Navigating the Future: Gas Town and Bespoke AI Agents

0:00 Scott: And I'm just kind of thinking about how that's going to go. Like if all of our companies are out there having to build our own software factories,

0:05 our own agents,

0:06 all this stuff, there's a lot of risk in that.

0:08 I'd love to have the wisdom of the crowd.

0:27 Hello, Brad.

0:28 Brad: Hello, Scott.

0:29 How are you doing?

0:31 Scott: I'm happy to be at another week of How Many CTOs Does It Take podcast.

0:36 Brad: How many CTOs does it take?

0:38 Scott: It takes two,

0:39 you and me. We

0:40 Brad: There

0:40 Scott: can solve

0:40 Brad: you go.

0:40 Scott: any problem together.

0:41 You,

0:42 me, and an AI agent can solve any problem in the world.

0:45 Brad: Exactly. So welcome everybody to another episode of the How Many CTOs Does It Take podcast?

0:50 I'm Brad Hefta-Gaub.

0:51 This is my trusty co-host and

0:53 Scott: Scott Porad.

0:55 Brad: Scott, what are we going to talk about today?

0:58 Scott: We're going to talk about a LinkedIn post that I saw of yours,

1:01 Brad,

1:01 that went viral.

1:02 Brad: That went viral. It did. Actually, I've had a couple that have gone viral in the last week.

1:06 So it's been a strange experience for me. I'm not used to my LinkedIn posts going viral.

1:11 Scott: You're internet famous now.

1:12 Brad: Oh,

1:13 I don't know about that.

1:14 Scott: Let me read to our audience what you had to say.

1:19 And then I have a comment or

1:20 Brad: Yes.

1:21 Scott: two that I'd like to explore.

1:25 Brad: Like the 56 other people that also had many comments on this post.

1:32 Scott: Let's see how many comments are on this post.

1:34 Brad: 57 comments on this post,

1:37 88 likes.

1:40 So.

1:40 Scott: Okay.

1:41 Software companies, according to an author named Brad,

1:44 have to go.

1:46 Build generalized products because software was expensive to write.

1:50 The labor costs only made sense if you could sell to a million customers.

1:55 So you built configurable platforms with massive option services, then hired implementation consultants to bend them to fit each customer's.

2:05 There's actual workflow.

2:06 AI just collapsed that math.

2:10 Satya Nadella, CEO of Microsoft,

2:12 said it bluntly on the BG2 podcast.

2:16 Is that a different podcast you have, Brad?

2:18 Brad: That's not my podcast.

2:20 That's not my podcast.

2:22 Scott: We should get Satya on our podcast.

2:24 Brad: Satya will definitely join the How Many CTOs podcast for sure.

2:27 Hey,

2:28 Satya, I know you're one of our 39 listeners,

2:30 so just hit up our producer and we will definitely have you on.

2:35 Scott: SaaS apps, according to Satya, are essentially crud databases with a bunch of business logic.

2:41 Business logic is all going to these AI agents.

2:44 Mark Cuban has been amplifying the same point,

2:47 noting that 30 million U.S. solopreneurs and SMBs were never well served by enterprise SaaS in the first place.

2:56 The math just didn't work. Now it does.

3:00 Custom workflows for a regional law firm,

3:03 a vertical specific CRM for the dozen companies in your industry that hate Salesforce,

3:08 internal tools for every SMB needs.

3:12 need but nobody could profitably build.

3:14 SaaS got fat because generalization was the only way to recoup engineering costs.

3:20 That's no longer true.

3:22 The next decade of software is going to be a lot more bespoke.

3:27 Okay, here is my reaction to that.

3:30 Brad: Tell me.

3:31 Scott: It's got a few parts.

3:33 Part one,

3:33 agree.

3:36 Brad: There we go.

3:37 End of podcast.

3:39 Scott: Part two,

3:40 sort of.

3:42 Brad: There you

3:42 Scott: We're getting a podcast.

3:43 Brad: go.

3:45 By the way, just to be clear,

3:47 you know, whatever, this is a, you know, I think actually when I wrote this, like I wrote it a couple, I had to like, I had a version of it that I wrote and then LinkedIn was like,

3:56 oh, that's too many words, you can't post that.

3:58 So then I had to like figure out like, how do I cut some of this out?

4:01 But there's a lot of nuance going on there.

4:03 So I will say that, I will say that like,

4:06 I will argue with, there's many interpretations.

4:09 of this as evidenced by the 57 some odd comments that that have been a combination of there's definitely been a combination of haters and lovers of this post right so people have piled on with you're an idiot that's never going to work to this is exactly what I'm doing and I'm making money doing it so

4:27 Scott: So the two examples that you gave are actually are perfect.

4:32 like segue into the question I fundamentally have or the thing that isn't working in my head,

4:38 but maybe you could answer it.

4:40 Custom workflows for a regional law firm is one example.

