AI Agents Need Guardrails. Your Software Rebuild Needs a Plan.

[00:00] Dave Pengelley: Now is it's that pub chat style of podcast, right? So we are talking, connecting as mates, talking about stuff going on. Um, and people will either like our our vibe, the cut of our jib, or they don't. But I think the random banter bit that we're doing right now is just part of that, you know, people were learning who we are and whether they actually care about what we think about any of the things that we talk about. So [00:26] Richard Webbe: It's the it's the virtual water cooler. We lost that conversation. When COVID came along, and then when companies became global and had matrix management, again, we didn't have the water cooler conversation because quite often your boss was in Europe or in the US or in another part of Australia. And so when we get this sort of conversation going, I think it's so important. [00:46] Dave Pengelley: But but [00:47] Richard Webbe: Because [00:47] Dave Pengelley: from a podcast point of view, like like if you're listening to a podcast, you'll stay or go based on whether you connect with the people talking and whether you like their personalities or not. And if you you sort of like I I don't mind listening to two blokes talk and text sometimes with um Stephen Fenneck and Trevor Long [01:01] Richard Webbe: Yep. [01:01] Dave Pengelley: because you know they've they've obviously been mates for decades and they get along well and it's just fun banter as well as you hear about the latest smart televisions or whatever they're talking about on any given week. Top gear was never really about the cars at the end of the day. He washing it for those three buffets. [01:17] Richard Webbe: That is so [01:18] Dave Pengelley: Um [01:18] Richard Webbe: true. [01:18] Dave Pengelley: not to put us in the same caliber and category as the uh buffets from uh the grand tour, but um [01:28] Dave Pengelley: exactly so. That's that's that's why we we do the random bantery stuff at uh at the beginning, something different. I'm I'm I'm I'm gonna do some dances. I'm at my standing desk, so we're gonna try something different. [01:40] Matt: Oh [01:40] Dave Pengelley: That's [01:41] Matt: wow. Interesting. [01:41] Dave Pengelley: yeah. Yeah. You you do you you got a standing desk? Do you stand often, Matt? [01:46] Matt: Uh I'm on and off all day. Like my [01:48] Dave Pengelley: Yeah. [01:49] Matt: you know, so this is now standing and and this is now back on my stool. [01:54] Dave Pengelley: Oh you [01:55] Matt: So [01:55] Dave Pengelley: got a stool. [01:55] Matt: yeah, [01:56] Dave Pengelley: Okay. [01:56] Matt: I've got a [01:56] Dave Pengelley: So [01:56] Matt: I've got a [01:57] Dave Pengelley: keep [01:57] Matt: a [01:57] Dave Pengelley: it in [01:57] Matt: curved [01:57] Dave Pengelley: high [01:58] Matt: bottom [01:58] Dave Pengelley: mode. [01:58] Matt: a curved bottom stool. So [02:00] Dave Pengelley: Oh, [02:01] Matt: it's got like a you know, it doesn't balance. [02:04] Dave Pengelley: like the fidgety stools for the the kids at schools that [02:07] Matt: Yeah, [02:07] Dave Pengelley: keep moving. [02:07] Matt: that's me. That's me, I'm a kid at school. [02:11] Richard Webbe: What's really important about, and as we get into uh another one of our sidetracks, I remember when I was first selling and I was first reaching out to build relationships with people on the phone, and a guy came up to me and said, now Richard, and he gave me a really because we didn't have mobile phones here, gave me a really long cord for my phone. Said, now every time you talk to someone, I want you to be standing up and with your phone. I said, why? He said, because you're an animated person. And when you're settling down, you're sitting, not settling, when you're sitting down, you stifle your personality. And then I used to do a lot of radio with uh Max Walker and a few other characters. And uh actually we went on on um on those two comedians that we've run love's show, and um uh Max turned off the mic for a second and said to me, Now don't stop. smiling. He said, you sound a lot better when you smile. He said, you can hear it coming in. So there's so much about what we do on our physical setup and our actual, you know, actual setup. So I'm I'm I'm a standing desk person. I got one because I had a saw pack. And then I don't have a saw back anymore, but I still use it because I love it. [03:19] Dave Pengelley: Sonny, we like you true, it's true, like like um being up on a mobile and your your brain works differently. We uh the officer used to work at there was an account exec and he'd have His little Bluetooth little headset on, and the sales the sales team desks were on the other side of the building from pre-sales where I was. But they this guy would just walk around doing laps and talk really loud on the phone, and we would all be like, oh that guy, go back. [03:46] Richard Webbe: Yeah, [03:46] Dave Pengelley: Yeah, it was. [03:47] Richard Webbe: yeah. You've got to be careful who you're doing it to, like when you're enjoying your music. [03:52] Dave Pengelley: Yeah, I mean, I I love my AirPods for music, but with the people that jump on the train and they just have their speakerphone or their little Bluetooth speaker cranking in a public area, I'm like, oh so inconsiderate, especially 'cause it's always rubbish music in my opinion. [04:06] Matt: Have you guys come across uh anyone doing like a voice transcription to their agents on the train or anything like that yet? You know how you have the people on the phone to like their partner or whatever and they're just being loud and annoying. [04:17] Dave Pengelley: Yeah. [04:17] Matt: Um I I've not come across why I don't really frequent public transport right now, but um, I think that'll be funny when you start hearing people say, like, no, change this component to this thing, and I want to use this. Just yelling at their [04:30] Dave Pengelley: I [04:30] Matt: jippity on the [04:31] Dave Pengelley: I [04:31] Matt: train. [04:31] Dave Pengelley: I hate taking phone calls on the on the train. Like I generally almost won't talk on the phone. Like Richard, you call me once or twice and we've had to have a chat on the phone because I'm like, ah, if I don't talk to him now, I'm never gonna talk to him. Um, but I hate it. I hate it for that reason because you've got to talk a little bit louder on the phone than if you're just next to someone and it just feels intrusive and invasive. [04:52] Dave Pengelley: Yeah, [04:52] Matt: I think [04:52] Dave Pengelley: not [04:52] Matt: anyone [04:52] Dave Pengelley: a [04:53] Matt: who's worked in like team environments where you go to your team room on your breaks and you know, you just want to go in there just to get away from everything. Eat your lunch and just I don't know, look at something on your phone. And then some dickhead walks in on their phone on their phone, just sitting in the team room, just like talking out loud. Can you not? [05:09] Dave Pengelley: Well, I mean, this is this is why I think I've told this story. Like with my um my old team, I do instantiate quiet time. So from one o'clock to like four in the afternoon or whatever, I would get little ropes and cordon off and I put little signs up that said quiet time, don't disturb my team. So they could actually just focus on getting their work done. And if you wanted to talk to them, like arrange to walk away from the desks to have a chat with them. I wasn't stopping my team from interacting with other people. But I'm just saying if you want to have a chat with them, don't disturb everyone in the open plan office. Go take them to the kitchen, to the water cooler for a chat. [05:41] Matt: Yeah, [05:42] Dave Pengelley: So [05:42] Matt: exactly. [05:44] Dave Pengelley: uh yes. Uh Joe is with us this morning. Welcome, Joe. Good to see you in the comments saying hello. Um, we will get on AI things. Now we'll we'll we'll play the normal trail of the bumper and then we'll get in because I have a special bumper that I want to show. Uh you've seen it. Matt, but I don't think Richard has. But anyway, let's kick off the AI operators weekly live show. [06:07] Dave Pengelley: Anyway. [06:18] Dave Pengelley: Um, it's no secret that I'm a fan of Hermes Agent. I use Hermes Agent all the time as my primary go-to. Uh a little bit of Codex app or Chat GPT app, or whatever they call it now, but a lot of my stuff I'm doing. Via Hermes and they put out a challenge on the weekend. They were giving me 48 hours of the new Flux 3 video algorithm via Black Forest Labs to produce a video for their competition. And I was like, you know what? It's free flux. I'm just gonna crank it. Uh, and so because Hermes, Hermes, I have a whole profile set up for this podcast. And so, like, when I finish this podcast today, I download all the recordings and my AI has set up a whole pipeline of scripts that I run. Where it transcribes it, listens to what we say, and then gets the AI to sort of break out the storylines and build the comics that we do and then produce the comics. And so I've got all these AI things pop all by Hermes. So my AI operators, Hermes, knows a lot about our show because it's read all the transcripts and it's built all this memory and knowledge. And so I said to it, hey, here's the ex post announcing this competition. I'd like