The Future of Agentic AI with Rory Richardson | AWS Humans In The Loop Podcast

0:00 AI isn't just a code generationtool. I mean, that's certainly a big piece of where we're seeing I get applied today. the most mature piece. I mean, do you know why it's the most mature? Enlighten me. there's no adjectives. In programing languages they're incredibly regular and consistent. So the application of AI, you get a lot better results than you do initially with natural language, because we we have adjectives and a lot of funny ways of saying things, right? But with programing languages, it forces a level of precision in order to get to the ones and zeros. So while developer tools are the most mature out of most of the workloads that I'm seeing, I think it's just a harbinger of what's to come. Welcome, everybody to the AWS Humans in the Loop podcast. I'm your host, Chris Shea. I'm a solutions architect at Amazon Web Services. And joining me today is Rory Richardson, Director of Agentic AI GTM at AWS. Thanks for having me, Chris. I always love to come and talk about what's new in technology.

1:03 Great to have you on the show. Tell me a little bit about yourself and how you came to where you are today. I would say my superpower is I'm a voracious learner and I grew up in Paris, Texas, and we we didn't have a lot of money, but my family always had a philosophy that everything was figureout-able. And this is pre-Internet, because I'm old. I'm old enough to have COBOL jokes. But with this attitude that everything is figure out-able, it just meant that there was never a door closed to me, because it just meant that I would rebuild the door somehow. and that was just how we were always taught. Because, you know, we're out in the country in the middle of nowhere, and there's no there's no system to save you. There's no expert. Everything you you have to figure everything out. And so I would say that has been the hallmark of my entire career. even at AWS I've had a weird career because I tend to change technology stacks almost every year.

2:04 because I like to work on stuff we haven't figured out quite yet. And then once we figure out, like the repeatable motion of how to talk to customers or what customers really want or how, to deploy faster, I get seduced by another technology and try to do something else. So I would say that's the hallmark of my career or what I've done here, is I tend to be enamored with what technology is, but also what technology could be. And there's always that tension between those two worlds. And that's where I like to. That's where I like to play the most. That's where the fun is. Yeah. So you've been with AWS 12 years? 13. 13years.

2:48 13 This month I accepted my offer to AWS on April Fool's Day. I don't know if the joke's on y'all or me, but it was on somebody. Okay. So we're almost down to the exact day. Yeah, that's that's exciting. yeah. So I typically like to ask, hey, what's like changed the most? But for you, I'm curious. Like, what hasn't changed? Culture, so, like, I think there was a sort of a shift in culture during Covid. and during our high growth periods because, you know, as soon as you bring a bunch of folks from the outside, it takes a while to sort of acclimate to, culture. But I, I tell all of my customers, you know, when they are asking, how do they prepare for the future? Or, you know, how are they incorporating agentic AI into their organizational behavioral change? And it's like, like Amazon's got a superpower in its culture. And if you you celebrate things like learn and be curious, that means you've got people that are pretty agile, right? They they're not they're not they're not waiting for someone to train them. And so the that notion of top down training doesn't really exist. Very like there's not a lot of that at Amazon because most people are self learners. So I would say the the thing that hasn't changed the most is really how much the leadership principles as touchpoints still continue to be AWS like secret sauce and being able to futureproof or attack ambiguous problems and have high degree of ownership for whatever we take on. That's awesome. And it's good to hear that that culture has really been able to persist throughout your time here, you know? Very good to see. I think it's definitely evolved. I mean, over the course of 13 years, our leadership principles have changed, right? Of course. the one that I missed the most, was vocally self-critical. I loved interviewing for that one because it made people so nervous. But at the same time, where we kind of decided on our leadership team made the decision on was that it's impossible to earn trust if you're not vocally self-critical. but I still like it as a leadership principle that I find all of the truth seekers on my team are rarely.. they're not trying to gloss over information, they're really like, this is what I know. This is what I don't know, and this is what I need to know. and that framework and that mental model, speeds conversations and creates innovation. you just you check your ego at the door and go as fast as you can. That's the best way to operate. You know, I think that's that's a really great way to describe it. So. As director of Agentic AI Go to Market. What's your day to day looking like these days? Honestly, if I take a vacation for ten days, I. It feels like you wake up in a coma.