4:44 Vertical specific CRM for a dozen companies.

4:47 Okay,

4:47 so the vertical specific CRM makes a lot of sense to me.

4:51 Brad: Yeah, like let me give an example what I mean by that, right?

4:53 Like, I don't know anything about being a plumber, but I would imagine that

4:58 A custom-built CRM for plumbers would have a lot of very specific features to plumbers that wouldn't matter to veterinarians.

5:09 Right.

5:10 But could be but to get that in Salesforce,

5:14 like, yes, Salesforce has all the bells and whistles and you could you could definitely configure it and you can hire Salesforce,

5:19 you know, deployment engineer to do it. But like the average plumber is just like, screw that.

5:24 I just, you know, I need to be able to, you know, reach out to the people that I regularly clean the pipes of and and give me a version of that that's specific to my needs.

5:34 That's that's the example I'm.

5:35 I'm used that's kind of my core example there

5:38 Scott: So there's, I think that's a workable example.

5:42 And let me give you a few reasons why I think it's workable. Because you've got a fair amount of customers, a dozen, whether it's just one person or whether it's just a few people,

5:51 they're going to be around to maintain and support that software for a while.

5:57 And in addition.

5:59 The plumber doesn't quite know what they need.

6:02 I mean, they can make feature requests and say this is how I want it to work.

6:05 But there is some value to the 13th plumber of having a platform which is.

6:15 suits them out of the box as opposed to having to sit down from scratch and figure out what do i need in a crm it reminds me of a dealbert by the way where

6:26 Brad: Yeah.

6:27 Scott: uh i don't remember the name of the program manager in deal where the product manager but there was an intern named asok and the product manager was saying to to the intern what kind of software do you want me to build and the intern said well i don't know what kind of software

6:41 I don't know what I want until you tell me what it's capable of.

6:44 It's the same kind of thing in the sense of you could sit down with a plumber and say, I'm going to design a CRM just for you.

6:50 But there are all sorts of things that a plumber might need in a CRM that they might never think of and you get some wisdom of the crowds.

6:57 Brad: Yeah.

6:58 Well,

6:59 and just to clarify,

7:00 like I am.

7:02 not trying to say that the cost of building collapses so much that you could go build a completely bespoke CRM for each plumber, right?

7:14 Scott: Mm

7:14 Brad: Like I agree

7:14 Scott: -hmm.

7:14 Brad: with your point that like plumbers in general will have some needs that there's enough commonality across them that you'd be able to.

7:26 build build the plumber crm and it would be cheaper for you to build it than it and cheaper for you to operate it for for this this niche audience of smes

7:36 Scott: Mm-hmm.

7:36 Brad: then it would be for each of them to get a salesforce license and a salesforce custom customization engineer etc so that's that's like the simple that's kind of like the simplest distillation of the that that point that i would make and and if you actually like

7:52 Like the Cuban was on the,

7:56 you know, the.

7:57 tbpn vlog video podcast and this was like back in i think it was actually in december so it's not like this is brand new and he was basically like saying hey if you're like a young scrappy entrepreneurial you know fresh out of college go go get embedded in small businesses learn their learn their businesses and then bring your bring ai skills to bear to basically

8:24 solve problems for them and there's and that's a healthy business that's not going anywhere is what his point was and and so i'm kind of just riffing off that point a little bit

8:34 Scott: Totally get it.

8:35 Okay,

8:36 so the other example,

8:37 custom workflows for a regional law firm.

8:39 This is the one that puzzles me a little bit.

8:43 So let's say I'm a guy.

8:44 I'm an independent software developer.

8:47 I get hired by

8:49 a reputable firm of you know poor Adam has to go to build some custom workflows and so I go in now I've got it I've actually got to have some real product management skills to go in there and understand what they need then I got to build it for them and you know then it's there

9:11 And maybe it's doing its thing.

9:13 I suspect the industry is moving so fast that it's going to need to get updated quite frequently.

9:20 And, you know, am I going to be around to continue to support this law firm or in another 18 months?

9:29 are they going to come to me for support and say hey we want you to do something or this is stop working all of a sudden and

9:36 Brad: Yeah.

9:36 Scott: where am I where am I going to be am I still going to be doing this thing of making software out for little companies or will that trend passed and I went and got a job at meta or um you know who knows what so

9:48 Brad: Yeah.

9:49 Scott: I just that's the one the the the relatively bespoke software for like

9:54 everybody getting their own software at their own firms for their own things and that firm that they're going to start doing their stuff whatever they use for this custom workflow is going to be some wacky way and all there's going to be all this differentiation

10:05 Brad: Yeah.

10:05 Scott: across the regional law firm community and it's going to be a mess so i just that's the part i'm having a hard time getting my head around what about that one

10:14 Brad: And to be fair, that was definitely a thread in some of the comments of the haters.