to make a new bumper for the show. And I expected it to come. Back and asked me a few questions and it went and just did a first draft straight away. And I was like, okay, didn't quite nail the brief, but because there was no brief. But it had Rubot, our kangaroo mascot, doing a few random things. And I was like, that's that's not bad. Let's let's run in that direction. Uh and so I gave it a bit of feedback back and forth, and we landed [07:52] Richard Webbe: Here we go. [07:53] Dave Pengelley: with this guy, uh, which is there available if you follow AI operators pod on X. You can Can see this. [08:00] Matt: Well, look, you [08:01] Dave Pengelley: Um, [08:01] Matt: got a mad comment from some really like interesting person. [08:04] Dave Pengelley: pretty crazy. That's free from Slager. Nice. Uh, [08:10] Richard Webbe: Oh, [08:10] Dave Pengelley: let's watch. [08:11] Richard Webbe: bad that guy. [08:13] Dave Pengelley: All right, let's let's let's let's see. This was free credits. This cost me nothing to make, and I didn't have to explain a lot of what I wanted because um the homies agent already had all the context about what we do. [08:25] Richard Webbe: But it did cost you tokens for a lot. [08:27] Dave Pengelley: No that well. [08:48] Dave Pengelley: So no, like I could have iterated, I could have spent I was like this was on Saturday night, I was trying to do family time and other bits and pieces, or I didn't spend a million hours really tweaking it hard, but just going, you know, I don't like that. Change that transition, make sure you use our C and O. The first one was Rue just doing five random things. I was like, that doesn't connect to anything. I wanted to finish on the logo, maybe do it, get him set setting the show up, and then send the signal out to the show. And that's like kind of that capsule thing that it did. But you know, for pretty lightweight prompting, just even Rubot. I'm so impressed. Like, I've got this one-page sheet that I have of like describes what Rubot is, and it used that and generated all of that. That's just wild. [09:30] Matt: cool like the like the [09:33] Matt: even just the idea of that being free like it's nuts even um like the minimax team just put out h3 which is something you can run locally to do video and it's it's actually like acceptable local uh but just blows my mind [09:50] Richard Webbe: Well it's funny, you know, I I've we've spoken about this a couple of times. Uh my children are very creative and they're eleven and fourteen and And uh, I think they're pretty smart, but I'm biased. And they hate AI. And the general consensus around the world from the youth is they hate AI because they see it's cheating the creative process and taking it away. And they're saying, and uh, I've mentioned this a couple of times, it'll be published soon. Uh, Dave's being very good in helping me publish my book. And uh they looked at the front cover, which has a real picture of me on it, but uh, I did get some help producing the font. Actually, most of us done. By day, but I had some help for it. And they're looking at me and go, Dad, we're not going to read your book because it's AI. And you've got it all done for free. You're supposed to use a creative or a design person to do it or something like that. And it's an interesting scenario that in the old days the youngest would run to the technology, but nowadays, a lot of children coming through school hate AI. They think it's going to ruin their lives. [10:50] Matt: It's the same with every industry though, right? Like you look at uh at a like an old school blacksmith, you know, he didn't go to Bunnings to buy his hammers. Yeah, you forged them all and and all the weird little tools and all the [11:03] Richard Webbe: Exactly. [11:03] Matt: strange little nails and [11:05] Richard Webbe: Exactly. [11:05] Matt: hooks and all the bits and pieces. So y yeah. [11:08] Richard Webbe: And you know what, you know, you're exactly you know what infuriated me? I'll get a bit animated here. What's been infuriating me the last couple of months is how all the all the protesters have moved off foreign wars and our weather, and they're now trying to demonize AI data centers. I mean, oh my God, what is wrong with these people? They're using it. There is nothing dangerous or bad about they're trying to come up with as many excuses as they can to say, oh, this is horrible and we shouldn't. I suppose it's just kind of like the kids worried about AI. I suppose it's just the community getting a bit scared and wanting to push back. But if I was to have, you know, one of them, one of the ladies on the news this morning referred to it as heavy industry. I said, Well, if you think I'm saying this, if you think that's heavy industry, let me place you next to some of the factories in Melbourne, which really do have. Heavy industry, and this person's bought a house next to a business park. They're complaining they're putting in a data center. It'll be quieter than a normal office. They're just, I don't get it. It is, it is, I suppose it's just a natural reaction from humans to not like change and be fearful of what they don't understand. [12:17] Matt: Yeah, and just to want to whinge. Like humans [12:19] Richard Webbe: Yeah. [12:19] Matt: just naturally pick on stuff and want to whinge, you know, whether you want to hear it or not. Um, that's that's kind of like where the concept of reality TV comes from. Like, people want to watch other people whinge like it's such a strange thing like the whole like point of the news is oh look what so and so did you know like wow like no one cares stop that just be do do what you want to do like build the world the way that you want to want to build it I don't know [12:45] Richard Webbe: We all got trapped in social media because social media feeds us stuff that we want to be enraged about, right on the money on what you said. So imagine if we had a feedback loop of a program where the script was written by AI depending on the response from the audience. No idea that [13:00] Dave Pengelley: So [13:01] Richard Webbe: we [13:01] Dave Pengelley: there was [13:01] Richard Webbe: create [13:01] Dave Pengelley: there was um on although there was a uh theme on that that I saw pop up and people complaining on LinkedIn [13:07] Richard Webbe: a [13:07] Dave Pengelley: around [13:07] Richard Webbe: lot [13:07] Dave Pengelley: Sam [13:07] Richard Webbe: of [13:08] Dave Pengelley: Altman made a comment, um Mr. OpenAI, about I found a great new use case for [13:12] Richard Webbe: the [13:12] Dave Pengelley: AI this week that [13:14] Richard Webbe: same. [13:14] Dave Pengelley: um you could hook it up to your calendar, your with uh your kids' calendar, and uh then it can see if they've got a soccer game or an art class or whatever, and then it could record like um do like a little mini podcast about their activity for them or something. Then people like, what didn't you? Talk to your kids. How about that? Um, [13:34] Richard Webbe: Exactly. [13:34] Dave Pengelley: so I think [13:35] Richard Webbe: And [13:35] Dave Pengelley: I think [13:35] Richard Webbe: you [13:35] Dave Pengelley: the [13:36] Richard Webbe: turn [13:36] Dave Pengelley: as find [13:36] Richard Webbe: through [13:36] Dave Pengelley: find the middle ground. My my kids aren't as anti-it as you are, Richard, as yours are, Richard. Um, they're a bit ambivalent around it. It's not they'll use Copilot here and there because that their school has [13:47] Richard Webbe: case. the [13:47] Dave Pengelley: office stuff, their emails, um, [13:50] Richard Webbe: Yeah. [13:50] Dave Pengelley: or they're allowed to use at school, but they're not you're really for or against it, they're just like it's there. Um, but it's funny because your uh cover, you mentioned your book. cover and I did a the first draft of it which was very simple then you you had the preview copy the last week that you showed us with the uh just title name picture of you and you were like oh I've designed I've designed a new one Dave and uh I [14:13] Richard Webbe: Sorry. [14:13] Dave Pengelley: was like did did you design a new one like or did you get AI to generate one and you're like I got ai to generate it and I was like [14:20] Richard Webbe: Well, hang on, to be fair, the original ones were designed by me, and then I got that design and I dropped it in to AI. And I dropped in the new photo. I said, could you help me arrange this so it looks like a professional book cover? And that's that's it. So it's not all AI. [14:37] Dave Pengelley: So what what what did you because this was what I said to you? I was like, because I can't just drop an AI design on the on the book cover. Like it needs to be higher fidelity, like the [14:44] Richard Webbe: Yes, [14:44] Dave Pengelley: rendering [14:45] Richard Webbe: I [14:45] Dave Pengelley: and the [14:45] Richard Webbe: know. [14:45] Dave Pengelley: rasterization [14:45] Richard Webbe: I know. [14:46] Dave Pengelley: of the AI stuff. So I spent you know a couple hours on the weekend manually trying to reconstruct this marvelous AI-generated image with tight fonts and the different different dimensions [14:55] Richard Webbe: You did a wonderful job, sir. [14:57] Dave Pengelley: in Canva. Um because Because I needed the high fidelity. I needed like, you know, vectors and rasters and so on. [15:04] Richard