5:35 Like, if. Because everything has changed so much, it's like this Rip Van Winkle moment. Because everything is changing so quickly. Our development life cycles for new services have been compressed so dramatically that I mean, off the top of my head, I can think of one service that was developed in ten days. So that means that it's pushing the limits of some of our mechanisms, like take PRFAQ's normally, you know, you have an idea, you write the press release and the FAQ and you go to your leadership and say, okay, I've got this idea. And they talk about it for a little bit and they're like, all right. We think this is a good idea. You get a team together and you start building it. But when you're able to go from a prototype into a working product and then into test and hardening it to production in such a short amount of time, I think one of the biggest challenges we have is getting on an executive calendar to get permission to do the thing you've already built. Wow, that's a good problem to have too. It's the future. I mean, it changes our relationship for development and code really fundamentally. Most of the executive conversations that I have always end up with in the realm of philosophy or sociology, you know, because how is this going to impact our kids or, you know, but don't you have to know the fundamentals? I am not a philosopher, nor a sociologist, but that's where the conversations happen, because you want a future proof. But in order to do that, you have to think about how humans are really interacting with technology in order to create systems and technology that work with humans. But a lot of a lot of people all over the world are wondering about what are the impacts to how we work when it's so easy to get at information, knowledge, wisdom and analysis. Like, my kids are growing up in a world where the power of the sum of human knowledge is in their pocket. So what's the point of memorizing anything? What's a different way to grow up? Like they learn faster, but they don't memorize much. Another thing that's changed.

7:52 especially it's true of my kids, but I'm seeing it really in work as well. Is the value of intrinsic learners versus extrinsic learners. And extrinsic learner gets feedback from a teacher. Is looking for a grade. Looking for approval, looking for an attaboy. And intrinsic learners just love to learn and they go after it. I think we are shifting dramatically towards intrinsic learning. more so than top down learning. What do you think the impact of that's going to be?

8:23 I see this is where it happens. And then I have to pontificate about what I think about the future. I'm a technologist, but I do think, how we've made decisions, has been compressed. You know, the we if you're able to, compress abstraction layers, you're really able to compress the entire life cycle of a decision or a project. And that's effectively what's happened. Because if you think about how technology is made, there's a lot of it requires a lot of expertise in bespoke disciplines, right. In order to get something out the door. What if it doesn't? What if you're going from ideation or intent of what's in your head to ones and zeros, and you're not passing through the abstraction filters of language, written language, spec, writing code, deploying the code test. And like, we have disrupted a very linear process. So if you're able to compress all the abstraction layers, you really are going from this idea in your head to ones and zeros.

9:33 That also means that this source of innovation for a lot of companies that I'm talking to, and that we're working with, isn't necessarily engineering, it's finance, marketing, customer support It's wherever you see the problem. You have an idea for the solution. You can start building it, prototyping it, and testing it yourself. You're not you're not waiting for permission. And that is very chaotic. But it's also a really fantastic opportunity. Well, yeah, I think what the entire philosophy of like move fast and break things, right? I think that's that's kind of how I try to try to operate. That's my that's my modus operandi. And I think it's served me fairly well so far. But yeah, you bring up the software development lifecycle, and I'd love to get your takeaway on what the future of that is going to look like with AI, because I think we're seeing a fundamental shift in the way code is being written now. Software engineers do their day to day work. Yeah, we we we definitely are. I think Swami Sivasubramanian talked about this in his keynote at re:Invent that, we had this project called Project Mantle, where we needed to rewrite big chunks of Bedrock, and what that team was able to achieve was so fundamentally different than anything else that we had developed, because they were estimating maybe 30 folks over a year to develop, and they were able to do it in about 76 days, 200,000 lines of code. and this wasn't this wasn't code bloat. Like, these were very senior developers, and it really pushed the envelope of how AI could be incorporated into the SDLC because certain things were extremely compressed. like, for instance, two years ago, if you needed to change direction of a project, you needed to rewrite the spec.

11:23 Usually that's the product manager rewriting the spec. And so you, you talk to them and you say, like, this is the direction that we need to take. The product manager would take sometimes a couple of weeks to turn around a new spec. That's it's like a couple of hours now. And so you have to embed the PM within that team because your changes are so quick. It also really identified the need for the humans to work together and to collaborate better. because right now, like, if you, if you ever vibe coded. Oh, absolutely. Okay. It's kind of a solo activity. Yeah.