10:19 on the on the thing saying oh this will be just total chaos and like why you know you know if you think sass is bad then imagine like all this bespoke stuff that'll just be chaos and i think i think that they are leaning into that my now my only defense of my comment would be simply my defense of my post would be well

10:36 I'm making a I said the next 10 years,

10:40 right?

10:40 So I just think like, I

10:42 Scott: Yeah.

10:42 Brad: mean, there's a big picture here. And the point I think the point that I'm making is that or the point I feel confident in is that this stuff has changed so it's gotten so much better in the last two years.

10:58 that you and i've been paying super close attention to it that there's no doubt in my mind that the the build cost is going to continue to collapse and therefore the idea that um you can build these custom things and it's a lot you still got to have good strong engineering best practices but it just is much cheaper to build things

11:24 Scott: Yeah, I'm having a hard time even conceptualizing what it's going to be like in five years.

11:31 Brad: Yeah, agreed.

11:32 Scott: I can't imagine it really.

11:34 I mean, I can kind of imagine some things,

11:36 but.

11:36 Brad: I mean, we could talk more about this question of like the post that went super viral, but there were two other posts that I posted this week that were in relation to this topic that also went viral themselves.

11:47 Not quite as much as that one,

11:49 but still pretty big numbers from my perspective. And those were more engineering focused posts.

11:56 And one of them was very specifically kind of.

11:59 kind of pushing back on the i run my 20 agents overnight and they they write they make all this magical stuff for me and i'm you know and i you know it's production software and yay and i'm pushing back on that saying like hey in my fractional cto work that's not what i'm seeing i am seeing teams that have fully adopted ai um

12:21 coding and they've got it built into their kind of AI oriented SDLC or their ADLC as you and I have called it a couple pods ago and they are cranking out lots of business value you know thousands of lines of code per developer of business value a day and of course you know I always joke about clock we all know clocks is a total

12:45 Bad metric,

12:46 right?

12:50 They are really, really, I mean, the costs for building is plummeted for them and that's causing them to build really fast. And when they're doing high quality engineering techniques around the building, they're getting amazing results.

13:04 Scott: Yeah,

13:05 that's right.

13:06 And that is where I'm going with TaskRabbit right now is what you talked about in this post,

13:10 which is getting the environment set up.

13:14 to really do the real deal and that's it that's going to be the new world and that'll actually that'll be i think what separate the good teams from the great teams

13:21 Brad: What I think is interesting about that is that set your environment upright.

13:25 Like that's got nothing to do with AI.

13:27 That's actually just got to do with like good software engineering techniques.

13:32 And I do think that

13:35 I mean, I'm not seeing people really I'm not seeing a lot of people be successful with this sort of dark factory or to say like, you know, like there's not humans watching the robots.

13:45 I'm not seeing anybody do that yet myself in my experience.

13:48 There's a lot of people claiming that they're doing it. Here's my guess,

13:52 pretty strong expectation.

13:55 The ones that are actually succeeding at that have built the harness around their agents in such a way that it's the same safeguards that would have guarded against humans writing code that fast.

14:07 Scott: Yes, that's right.

14:09 Imagine if you had a new employee show up on the first day and submit a 70,000 line PR,

14:13 what would you do?

14:14 Brad: Yeah.

14:16 Scott: Can we talk about the dark factory thing for a

14:18 Brad: second? Yeah, totally.

14:19 Let's do it. Cast town,

14:20 baby.

14:21 Scott: Gastown, did I talk to you about Gastown?

14:23 Brad: I mean, it's all over the place.

14:25 Tell me about your, you just learned more about Gastown, right?

14:29 Scott: I did a two-day intensive hands-on workshop with the founder and co-founders of Gastown.

14:35 Brad: Nice.

14:36 Scott: Do you think Gastown is ready for primetime?

14:39 Brad: Well,

14:39 I haven't tried to use it myself,

14:41 but I was reading articles about it.

14:43 Actually, this morning I was reading some of the more detailed articles about it, and I'm intrigued.

14:50 Scott: Well, I spent two days with it and it was very intriguing.

14:57 I have Plodits and what's the opposite of Plodits? I have compliments and complaints.

15:06 Brad: Okay.

15:06 Scott: It's cool.

15:07 I mean, it opened my eye. It brought my mind to a new level.

15:11 what this could be right and so like it was mind expanding for me and just for context this was a two-day workshop organized by the AI tinkers meetup which is a meetup group organized by my friend Joe Heitzberg and he's got chapters all around the globe and they organized this thing with

15:34 gas town the the founder the two co-founders of gas town there's like a ceo and a cto and that's why i keep on saying founder and co-founder but they're just two guys together so that was cool and i gave it a simple spec and i set about 15 agents about i kicked it off and it worked away for about 20 hours and it came back with seemingly a functioning product for what i wanted which was basically just a fantasy football

16:01 mobile app game um it didn't work perfectly because i didn't there were some things i didn't understand about how to use gas town or really even how to use a software factory i just gave it the spec and i gave it the command make the mobile app and then it went to work for 20 hours and i gotta you know that's not the way it turns out that's not the way you should do it i should have done it iteratively one step at a time just like we've done just like plan mode right

16:28 But I had on my machine running a planner agent,

16:32 a PM agent,

16:33 a designer agent,

16:34 five builders,

16:36 two reviewers,

16:37 and a deployer. It brought my machine to a screeching halt at times.