Webbe: Yeah. [15:04] Dave Pengelley: Um so like you can tell your kids, no, no, a real human, put that together. [15:10] Richard Webbe: I did tell them it was you. I said, I said, you know Dave, you've seen him on the podcast. He helped me generate this and do [15:16] Dave Pengelley: Yeah. [15:16] Richard Webbe: this. The reason my kids are so against AI is uh one of them's an actor and a budding uh writer, and the other one's an artist and a budding writer. And so they just [15:30] Dave Pengelley: Flicks for a couple hundred million dollars. [15:32] Richard Webbe: Wow. I bet you that cost him five bucks to build. Now that brings me to another point I want to raise with you guys, and I only raise these because you're the smartest guys in the room. I've been talking to a lot of people in the industry lately, and I've been getting across a lot of organizations, you know, uh Databricks, Pure Storage, you know, Anthropic, and I'm and I've got a lot of friends at work in those organizations, of course. And um I spoke to a couple of venture capitalists last week. And It seems to me the unspoken common thought is that software is going to crash to recession levels and internal building of your own software using AI, which is why AI was invented in the first place to write software, and using people like Matt and yourself to come and help companies articulate it. I mean, if you look at a giant ERP system, a giant word processing or communication system. Some companies are paying them millions and millions of dollars a year. I just grab a Matt and a Dave, bring them into the company. This is how we work. This is our value proposition. This is what we do. Well, can you just go and write software for us? Don't care about [16:44] Dave Pengelley: That's [16:44] Richard Webbe: anyone [16:45] Dave Pengelley: that's what [16:45] Richard Webbe: else. [16:45] Dave Pengelley: Matt does. [16:46] Richard Webbe: Yeah, I know. And I know it's I'm just one bring up, use uh API, and it's is the protocol MCP, Matt. [16:55] Dave Pengelley: It's a tape [16:56] Richard Webbe: The [16:56] Dave Pengelley: protocol. [16:56] Richard Webbe: context libraries, the matching of context libraries. [16:59] Matt: Yeah. [17:00] Richard Webbe: I [17:00] Matt: I [17:00] Richard Webbe: think [17:00] Matt: mean it [17:00] Richard Webbe: it's [17:00] Matt: depends. [17:00] Richard Webbe: called [17:02] Matt: Plug, you know, it it just allows agents to speak to APIs in a cleaner and more controlled manner. [17:09] Richard Webbe: Syntax matter, yeah, yeah. So so I understand that's what you know we do, that's what we're involved in, and I I love to do it on the business process side. But it occurred to me that software companies, their value is going to go through the floor. You [17:25] Matt: Not [17:25] Richard Webbe: don't [17:25] Matt: necessarily. [17:25] Richard Webbe: think so? [17:30] Matt: Scenario you put out, like why would a company even approach uh an independent developer that can rebuild something for them? Like, why would they do that in the first place? Means that there's something, yeah. Well, there's like either a cost-related issue or the the product, the product's not doing what they want. And if the product's not doing what they want and the company won't pivot and add those things for them, there's a few bits and pieces going on there. And my firmest, strongest belief here is that. Most enterprise software companies either don't work in the industry, that they're actually building software for their cost their customers, like the people that buy their tool, they don't use it. It's the same concept that we've spoken about before with some of these AI labs. You know, like do they use their tool that they're shipping to people to actually [18:21] Richard Webbe: Yeah. [18:21] Matt: run and use day-to-day? Um, and and that's the biggest difference. So, like you just said, if you approach an independent that actually cares, that actually Can come in and is equivalently capable, let's say, maybe less funded, but equivalently capable and cares about you and your business and your business processes, and making sure that your team can actually get appropriately lubricated and do the work that they need to do. That's that's the difference. [18:49] Richard Webbe: Yeah, there we go. [18:56] Dave Pengelley: And they're investing and they've got a team and funding and so on. It's why why does anyone go Get seed capital. It's because if you are trying to grow a product to be maximally effective in a reasonable amount of time, you need more than you know one guy with a clawed code session. You need a team in order to build it. Like, I know AI has changed the game, Matt, but still at the end of the day, like if you're trying to build a production thing and scale and have sales teams and support teams and all the other bits and pieces, you can't bootstrap that as quickly and effectively as a one person team as if you get a couple million dollars in funding to actually hire a group of people to work collectively. Collaborate collaboratively on it together. And that's a business building that thing and running that thing and focusing on product development, program management, product development, all the things that go on with that versus, you know, one potentially single point of failure. Um, and that comes back to all the disaster recovery stuff we talk about as well. So businesses do need to decide, are they in the business of being software companies and writing their own software? Or are they going to take good enough, you know, done well, done fast, done cheap, pick Pick the two that fit them both best and um build [20:03] Richard Webbe: And [20:03] Dave Pengelley: a [20:03] Richard Webbe: that [20:03] Dave Pengelley: bar. [20:03] Richard Webbe: yeah, but that's my point. And and I think you've both nailed it, and and and it's what you guys are in the business, or we're all in the business, of doing of improving productivity, efficiency, margin return to the investor. I just think some software companies, particularly the monolithic ones, have had it too good for too long. I mean, you know, they've got maintenance support contracts that run into the tens of millions of dollars. And why? Because they own the source code. If I Could grab, you know, you and Matt and say, hey, this is my business. I don't want to make you know Salesforce or HubSpot fit my business. I just want a CRM that works for my business. So you guys go vibe code me one. I understand we're going to have some money mistakes and it's going to bounce around. But you guys aren't going to charge me tens of millions of dollars. And because of AI, you can do it for me at scale. Now, the big proviso on all of this is where is the business process? Process, intellectual property sitting, the knowledge of time and motion studies of how to improve things. And all fairness to those monolithic software companies, they do have great specialists that understand if I'm running a steel company or I'm running a services company, they know what you need, right? But there are other people out there like that. So I think the answer to the question is: prices from software companies are going to have to come down substantially, particularly for support. I spoke to someone, I'm trying not to use company names, but I spoke to someone in Canberra the other day, and their job is to re-sign the maintenance and support for a couple of large monolithic software companies that are involved in the hypervisor game and stuff. And the costs are like, you know, what was 500,000 was three million this year. Yet there's no benefit to the customer. Three million. I just run and knock on YouTube front door and say, right, emulate that. Copy that, Vibcode be, you know, an app and an API that automatically adjusts itself between that software and that, and I want a screen that presents like that. And I and I know that you guys, I know with APIs and MCP, I know with AI, you can do that. Very reliable. There'll be mistakes, but will it be $10 million worth of mistakes? I don't think so. I think it'll be able to, you know, I could hire you guys for a lot less than $3 million a year. Could I? Maybe I can't. [22:30] Matt: If the if everything keeps scaling the way that it is right now, that might be my cost pretty soon. Um, [22:35] Richard Webbe: Well, tokenomics, as Dave always refers to, can can catch everyone out. I just think we're on the precipice in the IT industry of a massive shift. We were in hardware for many years. Oh, someone said on the SAS-to-ass agents as a service, I'm seeing a massive rise in interconnected debt. Yes, exactly. It's a technical debt now being created at scale. [22:59] Matt: I a good idea. that's think [23:00] Richard Webbe: Exactly. So all my friends who are in the hardware, networking, acceleration, sort of stuff, and cybersecurity, I might add, they can't get enough equipment. All the people [23:14] Dave Pengelley: But [23:14] Richard Webbe: in the software game that are trying to sign really expensive large maintenance contracts, they're all getting pushback. [23:20] Dave Pengelley: I think what what Doug's saying is is by rushing to take software as a service and create agents and get agents to do everything instead, that You've got all these dependency chains that are being created and it's creating technical