11:57 Right. It's not like you have ten of your friends together and you have a LAN party, a vibe coding, right? It does sound like a great time, but no, I don't usually do that. Yeah. No vibe coding with the boys. Sounds like a new podcast idea, right? I like it, but, But the technology of agentic IDEs are great at collaborating with agents, but not necessarily collaborating with other humans. Like, for instance, if you're deviating from the spec and if you deviate like,no, no more than like 20%, maybe it's time to resync with your buddies to say, all right, is this still the goal? this how we want to solve this or not? and I think, over the next three months, we're going to get a lot better at incorporating humans within, you know, sort of a mixed team with super agents or frontier agents and humans. Instead of it just being a solo. Like, let's just get started. Activity. humans are always better when they build together because they bring in different perspectives and different knowledge bases. Right. Yeah. And that's why we've seen things like pair programing in the past. Right. and, and whole programing languages were established, you know, like object oriented programing so that you sort of divide and conquer, you know, each person, like, had their own object and then sort of came together. But now it doesn't work that way. And so even the fabric of some of the languages that we're using is going to have to change to accommodate a very different way of getting stuff done.

13:30 I was just talking to an SDE before coming over here, and he works on Agent Core, and he wrote a 500 word prompt. It was very detailed and gave it to Kiro. And in the morning it was done and it was about 80% done. Like he needed to put the human in the loop. And he thinks maybe it was another day. And I was like, how long would it have taken you just, you know, the overnight piece. And it was like probably two days, two days. And he it happened while he was sleeping that that's a what a wildly different way to code. And so a lot of the software development pipelines that we used to have are, are pretty linear. I saw a meme, I don't know, last week that cracked me up and it said, epics aren't epic. Yeah, there's nothing epic about an epic. It's, it's a bit of a misnomer. But think about it. Think about, like, how much epics have been a part of our world for so long. But if you're able to compress the need or what a taking system was actually driving within the SDLC, then? Well, we have to think about this in a non-linear workflow. And in order to be able to optimize how people are working together. Absolutely. And, you know, I think again on the epic piece, too, like, I just think it wouldn't have taken off as quickly if it were called a miserable. But here we are, right? I think they oversold it. Yes. It was a little more than what they promised, I think. But I think you bring up a great point too, in that AI isn't just a code generation tool. I mean, that's certainly a big piece of where we're seeing AI get applied today, but. It's the most mature piece. I mean, do you know why it's the most mature? Enlighten me. there's no adjectives, in programing languages. They're incredibly regular and consistent. So the application of AI, you get a lot better results than you do with initially with natural language, because we, we have adjectives and a lot of funny ways of saying things,right? but with programing languages, it forces a level of precision in order to get to the ones and zeros. So while developer tools are the most mature out of most of the workloads that I'm seeing on generative AI or agentic AI that customers are building, I think it's just a harbinger of what's to come. today we launched DevOps Agent. I mean, it's so now now we have whole frontier agents or autonomous agents that are helping us with site resiliency. And it's amazing, like, site resiliency is a pain in the butt because you know that that means you're wearing a pager, that that's the job of the SRE. Yeah. Getting paged at 3 a.m. What if you had an SRE that never slept? I think I've known a couple of those in my day, but I don't think they were supposed to not sleep. yeah. And it's not healthy, but yeah, I mean, like, if you had something that could respond especially to low level interruptions and remediate them by the time you wake up. Wow. And, like, how cool would that be? Yeah. You know, even then, it's like, by the time you get paged and, you know, out of bed and blurry eyed, you're logging on. You're like, you can at least have something somewhat close to a root cause analysis tool. So, I mean, I've definitely been woken up at like 2 or 3 a.m. and my responses to the ticket were not, ooh, Shakespeare. Not really English, but, you know. Because you're angry and you're like, shut it all down. Yeah. Stop the noises Make the beeping stop. Yeah. I think the next three months we're going to see the rise of, autonomous agents, especially within the build phase.

