen. All right, how's this review coming out? Still grep tiles working? Usually, usually these take like 5 to 10 minutes. So it's also going to run its own, its own test suite right now as well, but for the sake of like demonstration, okay, so it says the peer preview should load the contact surface, contact should request more than Those 35 rows— gateway's unavailable. The UI should show explicit unavailable with retry. Okay. Um, Mark, yes, we're using, uh, Vercel for some things, um, and then Amazon Web Services for most of the backend. Vercel just for the frontend. Man, it's having trouble. I don't know why it would be having an issue. Let's see. Let's try and just get there manually then. Interesting. Interesting. Well, it works here. I have it loaded up in Chrome and signed in. Maybe you can try it on that preview link in Chrome instead of the in-app browser. And now script all doing still testing. Um, Mark, it is an agentic orchestration layer for— that'll live on top of any CRM. And, uh, so this is kind of like the back end of it that you're seeing on the screen right now, but is an agent orchestration layer that will live on top of any, um, any CRM. And then these 23 agents, um, that are built in Amazon Web Services, we will say, as you can see, we can build on— we're building on top of these right now. We're going to add AccuLynx and stuff like that after. We're testing inside environments we already know HubSpot and Zooper, right? But then it will essentially take actions inside the— inside that CRM, but it'll run on, you know, there's a permitting agent, a production agent, a rescheduling agent, a Google Local Service Ads agent, a voice agent, a deal follow-up agent, a production sales to production handoff agent. Um, so essentially it's a suite of employees built on top of— that you can basically connect to your CRM, and then it, it orchestrates on top of it. So it ingests the data from your CRM, and then it can take actions on it, push the information back to your CRM. I think it's not going to be working because it's working inside of a— okay, yeah, it's not working on an account that would actually have contacts in it. Yeah, see, the 10 has no contacts, right? What's grep tell saying? 4 to 5. The integrations error state path needs a fix. Performance— selected app schema or mapping failure can make all backend supported look unavailable. Shared proxy can turn successful non-JSON entry gateway body into— okay, okay, so this is what I was talking about before, the flow state. Integrations page, those problem solve requests, and list integration connections, get integration schema. So that's where it like gets the integration schema. This is where it gathers the schema of your CRM. So how— if you're on AcuLinks, you're going to have like a number of objects. You're going to have contacts, jobs, locations, um uh, estimates, quotes, invoices, payments, right? And JobNimbus, you're going to have contacts and jobs and work orders, right? Actually, in Enactlynx, you have work orders too. So basically absorbs the schema of that and then pulls everything in. And then these are these like error states that it's showing that it has for any failures. But it does perfectly become global outage. So we can say, okay, prompt to fix with AI. So I come back here and say, also, if you pull the PR again, Graphtile has some feedback and some suggested prompts to fix it. Thank you, Mark. I appreciate that, buddy. And so because I interrupted its polling, I think it stopped polling, or maybe I just— it was polling every X minutes. Yeah, there was no active sub-agents anymore. I'm guessing how this works. It's my first time seeing this, but I assume that it sets up this like kind of like Claude code would. It sets up a sub-agent to like pull the PR and wait for grep tiles response. Um, I may have to download the grep tile skills into Codex. I don't know if I did. I know I put them in— grep tile has like a skills file. So if you go to over here and go to grep tile, go to dashboard. And where was it? Settings, maybe. Did it when I set this up. No. Here, by the way, Mark, if you want to try grep-tel, you'll love it. And then I get a free— I get a free month or something. I don't know. Um, there was a place to download grep-tel skills. I could just probably find it. Yeah, right here. No, it's not. It's right— clicking too fast. There it is. Yeah, so I might have to load this into Codex. I may not have loaded this yet because I was kind of doing this as Club Code, but I haven't done it yet in here. Mark, I'll have to watch this video. What is this? Inspired me to give you credit for the new automation we've almost created. I did a video. Awesome. Broke the internet with it today. Broke ours. Just give me access to your account. I'll pay for 3 months. I'm sure an investor would not want me to do that. If there's an investor watching, I did not give Mark Heard access to our reptile. We're okay on the subscription. It's like 50— we have a startup discount, we get 50% off. You would probably get it too, actually. Little hack, go get, go get approved for GrabTile startup, you get 50% off. Okay, I pulled peer 1, other checks are green. Okay, please tackle the P1 and P2 errors and update the code and reply and resolve the threads. Yes, I definitely have to install the grep-tileskills because that's why it's doing this, and then it's going to compact my contacts right now. Let's get— add on a bunch of time. So, well, everyone, I hope you guys are enjoying it. Um, our viewership's been cut in half. Uh, so I'm assuming we're coming up on the end of people's attention spans. Been an hour and 10 minutes. I appreciate you guys being here. If you guys enjoyed the live stream, if you guys have questions, if you guys want to tell me why I suck or why this was awesome or share with others, um, please, uh, share that out. Help me out. I enjoy being here, getting to hang out with you guys. Uh, kind of— who's Mark? So yes, I'll be ending here in a moment, but I'm gonna read Mark's comment. Um, so you did a video last year regarding triggers back and forth from the CRM. I took that to my AI agent, connected to GoHighLevel, triage the call with automations. Hey man, that's the whole point. I love, I love it, man. I love that. And if you want something else cool, you can go to our GitHub, and we do have some public repos that you might like, such as our local service ad agent. So if you go to there, I'll post it in the, uh, I think this comment goes to YouTube. I don't know where it goes. It says all, so I think it might go to all. Um, but the Google Local Service Ad agent, you could add this. That one's free. Um, and there's also a Meta Ads agent. This one could use an update, but because they've now got the new, the new Facebook app, or developers' private in, in development app So this one could use a bit of an update. Someone contribute to it so I don't have to do it. It's just, I'm happy to let you be a— I'll let you be a maintainer. Just post a PR, make it better. Um, but yeah, thank you everybody for being here. It's been a pleasure. I, I really loved hanging out with you guys tonight. And, uh, sorry, Codex is new to— like, this new Codex is new to me. Um, so I, I'm not yet confirmed that this is better than Fable, but Um, I'll be up for a bit yet. All right, bye.