debt very quickly. So you're all of a sudden making big mistakes very quickly, and you're stacking on top of those because you moved so fast. You didn't have the foundation right. And so that means yes, you move fast, yes, you got a solution, but now it's got underlying technical debt and problems in the base infrastructure that do need to be addressed probably sooner than later. But now you're basically gonna have to go back and rebuild and re architect a whole bunch of stuff again because you moved so fast. And so I think this is all there's all good. Things that people need to think about. And it's one of these things people go, Yeah, I'll just vibe code a solution. But it is more complicated than that. There are so many interconnected factors here as far as, and this is what that's why we always say, hang on, why are you chasing that shiny object? What is the return you're going to get for doing that thing? Is that the right thing now? Let's structure this, let's build a plan, let's think about the order of execution and go, where are you actually going to see returns from implementing some kind of agentic changes or using an agent to build a new tool or swapping from this tool? to that tool. [24:30] Richard Webbe: Yeah, [24:31] Dave Pengelley: Let's [24:31] Richard Webbe: there's a there's a [24:31] Dave Pengelley: make [24:32] Richard Webbe: happy [24:32] Dave Pengelley: sure it [24:32] Richard Webbe: medium. [24:32] Dave Pengelley: pays [24:32] Richard Webbe: You're [24:32] Dave Pengelley: for itself. [24:33] Richard Webbe: you're and Doug exactly right. It's all about return on investment, total cost of ownership. Can you support it? Can you sustain it? And don't move too quickly. [24:41] Dave Pengelley: So Richard, Richard, you you referred to someone, I'll I'll I'll save name, but um someone you know uh reached out and had a chat with me the other day and um they're in in program and project management. And they um they were asking sort of around AI and how AI fits across that sort of industry. And I think like most people They were under the mistaken impression that to provide value in this AR world, they need to be an expert using AI project tools. And I said, hang on, no, no, no. Your value is a project and program manager where you can slow things down and you can actually make sure you've gathered requirements and understood the goal of the project and make sure that you're executing things and you've assessed and understood all the workflows and things, and make sure that you do that in a structured, managed way to actually drive an outcome that you've measured your baselines and you know what success looks like. You work the project towards that, that is your value, not that you know how to use some AI tool, but that you can make sure when anyone's chasing any of these AI genies, that they're actually doing it in a structured, meaningful, purposeful way that's going to deliver value, not just be like we talk about just theatrics. [25:47] Richard Webbe: Exactly, exactly. I always think of um the stand-up comedian who's quite known in Melbourne and Sydney, I think, uh Elliot Epstein, who's now a business consultant. And he's hilariously funny. So if you ever want to hire, oh Elliot's a friend, if you ever want to hire him. He comes in, sits your sales team or your group team or your project team down and he just starts talking and he throws jokes in. So it's very entertaining and very good. And his saying is you've got to know your business better than your customer. So if you're in the IT department, right, and you're going to make these changes or you're going to start vibe coding or moving things around to save money, you must know the business better than the technology before you start applying it. And as you said, Dave, you don't need to be a Specialist on AI to be a project manager that leverages the use of AI. [26:35] Matt: Exactly. It's there's fundamentals with all of this stuff. You don't have to become an expert in anything anymore. Like you just gotta be an expert in you and chase your interests and values. Like I find it incredible. Like we talked about the younger generation before. I did a bit of a uh like an internship with one of my younger brothers. Brought them in, tried to give them a bit of a an insight into what I do, you know, just the base fundamentals of of what's going on. And he is quite, he's quite a technical person. He he was at uni doing cybersecurity. He's deferred and he's doing other things in the meantime. But, you know, and I'll show him what I do. And I found it absolutely fascinating that he was leaning towards that opinion as well. It's like, I don't, I don't know enough of this yet. I don't like, I need to go and do a course, you know, I need to go and And read some guides. I need to find forums. It's so funny. Like he was reverting to this old, you know, dogmatic institutional knowledge method of like how you learn things. I'm like, dude, you literally have the source of every piece of information you need to do this right here, plus my you know, appendix and resources for you as well. And yeah, it was incredible. And it [27:55] Richard Webbe: It's [27:55] Matt: just [27:55] Richard Webbe: like [27:55] Matt: it [27:55] Richard Webbe: it's [27:55] Matt: was [27:56] Richard Webbe: like your [27:56] Matt: like [27:56] Richard Webbe: hammer story. It's like your hammer story. He to be a carpenter, you don't know. Ha they don't need to know how to build a hammer. [28:02] Matt: Yeah, exactly. And like just to give Doug some credit as well, again, on his comments there, it's so true. And it's exactly what we well, it's what I try and speak about a lot is you know, that whole idea of technical debt is because you've moved too fast and you did not have a solid grasp of the principled foundational fundamentals. You know, the whole idea of actually building things properly. You only get technical debt if you don't understand the system. [28:30] Richard Webbe: Formed an IT opinion group, believe it or not. And everyone was running around and go, you know, I've got a DEC 360, or I've got an IBM 380, or I've got a, you know, everyone was showing off what hardware they'd bought. And they actually had to come out with a white paper that says, look at your business, find the software that will do what your business needs, then buy the hardware. And I remember back then looking at it thinking, that's so insane having to tell people, don't put the cart before the horse. And it is, you can put the cat [29:00] Dave Pengelley: I [29:00] Richard Webbe: before [29:00] Dave Pengelley: mean, [29:00] Richard Webbe: the [29:00] Dave Pengelley: that's [29:00] Richard Webbe: horse. [29:01] Dave Pengelley: that's you meant you mentioned Clarkson in his farm earlier. Like he went and bought the massive Lamborghini tractor that didn't fit in his chair that was too big for what he needed for half the stuff, and [29:10] Richard Webbe: That's exactly [29:10] Dave Pengelley: because he went and [29:11] Richard Webbe: it. [29:11] Dave Pengelley: bought the hardware without understanding the job that it had to do. [29:14] Richard Webbe: It is, it is. So what's your problem? What's the problem you're trying to solve? Map that out for your business. Map where the benefits are and work backwards from there. And as you guys say, some along along the way, there'll be some guys like you or a bit of tools in AI that will help you build The bits that need to connect that together. [29:31] Dave Pengelley: But but people I get are getting confused and they don't understand. Like, what is an agent? When do I need an agent? Um, what what is the model? Which is the best one, which one's gonna do the job the most? Like all these different questions. And and Matt, you you you like this one. I I think this is a pithy truism or aphorism or whatever the that little sentence fragment's called. But uh they were asking me about um agents, and I said, and I as as we talk about on the show all the all the time, right? Like, this is just stuff we talk about, like models versus. Harnesses versus so on. And I said, every agent is a harness, but not every harness is an agent. [30:09] Richard Webbe: Oh, that's some wisdom in those words there, Dive. [30:12] Dave Pengelley: Yeah. Because you know, chatgpt.com is a harness, the web UI that you use, but you wouldn't necessarily call that an agentic interface. It's getting there with you know all the connectors and plugins and so on. It's much more than it was two years ago. Um but yeah, you can certainly have harnesses that are not necessarily agentic. But if you're running an agent, that is essentially a harness that's controlling the uh the model. [30:35] Richard Webbe: You know, that's something I did the other day that I really found, you know, I I know we like to talk business productivity and we also like to talk personal productivity, but I was going to have a meeting with someone, and I wasn't sure how the meeting was going to go. So I got onto Claude, don't mind giving them a plug, and I said, Okay, and I used the voice activation parts as my uh interface, and I said to it, I'm about to have a meeting. Meeting with this person, their angle and where they're coming from and issues are these, mine are these, I need this outcome. Can we please practice how the conversation might go? And