16:43 So we've got to figure out that problem.

16:44 That's a new problem we're going to have.

16:45 Like I couldn't run five gas towns at once on a MacBook Pro.

16:49 maybe if i had like 64 gigs of ram right now or something i learned some i learned some cool ways how to nudge it um like it would get stuck and so like i told it stuff like okay i want every i started to build up this i know it was a cron job under the covers but i started to build up this thing like every five minutes i want you to do x and at first i was just like i just want you to tell me your status and then it was like tell me if you got any questions you need anything to block you either or you need

17:16 you need me to help you move you forward and then i noticed sometimes some of the agents would just stop and so um i said hey i want you to just nudge the agents every five minutes and be like get back to work kind of thing you know and so it was all really cool it was all really neat to see how this is all going to unfold um i thought two or three negative things i mean the gas sounds super raw so i got to give it a break i mean they released it on march 1st or something

17:41 or something like that so i

17:42 Brad: Yep.

17:43 Scott: think version 1.0 came out last week was beta like three weeks i mean they're using gas town to build gas town i'll say just as a personal preference they're using this city metaphor with the mayor and rigs and all these different words and i just think that's that's a turnoff for me because i'm having to do this translation what's a rig

18:04 oh rigs is the app that you're building okay then why don't you call it your app right what's the what's the mayor well the mayor is kind of like the supervisor of the whole thing it's like okay like why don't you just call it the supervisor you

18:14 Brad: Right,

18:14 Scott: know so

18:15 Brad: right.

18:15 Scott: um so i've never been a big fan of like the sort of cutesy

18:21 nomenclature for a thing and i think that'll be an i actually think that'll be an impediment to gaston because i think across the industry you know big companies won't want that kind of nomenclature in their environments and then the other piece is like how long before all this stuff is just in cloud code cloud already has this thing called agent teams

18:43 Brad: Yep

18:43 Scott: Actually, I don't know if you've used it. I haven't used it yet.

18:46 And there are many other competing factory frameworks out there.

18:50 But it showed the skeleton of how it's going to work. There are going to be agents,

18:54 there's going to be a harness for them with an orchestration between them, and there's going to be a communication mechanism between them.

19:02 I think those are going to be the puzzle pieces.

19:04 And then, you know, you'll be able to customize all your agents and customize all a million different things.

19:11 It's going to take time to set up your factory. I think that's the last thing I learned.

19:15 There are factories out of the box,

19:17 but then you're going to have to go about building your own.

19:23 factory customizing that factory your front-end agent and my front-end agent are going to be well this factory I had just had a builder agent and someone might make a front-end agent and a back-end agent that are more specialized

19:35 Brad: Well,

19:36 and I can imagine a world where you had two reviewers,

19:40 but maybe you're going to have a security expert reviewer and then you're going to have another red team.

19:48 reviewer and you're going to have a compliance

19:50 Scott: Yes.

19:51 Brad: officer reviewer and you're going to have a you know design reviewer or whatever

19:55 Scott: I thought it was really cool.

19:58 It really helped me see.

19:59 We've talked about learning here.

20:02 it you know the ways people learn I'm really a very much a kinetic visual learner and

20:07 Brad: Yeah.

20:07 Scott: so just seeing it seeing you know the guys who did it put it on the screen and say this is how we did it getting to go through a lab and just you know build something on my own was really helpful it really helped me just like oh okay I see where the future is going with this thing like I saw it before like intellectually but now I get it like in my kishka's like oh I get this you know like

20:27 like here's a question for you i know your personal perspective on this because we've talked about it so much like that you

20:35 uh believe that this is the trend is happening and you're gonna you gotta get on the bus or or not um but here's my question for you

20:44 The three posts for me that went viral this week all had this notion of talking about really using AI in production code.

20:53 I did express some skepticism of the are the agents really ready yet, right, to be just let free.

21:03 You know, and I've obviously talked about like with this small business thing, I talked about how, you know, cost of building is collapsing. And so, you know, that just opens up more opportunities.

21:12 Okay.

21:13 So that was kind of the themes of my posts this week.

21:16 I was genuinely surprised by the number of

21:25 very experienced senior operators in my LinkedIn network and that got that caught this on LinkedIn who were still completely poo-pooing AI as being able to write production code I was shocked by that and I'm kind of curious I don't know what do you

21:50 What are you seeing?

21:51 What's your take on this?

21:53 How is that still a thing?

21:56 Well, I've got a few answers to that.