17:17 and the operate phase. So there are a lot of tasks that we can assign, frontier agents now that we hadn't been able to in the past. for instance, the security agent is able to do pen testing, and pen testing is one of those things. don't actually have to sit there and watch in order for it to, conclude, like how vulnerable your site is. Whatever. or, I've already mentioned, you know, going into YOLO mode, that's what we affectionately call like just sending, instructions or sending the prompt and saying, go ahead and complete the code. That's YOLO mode, right? Because you're not you're not checking every change, but you have to check all the changes once everything is complete. it's like a one shot, but you're verifying everything that was done at the end of the road. We're still not past human in the loop by any stretch of the imagination. I get asked a lot, from our customers. It's like, how do you know? How do you know that what AI is suggesting is the right thing? And I'm like, that's the developer. I can't fire AI, but I can fire a developer. There is still a level of responsibility. But sometimes AI is really misunderstood or misused, like especially if you don't work with it a lot. Some people think AI is deterministic or that you should trust it. that's not exactly right. You know,it, it it gets you 80% there in 99% of the cases that I work with. And which means that you're like, okay, that's not quite right, but it's only 20% of the work in order to change it. And so it's still a huge productivity boost. But what it isn't great for is if you treat it like a SQL statement. Right.

19:07 If you treat it like a SQL statement and this is what all the data I got from the database, you don't go by and you don't need to verify that. Right. Because it's it's very computational. Yeah. Yeah. When you are running a summarization, it may take the wrong path in the summary. And so the whole thing is wrong. And so you've got to be you still have to have a human in the loop. So like a lot of times when people treat it that way, they want to ship it their their AI responses immediately to our most senior leadership And I'm like, oh, please don't do that.

19:40 It's not that I don't love AI, I do, I really do. But if you're using it as a deterministic output output like a SQL statement, you're going to it's going to mess you up. It's going to it's going to give you some weird results. I've done enough workshops where you'll be, let's say, like typing a specific prompt in the Kiro, and it generates code that isn't going to look picture perfect, exactly like the workshop instructions. People will be like, hey, like my code is different. It's like, yes, that's the point. It's it's non-deterministic. not going to get an identical response with every prompt. Yeah. It's it's like it's like asking your friend that has a lot of cough medicine going through their system. This is like sometimes is it. Sometimes it's different. Yeah. You know. Sometimes it does different things, you know. But much like if you ask two different developers to solve for a problem, right, you might. At different points of time, different times of day, It it's rather organic that way. that's why I over even now I, you know, some folks on my team started shipping the AI slop. It drove me insane. I was like, are you trying to make yourself redundant? I mean, I can write a prompt. That's not your job.

20:45 Your job is to use AI to do your job faster. But you absolutely have to verify this and contribute to it and polish it up before you ship it on, because that's your value within anything. that's the most humanistic side of what we do. We can't. We're like, because AI is not it's not a machine. It's not computational. It's not deterministic. Don't use it like that. Just use it as an accelerant and a democratizer, and it's fantastic. Use it as something deterministic, you're in trouble. Yeah, 100%. And viewing it a little bit more on like specific use cases that you've seen that have really resonated with customers. Yeah. even just in like, your own life, I know. Like kind of showing. Showing your children kind of the way that the AI movement is really changing. The way that people are learning and doing a lot of different things in the space.

21:38 Like what's really resonated with you? one thing that's changed rather dramatically for me. And mind you, I have two kids, a 14 year old and a 16 year old. My 16 year old doesn't think in apps, like I'm old. We went through the 2000 when we kept saying, oh, I wish there was an app for that. Yeah. My 16 year old does not. He wishes there was an MCP server so that his agent could talk to this, and he he was out of the process entirely. That's how this next generation thinks. And we're slowly catching up. Like internally Quick allows you to create your own productivity agents that are customized to how you think and how you operate. You know, personally, and I'm seeing the rise of citizen agents. even within our organization. Citizen agents. Can you walk me through that I just made it up. So let's not get too serious and stuck on this and the things I just make up. But what's happened is, like, let's say my Sales Ops leader creates her own agent. And by the way, she anthropomorphizes it. I think her agent's name, Sarah, gives it a gender, you know, because, you know, Sarah has a personality because it's trained on everything that the sales leader actually cares about and prioritizes, which I think is hilarious because, you know, back in my day, 13 years ago, we used to say, you know, servers aren't pets, they're cattle. You know, because we used to name servers. Look, this new name, this one's named Gandalf. And. Right. But then when the cloud and being able to flip into a new server and have failover so quickly, we said no, no, no servers or cattle. Yeah. This is now, instance x907348. Like, it's a little more anonymous. Agents are moving the other direction. They're moving from cattle to pets. Yeah. So now it's not just a, you know, agent XYZ, it's Gary. It's Gary. Yeah. and a lot of that is driven by a few things, primarily, how memory is allocated to the agent, how the session state is stored so that it can maintain that personalization. So it's allowing more hyper personalization than we ever thought possible, because you're able to hold that information from one session to the next versus reloading all of that information, which has a latency and cost tax to it. Whereas having a stateful runtime means that you can pause it and pick it back up later. And I think that is a huge difference, in personalization. I was thinking about the shopping experience on Amazon because it's still a cascading style sheet, right?