we did it. While I'm driving all the way to the meeting, and it was astoundingly good. [31:18] Dave Pengelley: It's [31:19] Matt: Yeah, [31:19] Dave Pengelley: great. Yeah, [31:19] Matt: I [31:20] Dave Pengelley: it's [31:20] Matt: agree [31:20] Dave Pengelley: great. I [31:20] Matt: with [31:20] Dave Pengelley: agree. [31:20] Matt: that. [31:21] Dave Pengelley: Yeah. [31:21] Matt: Like I actually have another similar a similar win. Um my father, who is in his mid-70s. Um I've recently exposed him to to Jibity and I got it to got him to download it on his phone [31:36] Richard Webbe: Very quickly, [31:37] Matt: and [31:37] Richard Webbe: watch Dibity, just for everyone, watch Dibity. [31:39] Matt: uh it is the colloquial term for Chat GPT. [31:42] Richard Webbe: Oh good. Okay, thank [31:44] Matt: And [31:44] Richard Webbe: you. [31:44] Matt: um [31:44] Richard Webbe: I look like such an idiot. Go on. [31:47] Matt: and anyway, my my dad struggles to articulate things and I was putting him on the spot and making him kind of embarrassed to to speak to this thing, you know, this this computer on his phone. And I was like, just talk about something that you want to work on. Like, what's that solar project that you've got going on? And we activated the the voice mode, and and I was like, just tell it what you're thinking of doing. And then the new live voice mode in that ChatGPT app is very, very good. And it put him so much at ease. And he was able to just think through his project that he's got going on. And he he made massive mental strides to start moving this thing forward. Whereas previously, you know, that that big sheet. That big fear aspect of, you know, what do I do? I don't know. Where do I go? Do I have to speak to someone else? Like, how do I do this? I think that level of just unlock is very, very, really powerful. And yeah, the live voice stuff for rehearsing and just speaking your mind is really cool too. [32:49] Richard Webbe: It is, it is very cool. And they were talking actually, funny enough, uh the technology specialist on uh I think the Today show this morning. When you write something, you use a different part of your brain. To when you speak about something. The writing is about planning. The speaking is about, I need to understand this right now. It uses different parts of [33:09] Dave Pengelley: But [33:09] Richard Webbe: your brain. [33:10] Dave Pengelley: uh the [33:10] Richard Webbe: And [33:10] Dave Pengelley: today's show must watch our podcast because we talked about that last [33:13] Richard Webbe: we [33:13] Dave Pengelley: week. [33:13] Matt: Yeah, [33:13] Richard Webbe: did. [33:13] Matt: we did. [33:14] Richard Webbe: We did. I'm pretty sure they did. I'm [33:16] Dave Pengelley: Yeah. [33:17] Richard Webbe: pretty sure. [33:18] Dave Pengelley: That that that's Trevor, I think. Anyway, um Matt, [33:20] Richard Webbe: It [33:20] Dave Pengelley: were [33:20] Richard Webbe: was [33:20] Dave Pengelley: you gonna say something about the agents and harnesses stuff? [33:23] Matt: I [33:23] Dave Pengelley: You look [33:23] Matt: was, [33:23] Dave Pengelley: like you're [33:23] Matt: yeah, [33:24] Dave Pengelley: better. [33:24] Matt: I was gonna briefly touch on it. I don't want to go too super technical, but you know, [33:28] Richard Webbe: do [33:28] Matt: if [33:28] Richard Webbe: it. [33:28] Matt: you if you think of when the the old version of chat GPT, let's just say the old, old version where you are just, it's a chat bot, right? You know, you you send a message and you get a response. There's nothing else that happens in between other than the LLM inference. You know, the difference between that and then an agent is there's you know a question. It's it the question to the to the LLM is are we done yet? And then it'll it can say yes or no. And it just keeps looping until it says yes. So that's an agent. And you know the what what does it do? Those loops, it it's causing it's calling tools, and you know, those tools could be MCPs, it could be whatever. So it, you know, the agent is just a loop. The harness is when you add control to the agent, it's [34:12] Richard Webbe: Uh [34:12] Matt: when you add guardrails, it's when you add specific cases and instructions and hooks and all these other little bits and pieces that can happen in that agentic loop. That's [34:22] Richard Webbe: That's [34:23] Matt: probably [34:23] Richard Webbe: a [34:23] Matt: the [34:23] Richard Webbe: great [34:23] Matt: best [34:23] Richard Webbe: definition. [34:23] Matt: way. [34:24] Richard Webbe: And that relates to what we were talking off air when Dave told us to stop because we're sp talking about all the good stuff. And that is if we um, you know, with all the breakouts of AI agents breaking out and doing illegal things and stealing code and and you know, breaching all these things, we were talking prior to the program, it's like they're like two year olds, or it's like water. You need to self contain what it's doing. And I think we use the example as if you ask someone to go and build your garden because you want all greenery and more light, but it has no guardrails because it has no moral compass. It's more about telling it what You can't, it shouldn't do than what it should do. That you've got to be careful of. For instance, if I let a gardener go loose in my garden here and said, Look, I want more light, I want green there, I want green there. It could very well knock down the house next door to let in more light for me. It could very well steal all the prants from all my neighbors, right? And and that's what happens if we don't, as you said, with the harness have the right guardrails and the right guards that match our moral imperatives and our laws, we could end up in a lot of trouble. [35:30] Matt: Very interesting. I I really liked your analogy about the water too. Like if you have a a funny-shaped pond or a bucket that's, you know, a really weird shape, you know, that water will flow based on gravity wherever it goes. But you [35:43] Richard Webbe: Yeah. [35:43] Matt: know, it's it's a it's just a good analogy. I like it. [35:48] Richard Webbe: Yeah, oh [35:48] Dave Pengelley: Yeah. [35:49] Richard Webbe: thank you. [35:49] Dave Pengelley: We we I mean uh we we're we're seeing as as this sort of story continues and we're and we've been talking about this for months now, that the harnesses of where the war's gonna happen. Like the that's the moat. Um Models are all catching up. The models are all quite good. Uh, and they'll get the answer. They might take a few more turns, they might reason loop a few extra times, right? But whether it's Kimmy or now Quen or GLM or Opus or Fable or Terra or Sol, like they're all getting to a point where with enough coaxing, you'll get good outputs, you'll be able to create code, you'll be able to generate images, you can do all the different things, right? Generally speaking, that you need. Um, but it's how do you manage and control that? How do you manage and control that across multiple teams and that sort of scaling issue and so on? And then the Harnesses, whether that's Claude Code or Claude Manage Agents or Chat GPT and what they're doing with their shared sites and other bits and pieces, or Hermes. And I tried building my own one, Sargent, which we've talked about, which died. And that's interesting because I was trying to build it as a one-off guy who didn't really understand the depth of what I was doing when I started. I just saw OpenClaw and went, Oh, oh, I can do one of those. No, I couldn't. Uh, and then like Hermes has this whole research team, news research behind it. And I saw today um a guy I know on LinkedIn. In sharing this thing called KiroCrew, [37:05] Dave Pengelley: which came off the back of a group at Amazon. So [37:09] Matt: Yeah, [37:09] Dave Pengelley: uh at AWS, [37:11] Matt: something. Yeah. [37:12] Dave Pengelley: um, they um have spun out this thing called KiroCrew, which again is another one of these um tools for having whole teams of people running multiple sessions across remote agents, persistent working, shared context across different team members and stuff. And I think this is we're gonna see more. Of these sorts of things too. That like how do you get not just a standalone coder, software engineer vibe coding on their own stuff, but how do you actually start managing that across fleets of agents, fleets of people, all working together across common goals to build something bigger? That's the kind of thing that's going to enable, you know, as you predicted, Richard, the death of software by having better automated factory systems for managing, developing. And so Kiro's the latest one, obviously. There's things like um was it Gas Town, uh, which has been out for a little while, that is in that idea [38:06] Matt: There's [38:07] Dave Pengelley: of [38:07] Matt: there's too many. You can't even say that Kiro is new. Kiro's been around for for [38:11] Dave Pengelley: no, [38:11] Matt: like [38:12] Dave Pengelley: but [38:12] Matt: a [38:12] Dave Pengelley: they've [38:12] Matt: year [38:12] Dave Pengelley: they've [38:12] Matt: now. [38:12] Dave Pengelley: just I've just hit a point where they're saying this is a major release or something like that. They're saying like it's been around, but four hours ago they're saying now 39,000 developers providing