21:58 The first is I'll start with something potentially a little humorous.

22:02 There's a famous quote by I think the author H.L. Mencken.

22:06 Right.

22:08 But it could have been by Upton Sinclair,

22:09 I believe.

22:10 Actually, I believe it was by Upton Sinclair,

22:12 now that I think about it.

22:13 And for those of you on this podcast who do not know who Upton Sinclair is,

22:17 he wrote a book called The Jungle,

22:18 which actually,

22:20 going back and reading that book, might be highly relevant to the times in which we are living.

22:24 Yeah, good call.

22:26 The Industrial Revolution, it was about the Industrial Revolution and the impact it had on workers during the Industrial Revolution. And we're going through an AI revolution.

22:33 And what he said, he had a quote that's attributed to him, which is, it's hard to get a person to understand something when their living depends upon not understanding it.

22:49 Yeah

22:50 That's my first thought.

22:51 My second thought is I run into people all the time who are naysayers.

22:57 they don't believe it's going to happen and they don't think production ready code they don't think it's ready and almost all of them I think almost all of them have not actually used it they've not really

23:10 They genuinely sat down with an open mind and tried to engage with it and said, can I make this work?

23:17 And when they did engage with it, did not follow the best practices for doing

23:21 Yeah,

23:21 so.

23:21 they were sort of like pre-biased to want to see it fail.

23:26 So they did stuff with it that they themselves wouldn't do themselves kind of thing.

23:31 Or, you know, they just did a little bit and they're like,

23:34 it isn't really working.

23:35 And I know that the people who I think most highly of and who I think are most talented with it, and frankly, Brad,

23:42 I include you in that group.

23:43 I know the people who are the most who are getting the most out of it have spent hundreds of hours trying to figure it out, working with it, learning how to work with it. Those people.

23:56 are submitting 70 000 line prs that are flawless all generated by ai right

24:01 Yeah.

24:01 which is insane that someone would do that but they've got the quality they've got the process and procedures in place they've got the quality gates in place

24:10 I'll tell you a fun story. For the first time, I've actually,

24:13 it wasn't 70,000 lines, but I was working on a tiny little, a tiny project on a side project I was working on and I had a

24:21 Basically had a agent write a B2C style like monthly subscription stripe integration,

24:30 you know.

24:31 thing for me so had

24:32 Mm

24:32 a very clean spec that i wrote and then had the agent go off and build it and um you know had that had the framework that i'm always preaching of having like you know strong quality gates and stuff and one of the quality gates i'm very used to is like i always give i always always always always always always let copilot run a code review on anything and the first time ever i had a pr

24:56 that was 24 000 lines of code and i found out that copilot has a limit of 20 000 lines of code to do to do copilot code reviews so

25:07 That's interesting.

25:08 i had to like go oh okay hold on a second well half that half that pr was actually unit tests so it was a pretty easy fix like okay like split split the unit tests off as a separate pr and

25:19 And I don't think we want huge PRs. I think we still want small PRs.

25:23 PRs.

25:23 So one of the things I'm very much seeing for real in the teams I'm working with is they're pretty much all pretty much the smallest PR I'm seeing at this point is like a small PR is now a thousand lines of code.

25:39 And most of them are more like averaging closer to three thousand lines of code.

25:50 they wouldn't want to split those up into 3,000 line PRs.

25:55 because the work product of what they're able to get done has grown to the point where you can essentially implement a whole epic.

26:05 Now, you still want to be iterative, and you still have broken it down, and you've still challenged the agent to break it down, and you've done a whole bunch of local testing and validating the plan well before it got built.

26:19 splitting it into three smaller prs or four smaller prs actually makes the job of the reviewers harder because they they have to review essentially the piece and like it's only meaningful once all the pieces are put together like it's it's it's not only meaningful but it's it's much more meaningful when all the pieces are put together so that you can

26:43 see the functionality manifest end to end,

26:48 right?

26:48 Mm-hmm.

26:49 And so this is something that I do think

26:52 I do think is challenging one of our standard practices that we would always say to people is like,

26:58 Mm

26:58 hey,

26:58 -hmm.

26:59 break it down,

27:00 get your back end working correctly,

27:02 have like super solid unit tests against your back end. If your front end is coming before your back end, then great.

27:07 Implement like mocks or something like that, that your back end, you know, if when the back end arrives,

27:12 it fully manifests the contract that you've agreed to,

27:15 then your front end will work perfectly. And that's how we would like.

27:17 would like break things down into smaller pieces and still get them over the finish line well that model which which was a which was a perfectly reasonable model that we've used for you know decades

27:29 That actually creates more work for us if it's possible to stand up the back end and the front end at the same time across the same platform and demonstrate it.

27:41 Right.

27:42 And this was actually something that remember when we when Justin Kaufman came on and he was telling us all about the project he'd worked on, he basically made the case for like,

27:50 yeah,

27:51 I'm like all about mono repos now. Like, I'm just going to like.