24:25 And there's a lot of recommendations based on my purchase history within Amazon. But what if it wasn't?What? Whatif it was a concierge? What if it was like talking to your best friend? And your best friend knows everything about you? You know, from high school onwards they know everything. The good, the bad and the ugly. You know all of your other friends. They remember the jorts phase. Everything. They know your purchase history and a bunch of third party sites. Then why would you have a cascading style sheet then? Like instead, it would just be like this. I would just go up to you and say, here's the deal. I'm gonna need a dishwasher. can you just just TLDR this for me. And the dishwasher shows up. That's happening. Like, I like, the way we interact with information and technology through these static pages, I think is going to fundamentally shift over the next six months. It's a very exciting time, but I think it's also very terrifying. Yes. I read a lot of dystopian science fiction as a kid. It never ended up well for the humans. Oh, goodness. What authors? Oh, look. I actually wrote down the canon of science fiction that all my generation of nerds read. And it's like, you know, Ender's Game and Hitchhiker's Guide to the Galaxy. Like these books from Gibson, Stephenson shaped this whole technical generation in how we think about the internet in the future. I and I look at my kids a lot and go, wait, what are you reading? You know what? Because I kind of want to know what's happening next. I want to, you know, read the next chapter. But I do think it's fascinating that we have a generation, that there's about 30 books that we've all read, and we all get the jokes from, you know, like, using the word grok because it's Heinlein. these are, these are within this population.

26:20 And, you know, when I look at somebody in their 20s and I think about what were their cultural touchpoints in order to direct what they're going to be creating. I find it absolutely fascinating. Like, there's going to be like a huge K-pop influence, apparently in the next generation of technology. I'm very curious to see how that's going to manifest itself. But at the same time, too, it's like despite maybe like a generational gap, you do have AI tooling there to serve almost as like a Babelfish of sorts, right? That's able to kind of connect those dots. I think we have that product. It's called Babelfish. Yes. Oh actually I think that's an internal project name.

26:59 But yeah, having a universal translator between things. Like here's another weird example of how AI has changed the game. I recently, was on a business trip to China and I never had an interpreter. So we were very lost in this oriented the entire time I imagine? No I used AI. Everybody was using AI like everybody, like the. This is where I learned that, that developers are just as sarcastic in China as anywhere else in the world. But I never knew that because I don't speak Chinese until this last trip. And I got to see the direct translation because usually the interpreters sanitize. Yes, some of the spicier things developers say, but this time being able to see it communicate directly with them. I mean, I just started laughing because I was like, I just want you all know y'all are the same all over the world. Yeah, no, the the foul mouthed PM is definitely a it's a universal trope. It was, you know, because they're vicious, right? And it's like, why would you even develop it this way? Don't you see how you're gonna run into problems in this? You know, like it? Developers are developers. We have a certain mentality to uphold.

28:09 It's great to see that that's a universal thing across the globe. Yeah, yeah, it's, It's everywhere. Yep, yep. It's it's a persistent fact of life. But I can tell you the next time I go to a place where I really don't speak the language, like I don't speak Portuguese and I don't speak Chinese at all. I'm definitely going with translation hardware of some sort because it's it's pretty fantastic. Is this, like, real time? Like, you have it in, like, earbuds? Like. Conversationally translate one of the meetings did because the SA had, Xiaomi, glasses. You know, they're they're like meta glasses. Oh, interesting. and she was getting audio. I thought about getting them, and I was like, I can't get a Chinese translation. Oh, totally. I can't do that. But, when I look at, like, the next evolution of, like, what these glasses are going to be able to do. some of them are giving you text that you're reading. but some of them are giving you an audio translation, and some languages are easier to translate to and from than, you know, others, like Portuguese and Spanish is not that hard for me because it's constructedsimilarly.