PRs and so on back to it. [38:23] Matt: It [38:23] Dave Pengelley: We're [38:23] Matt: brings up [38:23] Dave Pengelley: kind [38:23] Matt: a [38:23] Dave Pengelley: of [38:23] Matt: good [38:24] Dave Pengelley: happy [38:24] Matt: point [38:24] Dave Pengelley: with [38:24] Matt: though. [38:24] Dave Pengelley: it. [38:24] Matt: Like if there's constantly newer and newer ones of these and revisions of old ones. You know, is the harness actually any better than a different one? And what would be the reasons either for or against that claim? And is it the case that it might be that certain organizations and companies or businesses will prefer a certain harness? Or, you know, is it better suited to a slightly different perspective? Because, like you said, that that kind of harness thing, it's probably not something you want to go try and build yourself. Um, just the Amount of complexity and like Doug's technical debt. But [39:03] Dave Pengelley: Yeah. [39:03] Matt: the the idea of you know just using one in your particular situation and then going, wait, maybe I should use this one. Wait, maybe this one is better [39:14] Dave Pengelley: I think there [39:14] Matt: for [39:14] Dave Pengelley: comes a [39:15] Matt: me. [39:15] Dave Pengelley: point where you just gotta lock in and run with what you've got. Like for a project and go, okay, we'll do a uh uh what's the word after the project, uh review and I can't remember the name of it, [39:26] Matt: Another [39:26] Dave Pengelley: uh post [39:26] Matt: project. [39:26] Dave Pengelley: mortem. Um and and decide whether Whether we would continue on that for the next project. So you'd start small with like a small scale thing, potentially. But, you know, why would you pick one over the other? And just looking at the way they pitch them and the advertising for whether it's um Kiro or Gastown or Hermes, Hermes are pushing into this space. They talk about, you know, token efficiency, memory utilization, remembering stuff, um, using command line tools and using things in more deterministic ways that don't spend tokens. And I think as tokens, are tokens going to be more expensive? And is there going to be the massive token wars or is everything becoming commoditized? And we saw Chat Gippity drop the Lunar model down to like ridiculously cheap. Like, this is this, we're still in this flux zone where we don't [40:16] Richard Webbe: Very [40:16] Dave Pengelley: know is [40:17] Richard Webbe: a very [40:17] Dave Pengelley: where [40:17] Richard Webbe: dynamic [40:17] Dave Pengelley: where is [40:18] Richard Webbe: market. [40:18] Dave Pengelley: the cost going to be? And so, [40:19] Richard Webbe: Yeah. [40:19] Dave Pengelley: like, does it really matter how how important is token efficiency versus harness efficiency versus time? Um, Hermes keep pushing, you know, time to first prompt is 0.9 seconds instead of five seconds. It's like, oh. Okay, but that's the first prom, just get working and run [40:32] Richard Webbe: Yeah, how did [40:33] Dave Pengelley: on [40:33] Richard Webbe: how [40:33] Dave Pengelley: it. [40:33] Richard Webbe: does that help me? [40:34] Dave Pengelley: But [40:34] Richard Webbe: I Phil Phil just fill Phil just made a comment in the comment section, which I think is very relevant. Now I wanted to clarify what you said um earlier. When I say the death of software, I say the death of monolithic software is what I meant. Like really big, I buy one bit of software from everyone. And Phil's made the comment, of course, harnesses and you know, uh different agents and that are the new monoliths. And [40:57] Dave Pengelley: I [40:57] Richard Webbe: I think [40:57] Dave Pengelley: don't [40:57] Richard Webbe: that's very true. All we've got to do is work [40:59] Dave Pengelley: know. [40:59] Richard Webbe: out Uh, how fast to deploy? What are we is the problem we're trying to solve, and what's the return on investment over time of using that tool? And it's very dynamic market. So one minute we've got a tool that does this much. And as you say, Dave, next minute there's a tool that does this much. And you go, I want to swap that out for that. But as Matt was pointing out, how long will it take to do that and and and change it over? So we are living, that's why I think you guys are the most important people. At the cornerstone of this whole evolution that we're going through. And that is you need context of the tools and how quick to deploy, and you need that experience. [41:42] Dave Pengelley: So, [41:42] Matt: It's also [41:42] Dave Pengelley: Matt, I'm I'm gonna ask you for some live tech support on this uh while while we're on this topic, I think it's relevant. So I [41:47] Matt: a [41:48] Dave Pengelley: decided I was gonna try and rebuild a monolithic piece of software for a friend the other day. Um and I worked with [41:54] Matt: good idea. [41:54] Dave Pengelley: GPT Soul Pro Plus Mega on the web. Interface that I you can't access via the agent, I think. It's only via the web interface. And it went and built me a 70-page specifications document on what it would take to reproduce and recreate a better new modern version of my own version of this particular monolith. Um, and then I said, Okay, well, that's great. I don't need a 70-page Word document, I need to give this to agents. So it sort of built out a repo structure with agent MDs and did all the bits and pieces. Uh, and I was really big on some of the stuff we talked around around back pressure. Automated testing, really trying to make sure this was highly rigorously built so that way it wouldn't progress past gates without double checking and triple checking itself because obviously I'm not reading a line of code that it's making. So I put all of this into this build infrastructure and then went to the Chat GPD codex app. I went, you know what? This is, you know, I want to make this good for my friend. Um I want to use, you know, the best in class codex software. So I threw it in there. And not only did I burn through my token. Ridiculously quickly and keep running out of quota. Fortunately, last week Chat GPD just kept resetting it for us all the time. Um, but that was that was helpful. But I was concerned that like because it's doing so much work verifying and double checking and doing things, and and I had this issue obviously on my own local hardware stack where I'm in Windows, but there's WSL and Windows commands for a lot of this development stuff that not as good in PowerShell. So I was trying to, and I'm like, Am I wasting tokens? Because it's just trying to hack my or Operating system and grind commands that it can't do this text file bash operations and stuff properly in PowerShell and it's got to go to and so I'm like, oh, is this or should I have just spun up a fresh Hermes instance and just run Hermes? Like, what is the benefit in this instance of me going pure Codex app? And is that actually costing me more tokens and wasting my time? And then I'm stuck only developing technically. I know I can swap things, but trying to keep consistency. Um only using Codex and Hermes, I could plug. model in when I run out of tokens and keep going and it's gonna have all that continuous memory and context and stuff. So I don't know, Matt, you're uh you're you're big into the chat GPT and the codex land stuff at the moment. Where would like am I wasting well I'm not wasting time I'm just what what's your point of view on that little story I just told [44:19] Matt: For I guess just before I go down these rabbit holes, what question do you want like actually answered here? Like what's [44:27] Dave Pengelley: me. [44:27] Matt: the what's the takeaway? [44:28] Dave Pengelley: I'm I'm where where I'm talking On is do I continue just pushing through on the the Codex app, which I expected and was hoping, I was hoping like this is separate from everything else I'm doing. So I set up with clean accounts and a clean superbase and a clean GitHub and everything. And I use the the Codex plugins and connected it to Superbase and connected it to code to and it still wanted to then get CLIs and do everything else. I'm like, I gave you your own native plugins, what are you doing? And it just kept and so the I thought, you know, I'll do it through the The chat GPT professional open AI special development harness, and it was not as smooth and professional as I expected it to be. And so that's where I'm like left going. Maybe I should have just used Hermes. [45:12] Matt: Okay, [45:12] Dave Pengelley: Um [45:12] Matt: my answer is gonna be very annoying, and it's that it's based on your setup. It's not based on your prompts, it's not based on the lack of plugins, it's not based on the [45:25] Dave Pengelley: which [45:25] Matt: model, [45:25] Dave Pengelley: which which part of the [45:26] Matt: it's [45:26] Dave Pengelley: setup? [45:27] Matt: your your lack of attention to detail to the setup itself. [45:33] Matt: Now does that make sense? Because that might not. [45:37] Matt: The back pressure, the instructions, the guidance, you know, all of those various aspects that stop the thing from hunting. [45:44] Dave Pengelley: Cause I spent hours and hours and hours trying to get all of that set up ahead of the the hitting the go button, right? Making sure it had all like it had this whole phase zero of the setup to get all the connections in place, to