27:54 It's just so much easier to build features when the entirety of the stack is available for me to work in.

28:03 Mm-hmm.

28:04 Yes,

28:04 agree.

28:05 In fact, we're having this conversation about whether we should sort of move to a model right now.

28:10 Coming back to the naysayer thing,

28:11 the only thing I wanted to add was, you know, I was just talking to someone.

28:15 earlier stay at my organization who is who is on the leading edge and he has been also frustrated by people and he gave this analogy he said I feel like I'm talking to people who are saying I'm not sure if this airplane thing is a good idea and he's like he's like what I just flew here on an airplane and it was effing awesome right and to me it's about experience I've just the people I see the people who I I mean maybe this is a

28:40 a truism,

28:41 but the people who are really making it work had to spend time to get there.

28:47 Yeah,

28:47 And

28:47 I

28:47 now, is it that there were other people who spent time to get there and they couldn't make it work and so these are just some lucky ones?

28:54 I don't know. But these people

28:55 don't buy who that.

28:56 I see, the people who I see making it work also before AI were among the most talented engineers I knew.

29:03 Yeah.

29:03 So I see that those going hand in hand. And the last thing I'll say to the 39 people listening to our audience,

29:10 I haven't asked every one of these engineers,

29:12 but several of them are telling me.

29:13 or tell me i'm having more fun engineering than i've had in a long time right i mean to be a professional commercial software engineer it's a little bit of a grind right you work in a big company there's product managers there's meetings you

29:25 Yeah.

29:26 know you're doing stuff you got to spend time writing the test it's all it's all a lot of work right and i'm thinking one guy in particular who you know said to me like it's just i'm just get it's just fun to get so much stuff done i just

29:38 just feel like i have been released from

29:41 Yeah.

29:42 so much and the visual i have in my head is um you know play-doh you ever play with play-doh yeah

29:48 You mean the Play-Doh like what we played with as kids?

29:52 and

29:52 Yeah.

29:52 Okay.

29:52 you know how easy it is to make a snake with play-doh you just roll your hands together like this

29:59 I feel like it's like making the snake.

30:01 It's like, that was easy.

30:03 Yeah.

30:03 This is fun.

30:04 I just made something.

30:06 Yeah.

30:06 I don't know. But so that's what I've seen. And, you know, what are the reasons people that then there's all sorts of reasons people have for why it won't work.

30:16 That's fair.

30:16 That is fair.

30:17 Okay.

30:18 And I definitely agree with what you're saying.

30:20 And I think that that's what I'm seeing in the people that I work with, the ones that are basically open minded, curious and were already good engineers to start with were already creative.

30:29 Like it's just like, bam,

30:30 this is this amazing tool and I already knew how to write code and this thing is just helping,

30:35 you know, I can just get it to write code correctly the way that I would have written it and it's just making me so much faster. Right.

30:42 And,

30:43 you know, another one on this.

30:45 i would say is like one of my senior engineers i work with regularly he was like it's doing a better job of code reviews than i ever did right

30:52 Mm-hmm.

30:53 it's like yeah like it it is a it is finding the obscure things that were really hard to find you know and so you know again like i agree i clearly am a believer and a convert that this tool is really quite amazing

31:10 Now, in the last week, there have been some pretty colossal, some pretty colossal massive mistakes.

31:19 The startup that accidentally deleted their production database because they let Cursor get,

31:26 you know, had an MCP server to essentially the database and it like it literally just saw some it had some bad mismatch keys and so it decided it was going to fix those.

31:36 It's those mismatched keys,

31:38 and it basically deleted the database and the backup.

31:41 Look,

31:42 the AI agent isn't eager to please entry-level engineer.

31:47 I'm sure at some point in my early in my career, I deleted something I shouldn't have.

31:51 And over a generation,

31:54 my generation,

31:55 we built up all sorts of guardrails to prevent against those things.

31:59 Right,

32:00 and so I guess I'm throwing that out as an example of...

32:04 That's great fodder for all the naysayers.

32:08 Totally.

32:09 And to which my response would be, who gave your local development environment write access to your production database?

32:17 I mean, this is just basic development practice.

32:21 You wouldn't have given write access to your eager to please intern.

32:24 You wouldn't let your intern make a change to your database without having it looked at.

32:30 Right.

32:31 I mean, you know, one of the things I've learned, you've probably learned it too,

32:34 is when anytime somebody is doing a major change to our environment,

32:38 the first thing I ask is, what is your rollback plan?

32:42 Because things go wrong. So if you don't have a rollback plan before you go into doing it, my point is like, no, you need to get that together first and then go do it.

32:52 And that's Murphy's law in action, because when you get your rollback plan together,

32:55 then you don't need it.

32:56 Right. Exactly.

32:57 Yeah.

32:58 And

32:58 Premium.