29:15 Chinese is really hard because it's highly contextual. And so what was happening in the translation and what happens in the simultaneous translation is you almost have to wait until the full sentence is complete to know which word it was, because the words are so similar. so that was also fascinating to try to simulcast, because I was getting a really bad headache because you're like, something about your mother. No, it's about this CI/CD. Nope. It's about words change. It leaves you on the edge of your seat until the end of the sentence. You're like, oh, okay. All right. I thought you were insulting me, but no, you just needed to reboot that EC2 instance. Makes sense. It it is fascinating. I mean, but other places are still using, you know, interpreters. But I was kind of I went a week in China without, really having a good working knowledge of Mandarin or Cantonese. By the end of it, though, I could curse pretty well. Sounds like a successful journey. Like I said, it's talking to developers. Yeah, absolutely. Absolutely. So I do want to talk a little bit to you about agentic AI. Yeah. And I think that's it's kind of a trend that we're pivoting towards. But What services have you been seeing customers use that have really resonated in the agentic AI space today?

30:34 so you've got the, applying AI to the SDLC world, but then you've got. How to develop, deploying and run agents world, and those worlds are going to converge. But right now they're a little separate, because the things that you need to deploy and run agents are different than what you need for applications. So I found, I think Agent Core is brilliant. because in the true Amazonian fashion, it removes the undifferentiated heavy lifting of deploying and creating and maintaining agents, because there's a lot of elements that you need for developing an agent that are primitives within Agent Core. You know, like, where are you going to run it? Okay. Your going to put it in the runtime? How much memory are you going to give it? Okay, great. That's memory. How are you going to observe it. How are you maintaining identity. how are you creating policies so that the agent behaves within the parameters that you set for it? Not necessarily the guardrails that you put at the model level. So I find that Agent Core resonates with a lot of customers. And the trickiest part in these conversations is to figure out how sophisticated are they really, not how sophisticated they say they are. Yeah, right.

31:51 Everyone likes to I think, over represent where they are. Yeah. You. So you met developers. 1 or 2 in my day. Yeah, yeah. So, I started trying to qualify, like, their level of sophistication. like, do your agents talk to each other or do your agents talk to humans? Right. Who? Who is your agent, actually for? Have you built a super agent or, you know, an agent harness for your agents? It also gives you an idea of how many agents would like or within their environment, and then how they're coordinating and doing more complex things. have you thought about a genetic payments yet? Or how do you what is, how do you discover agents that other people create? These are all things that denote a certain aspect of sophistication because we're finding, like, for instance, startups will build one big agent that's good at a lot of things, whereas more mature companies want to build smaller agents. it's more similar to microservices. to build flexibility and agility within their strategy. That that seems to be a reoccurring theme.

33:08 But I mean, startups, of course, move faster, and follow less rules. So they're really fun to talk to as far as, like pushing the boundaries, for instance, of what they can do with memory. Whereas, for enterprises, they still ask questions about, well, how do I institute an organizational behavioral change? The thing is, like we're finding that top down motion doesn't work anymore. And so it's about finding and building tiger teams of people that are really savvy and really adaptable and letting him loose on a problem and then letting them brag about it because developers only listen to developers. That's just a thing. and it's true for AWS internally. Our most popular broadcast, which is our internal YouTube, is about like that project that I mentioned on Bedrock. because they got to talk about how great and how fast and how well thought out. You know, their creation was in that works. But if you're looking for we're going to hire a consulting company to come in and teach everybody how to use AI. No. Yeah. Snap your fingers. Everybody knows AI now. Problem solved. It doesn't usually pan out like that, does it? Yeah. I'm even finding that prompt engineering is devolving as a skill set, because the models have gotten so much better at clarifying questions that you don't really have to write the perfect prompt the first time. Yeah. You know, you can have an iterative conversation even before it hits the LLM, you know, like even at the MCP server level.