make sure it was wired up, and I thought I'd ticked all the boxes and Done everything I'd I I was meant to. Obviously I didn't because then it spent another 12 hours trying to process working through additional connection issues with phase zero. And maybe I've got because I automated too much of it and I've got some conflicting back pressures. I did also get it to use the Groc CLI in order to have a verification agent that's a third party that's doing stuff. So I've also harnessed that out, which is great. Um but I think one of the big cons the big issues I have is The whole Windows development thing. And I think the agents they prefer Linux backends and even the Codex app, even though it's on Windows, built into Windows, installs into Windows, it's a whole MSI installer. Once the agents start developing, they don't want to be in Windows. [46:48] Matt: So again, I'm gonna say it's a setup thing because if you have a task that needs to be done or a task that can be done, let's just picture a a corporate example, like you know a people example and if you go to a new company, you know, like like this place, um, if you go and and get a job at Bundings, you know, they're gonna make you sign a whole bunch of safe work practices and get you to basically study what your duties and your standard operating procedures are for your role. Now like why does any of that bureaucracy crap exist? It's to stop people from trying to figure things out for themselves and to make sure that they follow the predetermined path Path. Now, your agents need predetermined paths. Yes, they're intelligent and they can go and figure out new ways to do things, but that's kind of not what you want. You know, you want them to access your PowerShell through this method with this particular script, not by just YOLOing PowerShell script uh bash commands, you know. Um, and then the same thing, you know, between your WSL environment and Linux stuff. And if you did touch your Mac again, you know, all of these patterns need to be defined ahead of Time. [48:06] Richard Webbe: So Matt, you're saying to paraphrase, correct me if I'm wrong, it's a new way of thinking. And if you're using new tools, but you're thinking in old processes, you're gonna get yourself bogged down. [48:17] Dave Pengelley: I [48:17] Matt: Yes. [48:17] Dave Pengelley: think and and I th I I I get what you're saying. I didn't I I I assumed that and this is the problem, you know, you make a make an assumption that being in the Codex Harness, it would have enough innate knowledge about being the code. Codex harness that it would know things. That where when I told it, hey, you have the Superbase connector, but for whatever reason, it hasn't stored that, written that, updated that in the right agents MD files and the right specifications. So when it wants to go talk to Superbase or talk to GitHub, it has to reimagine that every time. And even though I yell at it and tell it, you have the access, remember you have the access. Because I was not specific enough in my agent setup around the environment. Environment and the environmental restrictions that it had, that's where I've caused the gap. So as much as I went to the effort of making sure it had the. [49:26] Matt: Yes. Imagine you you you want the agent to smash a paint. glass, what's it gonna do? Just charge at it with its head or it's gonna look around and like see, okay, there's a screwdriver here. I can probably try and like jab it at it. Or you're gonna say, all right, when you need to smash a pane of glass, there is the glass smashing hammer that is kept in this drawer and you need to use it with your right hand and swing it at this speed. That's the kind of like procedural um unlock of information. It's called progressive disclosure. You know, it's that that's the concept of a skill file as well By the way. So, you know, imagine your skill file is essentially an SOP for a situation that you may incur. And then if this situation arises, here's the instructions and here's the reference files that you can selectively go and read and learn more about. [50:17] Dave Pengelley: Yeah. [50:18] Matt: Um that that's that's what I'm talking about. And I spend [50:21] Dave Pengelley: Yeah. [50:21] Matt: so so long on that stuff. [50:23] Dave Pengelley: Yeah, I I th I thought I spent a long time on it, but clearly there were gaps in my setting up of it, even though I [50:30] Matt: The [50:30] Dave Pengelley: Yeah. [50:30] Matt: other thing too, like if I can just add on something controversial here, don't trust the agents. Don't trust that they know what they're doing. Don't trust that they understand their harness. Don't trust that their plan for you is actually correct or right or even feasible. Don't trust that any of their response to you is actually going to help you in the right direction. You need to force that. [50:55] Richard Webbe: I uh I um I like that statement 'cause people once They feel the technology is quite smart, they give it too much credit and they don't see the blind spots. The funniest story I have of that, no one died was an old lady who bought a camper van in the US and uh put on cruise control, thinking it knew to stop at red lights. [51:17] Matt: Hmm. [51:17] Richard Webbe: In the old days. [51:19] Matt: Interesting. [51:20] Richard Webbe: Yeah, [51:20] Matt: But [51:20] Richard Webbe: it [51:20] Matt: yeah, [51:21] Richard Webbe: didn't. [51:22] Matt: there's there's another quick red flag, Dave, and I'm sure people listening and watching have done this as well. Don't put plain text API. Tokens in your.e and v files. Just don't put any plain text API token in any of your files, regardless if it's in a dot file or not. You should be using some sort of secrets manager. There's plenty out there. Go research with your agents about them. I use BWS. It's not beer, wine, and spirits, it's Bitwarden Secrets Manager. And yeah, that as well as Parse. Parse is another really good one that's in Linux environments. So you should basically never ever Have a plain text API key in any of your files because this is exactly how these massive supply chain worms are getting spread. Uh you know, poor people at companies being maintainers for a product, they get their GitHub personal access token basically ripped from some supply chain attack, and then it gets propagated because they now have access to their GitHub on all their various things, and they can they can push a new version and it gets spread to everyone else when they install it. So Yeah, try not to ever, ever, ever use plain text tokens, [52:33] Dave Pengelley: Which [52:33] Matt: please. [52:33] Dave Pengelley: which which is actually, I think uh I misspoke. I didn't actually give it.e and v things. That's why I was trying to use the proper plugins via codecs where OWAST because I was trying to avoid all of that and I've used the codex harness for doing all its things. But it was one of those ones where you know I was like, I was I was my mate asked me to do this thing, and I was like, uh no, I don't I think that's too big. I think you know, one guy building that, and then I was like, ah, what the heck? We'll just run for it. And then I spent hours. Doing all the spec stuff and got to the point where I think this must be ready to go now. And then just kept hitting my head against the wall and burning all of my tokens so quickly. And then you're like, ah, now I can't do anything while I wait for tokens to reset. And so the project pauses and oh, I gotta reset. Good, I'll give it another run. And then ah, that still hasn't pushed a single table to superbase yet because it's still just fighting itself. And I'm like, is it me? Am I the problem? It's me, it's probably me. [53:26] Matt: I [53:26] Dave Pengelley: But [53:26] Matt: don't think [53:26] Dave Pengelley: um [53:26] Matt: it's you. Uh it's definitely not you. It it's [53:28] Dave Pengelley: it's [53:28] Matt: just it's just assumption. Yes, that's all there [53:30] Dave Pengelley: But [53:31] Matt: is. [53:31] Dave Pengelley: and and and the thing is because it is working so hard and doing so much behind the scenes and it is back pressuring itself, it started building some stuff. It has pushed a few web pages out. Like it's probably doing some really good stuff in the background that I can't see. It's just using a lot of SOL tokens in order to do it. And so progress feels slow because I'm waiting for tokens to reset and do things. So there's that impatience piece as well. If you're not willing to pay for as many tokens as you need to push forward on these things. So um yeah, maybe maybe it's working fine. I don't know. I I just I th I thought, you know, for all those people out there that are trying to build things, it was worth just maybe talking about them on the topic of monolithic software and what it takes to recreate it. Um, there's some effort involved. [54:14] Matt: It's [54:14] Dave Pengelley: It's [54:14] Matt: a [54:14] Dave Pengelley: not [54:14] Matt: huge [54:14] Dave Pengelley: just [54:15] Matt: topic. [54:15] Richard Webbe: It's not stable, of course. [54:17] Matt: Yeah, and and like we can keep bringing this up every week if you want, because it's it's what I'm the most passionate about. But like I don't trust any of my agents, you know, individually like at all. Like I trust my validation. Systems. I trust my my back pressure. I