32:58 if you have a, and whenever I see a plan of some sort where it says like, this is a point of no return once we go past here or whatever,

33:05 I'm like,

33:05 no, you guys need to go back and rework this,

33:07 figure it out. There's some other way you can do this, I'm sure.

33:10 Right.

33:11 You can make it safe because we've learned from all those production databases that we fat fingered on our own.

33:18 One of my non-technical buddies who's using AI pretty aggressively, and he's very technical actually,

33:25 but he's just not a classically trained software engineer.

33:28 He and his co-founder is a classically trained software engineer. They've set up their,

33:34 and I don't know if they're using Gastown or not,

33:36 but it sounds almost like they're using a Gastown type thing.

33:38 And they have two different code review agents and one

33:41 Mm

33:41 of

33:41 -hmm,

33:41 the code review agents they've given the persona of the the eager rising star with a bias for action and just get it done.

33:50 And then the other code review agent,

33:52 they've given it as the persona of the grizzled veteran that has seen so many production failures like,

33:59 you know, you wouldn't believe this war stories I have and they literally.

34:04 really set up they gave their coding agent their code review agents these personas intentionally to get like you know essentially antagonistic reviews

34:14 Okay, there's a few things I want to say.

34:15 I want to come back to the previous thing about rollback plans and just I want to give a nod to Mike Hansen if you're out there listening once Mike Hansen and I effed something up really badly in production and we got it all fixed and ready we had a script already to fix it and he did something which I thought was so wise that I've taken with me he was like we're working as fast as we can to fix this problem we're ready to push the button again to like fix it

34:40 And he said, you know, why don't we just wait for an hour?

34:44 Why don't we go get some food?

34:46 Just let's just see if anything comes to our mind that maybe we haven't thought of before we push this button again to fix the problem that we already made and then make another problem.

34:55 Right.

34:55 And I just thought that was so wise.

34:57 Yeah.

34:58 He was like, why don't we just wait?

35:00 I know we're in a hurry,

35:01 but why don't we just wait one hour and maybe we'll just think of something.

35:04 Yeah, there's a quote that totally resonates with that that I've heard from my mountain world is don't just do something.

35:12 stand there.

35:14 Yes, that's right.

35:16 That's right.

35:17 Okay,

35:18 so you're talking about the two agents.

35:20 When I was with the Gastown crew last week,

35:23 I can't remember the exact scenario they had.

35:27 I think it was similar to this code review.

35:29 They had three personas,

35:31 like these code reviewers.

35:33 They ran them all.

35:36 three on the three major models so gemini uh whatever open ai's got whatever whatever model gemini there's some gemini models some open ai models some anthropic models and then they had they had nine reports

35:49 Then they had three other agents that synthesized all these together. So then they had three reports and then I can't remember how they whittled it all down, but they had built this whole layered things of different perspectives, synthesizing different perspectives to come up with.

36:07 Yeah,

36:07 this is OK.

36:09 Yeah.

36:09 If all these different perspectives from these different models, because remember the models do stuff differently,

36:14 they've got biases built into them, you know, and so it's the same idea and that's probably where all this stuff is going to go,

36:19 which leads me to a question I've been sitting on for about 20 minutes now.

36:22 So okay,

36:23 code reviewer agent,

36:24 grizzled veteran,

36:25 eager,

36:26 eager intern.

36:29 When I was doing this training last week and we were getting it all set up and you're setting up all your agents,

36:36 you know, I was just copying.

36:38 agents that they had made for this workshop that i could just use to like use the thing and they talked about like look you're going to kind of create these for your own company for your rules for how you want to do your business in fact i was talking to another company today and they were i was talking to the cto was telling me how they created like you know front-end agents and develop different types of developer agents all this different stuff they'll specialize for how they like to do it at their company

37:03 And the thought that I've been kind of running through my head is,

37:08 do you think that's how it's going to work or do you think we're going to?

37:12 like is there going to be maybe maybe be more of a marketplace for agents where like i'm just either i'm just buying the grizzled veteran code review agent and it's got it's in github and it's got 50 000 stars or whatever so everyone thinks it's really good and we just like we've all as an industry agreed like this is the good one and maybe it's an open source agent so people can submit

37:33 because an agent's just an MD file.

37:34 People can

37:35 Yeah.

37:35 submit improvements to it.

37:36 I'm just kind of thinking about how that's going to go. Like if all of our companies are out there having to build our own software factories,

37:42 our own agents,

37:43 all this stuff, there's a lot of risk in that.

37:45 I'd love to have the wisdom of the crowd in that.

37:48 Yeah,

37:48 it's a great question.

37:49 I have a couple.

37:50 feel i think is probably going to be a little bit of both that's my answer you know i'm going to stick with that one it's going to be a little bit of both you know we've already been living this with open source right like there are

38:02 you know, open source frameworks like that are clearly benefiting from massive amount of support.

38:08 And so to your point, it's like the wisdom of the crowds.