34:46 and that that naturalness and how people talk, means that you don't really have to do prompt engineering as a discipline It's a really good point. You bring up, I've noticed myself, is if you look at prompt guides back in, even a year ago, it was a long, verbose, like, we're talking just massive text of context, instructions, examples. And now it feels like you don't really need to have as much of that know when you prompt. I would agree. I personally always thought prompt engineering was an anti-pattern, though that was a discrepancy in the how sophisticated the technology was.

35:24 Not necessarily a skill set. but that said, I would say that communication, expertise is still king, which I don't know. Like, old school developers kind of excelled at not communicating with other developers, right? You know, they they did their thing and they have their mind castle with what they wanted to create. And it's beautiful and it's well architected, but it's not, they're not they weren't as collaborative as this next generation of developers where, you know, they vibe code and communicate differently. Yeah. Makes a ton of sense. And how are we seeing In terms of how AI is changing the software development lifecycle. We have something called AI DLC. Yeah, here. Could you walk me through what that is? Now it's helping customers. The application of AI into the software development lifecycle. Right. So from where you're building an energetic IDE like Kiro, to how you're pushing that through to get all the way to a full code commit. and applying AI at each one of those stages, like the, the customer I was just talking to before I came over here, we were talking about how to change a linear pipeline of software development. Normally, you know, like I said, that's JIRA is the most popular, way, but it's also very linear. And so the compression of those stages, that's AI DLC. is how do we use AI for everything that is mucky within the software development lifecycle? But another term that we're seeing in the AI DLC that I don't think we talk enough about is how we are retiring tech debt at the same time. So, everybody has tech debt. I don't care if you're a startup. That's all on go. You still have tech debt. Everybody has tech debt. So what if instead of thinking about tech debt as, oh, I've got to remodel my house, I've got to find a contractor, it's going to go over budget, there's risk, blah, blah. What if it was just a handyman that you hire every two weeks and so you don't incur a mountain of tech debt? That's the philosophy that some of our customers are adopting by applying transform as part of the software development lifecycle as well. So let's say you needed a new feature request, and you're looking at the old code and where to put the feature in or the connection into the application. Cool. Well, while you're there, why don't you go ahead and run transform and update the code base? So instead of doing a massive jump from, well, in Amazon's case, and we're going from like Java seven, then you're always upgrading to the latest version of the code family. That's a massive shift in how to think about tech debt, because I can imagine a world not too far away from now where we don't have a mountain of tech debt. Not really, because we've been chipping away from it without it being like a major project, but more like, this is just what we do, right? Why? Why would you why would you mess around with integrating into an older, code edition when it's so much easier? You know, with the current one, it's only going to add like a day onto your project timeline. That's a massive mental shift from how we've been thinking about tech, debt and old technology until now. Absolutely. It's like we'll soon see almost the inverse of tech debt.

38:52 We'll start being able to build tech wealth in a way. But what are some of the decisions that you have to make when you're going through this process? Well, from a technology side, I think there are six sort of big decisions you have to make. And then there's a bunch of small decisions within each category. Most people are going to get started in IDE or CLI, right? You got to make a decision. Are you going to use like an agentic IDE like Kiro or are you going to do something else? What is that? Where are you going to get started. Next you're probably going to make a framework decision. in AWS we have strands, but sometimes you're going to use linting or crewAI. Another key decision. next you're going to need to put it somewhere. and so then you're going to make a decision of like, am I going to use an, a generic platform like Agent Core and take advantage of this primitives? Or might I do it myself? Or maybe there's another platform. next you're going to choose your inference. or model, obviously, we have Bedrock, which hosts a number of different models, but it's a key consideration. Generally, you can flip between the different models when you're in Bedrock. So it's a more flexible decision now than it used to be. next you'll make a decision about protocol, because, I mean, you know, an agent that doesn't talk to other agents is so 2025.