trust my loops. I trust everything that I've spent the most time engineering. I think we spoke last week about Uncle Bob, guy behind Clean Code. And it's kind of funny that he's pushing these concepts that now people are slowly adopting or extremely vehemently rejecting, which is the concept that like engineering now is less about what code's written and it's more about About how the code is checked and validated. And you should not really bother spending time fussing about how it's generated. If anything, you should be generating more and then deleting and reviewing and rejecting that through your validation systems. [55:16] Dave Pengelley: Yeah. [55:16] Matt: So yeah, [55:17] Dave Pengelley: Which which [55:17] Matt: that's [55:17] Dave Pengelley: is why [55:18] Matt: great. [55:18] Dave Pengelley: I spent so much time with this spec, trying to actually put in all those back pressure and unit testing systems and make sure the agents knew they couldn't push anything before it had ticked all these boxes of preparation and readiness and had passed it over to another agent for. review and come back again. And so yeah, try trying to take all that feedback on board, which is why it's feels so slow because there's so much validation and looping and making sure it's right versus and you know this is probably far more complex than anything else I've built before. So it is a different level of work versus some of my other systems, even my SAS, which was much simpler than this sort of monolithic data structure I'm trying to recreate, where I could iterate much more quickly and see results much faster on that and had a clearer vision for exactly what that was. Sorry, this is that's a whole different base. [56:03] Matt: 100%. [56:03] Richard Webbe: Yeah. [56:03] Matt: It's even crazier once the thing's live and you need to continue continue to build it once it's live. Um, you know, the the all of the benchmarks and all of the like hype news that you see is always people just greenfield single prompting toys. You know, the difference between a single prompt greenfield toy and actual real production hardened brownfield work is so different uh the the pain of scaling into that is is huge and that's that's the thing that i'm really really passionate about trying to spread awareness of and and help teach it [56:41] Richard Webbe: Isn't that the philosophy of agile project management? Uh, [56:45] Matt: maybe [56:45] Richard Webbe: people before process and technology. So just get it built and get it out there, and then, as you say, uh double check it and look into it. [56:56] Matt: yeah or or you you could you could think of like inner reality where it's it's It's safe to constantly push domain and main is currently deployed in production. What does it look like to constantly push to main? You [57:08] Richard Webbe: Exactly. [57:08] Matt: know, what's what needs to be in place to be confident, confident, to put something like a ticket in one end of your uh situation, and then the other end it gets pushed to main, and you're confident that enough validation and checks have been put into place that that thing's gonna work. [57:25] Dave Pengelley: Yeah, I saw um I mentioned um the psychology of Developers post from um a few weeks ago uh from someone on LinkedIn called um Annie Bella, who we're uh gonna try and line up and get on the show soon. But she's talking about um at the moment uh the whole idea of not even testing afterwards, but before you actually run the compile, um, mathematical verification of code and algorithms, uh, and there's this whole sphere of work. [57:58] Matt: That's what linkers are. [58:00] Dave Pengelley: Um yeah, verification without inspection. And so she's been talking a lot about this lately. Um this sort of mathematical proofing arena of stuff. [58:12] Matt: Yep, [58:12] Dave Pengelley: But [58:12] Matt: 100%. [58:12] Dave Pengelley: uh [58:13] Matt: I agree with it. You know, any possible way that you can check, you know, that your code is correct to begin with. But then also checking the the the the ramifications of mutations. So can you can you Pump uh a whole bunch of pressure into one of your API endpoints and see what happens? Can you like basically DDoS your own deployment and see what happens? Can you recover once it breaks? What's the cost of a mistake? What's the cost of a break? You [58:43] Richard Webbe: You're [58:43] Matt: know? [58:43] Richard Webbe: trying to ensure against the law of unintended consequences, and that is one of the hardest things to do. [58:49] Matt: Yeah, and expect them. [58:52] Richard Webbe: So that's why we're talking about whether it's a gardener fixing up your backyard or you're building a a piece of software to a large piece of software or an agent. To work on your financial position, you don't know what it doesn't know, and you don't know what how many look actually we're at a disadvantage with AI because it will come up with millions more permutations than we can think of that would be mistakes. And that's [59:13] Matt: Yeah. [59:13] Richard Webbe: why the guardrails are so important. [59:15] Matt: It and it can also say, Yeah, I know what the problem is, it's this, you know, do you want me to fix it? Yeah, all right, go fix it. And then it does it, and then [59:22] Richard Webbe: Oops. [59:22] Matt: that wasn't the problem. [59:24] Richard Webbe: Exactly. [59:26] Matt: Yeah. [59:26] Richard Webbe: Oops. Yeah. [59:30] Dave Pengelley: Yes, always always more to do, more to learn. Um, and if you're listening to this and you're already on the train, well done, you are still early. Um, and look, but for many of these things, there's not necessarily right or wrong, it's all just you know degrees of separation and working out what fits your situation, whether it's yeah, but but like you like like Matt was saying, like get in and start. Don't think you've got to do a formal course in order to [59:56] Dave Pengelley: experiment with these things. Get in there, ask your Which I said to your friend uh Richard, I was like, if you want to be using AI to help run project management, ask your AI what are some ways I could use this to help this. But remember, your value is not in being able to know which AI tools are good for project management, your value is in project management. [01:00:17] Matt: And and also asking really good questions during that. Like, you know, think about different realities that might be possible and say, okay, you know, that might be true from the past. What's the current best way for project management? You know, what if we were to reinvent project management? What would that look like? Okay, [01:00:36] Richard Webbe: Yeah. [01:00:36] Matt: now like delete all of those processes. What's the simplest and most efficient way of doing project management? And yeah, just spend the time asking all of these crazy questions, like you're trying to just make things appear, make something appear out of reality that doesn't already exist and see where it takes you. [01:00:58] Richard Webbe: And don't limit yourself to The problem you're trying to [01:01:01] Dave Pengelley: Yeah. [01:01:01] Richard Webbe: solve. [01:01:02] Dave Pengelley: Challenge the orthodoxies, but also recognize that you don't need to reinvent every wheel either. Is [01:01:06] Richard Webbe: No. [01:01:06] Dave Pengelley: that [01:01:06] Matt: No, exactly. [01:01:08] Dave Pengelley: yeah. [01:01:09] Matt: The [01:01:09] Dave Pengelley: All right. [01:01:09] Matt: terminology I use for that is solved problem. Like I I I usually say let's not inherit dogmatic institutional knowledge, but [01:01:17] Richard Webbe: Yeah. [01:01:17] Matt: let's also make sure that this isn't already a solved problem. [01:01:21] Dave Pengelley: Like a wheel. The wheel's [01:01:23] Matt: Yeah. [01:01:23] Dave Pengelley: pretty good. [01:01:24] Matt: Yeah, yeah, they're efficient. [01:01:27] Dave Pengelley: Yeah, we wheelie, we has. All right. And on that terrible, terrible line. Uh thank you, boys. Thank you, gentlemen. Thank you to our viewers all around the world for [01:01:39] Richard Webbe: Yeah. [01:01:39] Dave Pengelley: uh jumping in. Uh [01:01:40] Richard Webbe: Oh, [01:01:41] Dave Pengelley: bunch [01:01:41] Richard Webbe: Doug [01:01:41] Dave Pengelley: of people [01:01:41] Richard Webbe: Amber [01:01:41] Dave Pengelley: on [01:01:41] Richard Webbe: and [01:01:41] Dave Pengelley: the [01:01:42] Richard Webbe: John. [01:01:43] Dave Pengelley: Amber, Joe, Phil, Doug, we had a whole bunch of people contribute today. So thank you everyone for jumping on the live chat, interacting with us. We love your comments. Uh we love comments after the fact as well. So when you're watching this on replay, make sure you drop a comment for us as well. We'll look to Talk about those on future episodes if you want to ask us any questions on that. So make sure you like, subscribe, comment. We are AI Operators Pod. Um, follow us on the new LinkedIn. Uh, there's an AI operators podcast page on LinkedIn. Please follow that. Once that gets big enough, we can live stream to that, not just to my personal profile. Uh, follow YouTube, keep that growing. Get us followed everywhere. Everywhere you like social media, look for AI Operators Pod. And on that note, thank you. We'll see you back here next week. Shooting the breeze, talking more AI business.