38:10 And then there are still opinions and one

38:13 Mm

38:14 opinionated

38:14 -hmm.

38:14 version of a framework versus another opinionated version of a framework like and, you know, so I think there will be I think there will be room for diverse opinions in this.

38:30 And that kind of goes back to the point that I was making in my post that went viral is that the cost of building actually has gone down so much that you're allowed to be opinionated because it used to be it used to be cost prohibitive to be opinionated and now it's not cost prohibitive.

38:48 That's exactly right.

38:50 That's why you end up with products like Jira.

38:52 Yeah, exactly.

38:54 The thing that I feel pretty certain about that for the and I'm going to say near term, I don't know how long this near term is, but for the near term,

39:03 I feel like I've been saying this for a while. It's the K-shaped economy.

39:07 The people that were already.

39:10 good at building software and knowing how to build software in a high quality way those people are really and and then they they decide to adopt AI strategies they're the ones that are going to just get these massive outsized gains and then people in the lower half of the K are either going to be a combination of to your point people that aren't open-minded aren't curious maybe they were very talented

39:36 But they've just for whatever reason, reluctant to actually adopt and be curious about the tools.

39:43 That's part of it's going to be that.

39:45 And then the other part, I think, will be filled with people that don't have the fundamental engineering expertise. And so they're very much dependent on getting lucky.

39:59 And

39:59 Brad: Mm

40:00 Scott: not

40:00 Brad: -hmm.

40:00 Scott: being able to check the work of the agents. And at least now in the short term,

40:05 I think that's a very real, like there's not yet a solution to that.

40:14 And, you know, anthropic would say, oh, that's only six months out and we're going to we're going to you know, we're going to solve those problems.

40:21 I don't know. Six months.

40:23 I mean, it was certainly was a year ago that we were going to lose all software engineering jobs in six months. Well,

40:28 you know, past due on that one.

40:30 Yeah, well,

40:32 that actually leads to another thought,

40:34 which is I just sort of see all the software factory stuff moving into something like cloud code or whatever.

40:42 But what is you know, there's codecs over it up in the eye and I don't know who has them But don't you just think all this is just going to move into the platforms?

40:50 Sure.

40:51 But again,

40:51 my thesis,

40:53 that's great.

40:53 Yeah,

40:54 sure.

40:54 I think if you're trying to make a business out of building a framework,

40:58 I think that's probably not going to be that useful to you.

41:01 But I think that the frameworks are getting better and better.

41:08 Mm

41:08 And

41:08 -hmm.

41:08 using the frameworks well is just a huge opportunity.

41:14 yeah but your point that you were just making about being able to verify that the agent or the model does the right thing is real it's real here it's real with everybody who's using models you know like i i tell everybody that i get a chance to tell them like hey i looked this up on ai but you can't believe anything you get from ai so like take it with a grain of salt you

41:36 Yeah.

41:36 know when you want when you want another because you know you do stuff like

41:39 Okay, I'm just going to go right now,

41:41 ask Gemini, where is it? Where's my Gemini window?

41:43 I got Gemini open right now.

41:45 What day is it?

41:47 Is he going to get it right?

41:49 I don't know. Let's find out.

41:51 It actually got it right.

41:52 It is indeed the day that it said,

41:54 which I will not say because I don't know when we're going to release this podcast.

41:57 In case people wanted to know,

41:59 Gemini also told me it's also International Jazz Day,

42:02 National Bubble Tea Day,

42:04 National Adopt a Shelter Pet Day and National Honesty Day,

42:09 which is really interesting because that means all the rest of the days we can be dishonest.

42:14 Yeah,

42:14 I just asked ChatGPT.

42:16 what day is it it just came back with today is the date it

42:20 Yeah,

42:21 didn't give

42:21 I

42:21 me

42:21 guess they got that figured out.

42:22 it didn't give me it didn't give me a fancy details about what what happens on this day today

42:29 Well,

42:30 you know what today is for?

42:32 is tell me what is today it's

42:34 It's a great day.

42:35 a great day to smash that subscribe button

42:39 Great review.

42:39 Rate and review.

42:41 Awesome,

42:42 Scott.

42:42 It's great talking to you again.

42:43 It really was. Good talking to you,

42:44 Brent.

42:44 All right. Have a good one. Bye.

42:47 This episode of the How Many CTOs Does It Take podcast was produced and edited by Saffron Teftekau.

42:52 Our theme song was AI Generated by Brad.

42:56 Please follow us on all the social media platforms at HowManyCTOsPod.

43:01 You can find us on YouTube,

43:03 LinkedIn,

43:04 Instagram,

43:04 Facebook,

43:05 Threads, X,

43:06 TikTok,

43:06 and Patreon.

43:07 Or at our website at HowManyCTOsPod.com. And remember to rate, review,

43:13 and subscribe.

43:13 Subscribe on the podcast platform of your choice.

43:16 Thanks for listening.