40:08 and, you know, it could be MCP. It could be A-to-A, maybe even ACP, depending on the type of agent that you're making. And the last decision, it's the most important one. That'sdata. So data is the heart and soul of any organization or any agent really. Because that's how you're training it on your data. So whether your data is in something like S3 or RDS or Redshift, you're probably going to use something to unlock it, like S3 Vectors or MCP servers for RDS or, Redshift, in order to connect that back and train the model and then build the agent on top of and using that model in order to get more bespoke, more specialized and more curated responses. I've always hated store procedures because I feel like they're a barnacle on your soul. I mean, they're the definition of tech debt. but we still have them. You did replace them with Lambda. I saw you have a patent in place for it, so. I do, but that's old school, right? That's not using AI. Now, we can actually use AI to move that business logic out of the database layer and into the application tier. So you've been trying to slay this dragon for years?

41:20 Yes. I really hate it that much. It's an inspirational level of spite, I love that. It really is because I just found that you were locked into one query pattern. You know, with stored procedures. Like, I like relational databases, but relational databases are only great if you know every question you're ever going to ask. If the information first and you can build the perfect relational database. Nobody nows. Nobody works that way. I don't know the next ten questions we need to ask of information. So being able to create more flexible information architecture means that you have future proofed the treasure trove that you have within your organization. But if you think you're so arrogant and so smart that you know how to optimize based on this pattern, it's going to leave you rather inflexible to deal with change in innovation because you don't know what's going to happen next. and so I really like decoupled and more flexible systems, even if they're like slightly slower, which now they're not really, but they give you so much more flexibility in helping you think ahead for the future and avoid really costly mistakes that lock you into a pattern. So yeah, I'm really glad that we no longer write stored procedures. It just in general no net new stored procedures. But I also think it it's what really sort of locks people into certain commercial databases and ways of thinking about their information And if you can, if you can morph those out, you you can unlock, like I said, like huge swaths of information that is pretty buttoned down when you have store procedures. can hear a million DBA sighing relief across the world right now. They didn't used to. I mean, seriously, when I started working on this, like, I don't know, ten years ago, I was on non-relational databases and it was a pretty radical idea then to stop all stored procedures.

43:23 But they came around and said, oh yeah, we don't have to do that anymore. Collectively had Stockholm Syndrome. Yeah, right. I mean, some people have an entire career on stored procedure optimization, and I'm like, Eek! Eek, doesn't sum it up quite right. That's just different noises. I know I'm probably going to get a ton of hate comments about it, but I don't, I really don't like stored procedures. I don't like networking either, to be honest. You know, I think you're not alone in that sentiment. I think there's dozens of us out there, but, Not the folks with Cisco certifications. Yikes, yeah. Another weird thing that I've seen as a trend is, is just code replacement versus debugging. And this hurts the soul of a lot of people when I say this, but it's kind of true. So debugging is a whole skill in of itself, right? and there's some people that are amazing at it and some people that really struggle with it, but it's the cornerstone of writing great applications is debugging. What if you identified where the bug was, but you just replace that wad of code? See what I mean? Yeah. Like you felt that.

44:33 Yeah. Yeah. Just fixing what's in place. We're just going to throw the entire, replacing the entire chunk of code. Like, I mean, I grew up working on cars with my dad. I know how to rebuild a carburetor. That skill set is useless now. Because what happens for your car, right? You take it to a mechanic, you plug it into a diagnostic tool. It tells you what part is broken. You replace the part. You don't rebuild carburetors anymore because there's no ROI into that process.

44:58 What if code was like that? That is another trend I'm seeing and I can't believe like I'm saying it, but yeah, it's sometimes faster to just replace it with something that you know that works versus figuring out why the other one didn't. And if you were to tell a developer that ten years ago, they'd they'd have a conniption Dude, and when I tell my brother that and who's older than me, he dies a little inside. Just like, how is this possible?It's. It's insane to see the clip at which things are changing. And I think to, like, are there any common patterns you're seeing with customers who are really succeeding with adopting AI in their development lifecycle? Culture. I mean, seriously, it's it's not it's not whether they're astartup.

45:40 It's not whether they're anenterprise, it's whether or not, they have a, a culture that embraces play and experimenting. And failure. Like, that's a very forgiving and advanced sort of organizational culture. And if they have that as like the cornerstone, then they're, they're moving really fast. It's usually a high trust organization. They, you know, they trust each other. They're not gunning for each other. so that emotional maturity of humans actually plays a huge part in how you adopt and experiment with technology. It takes a lot of humility in a way, to be able to embrace AI effectively. I feel. I would agree. I'm, I would say, you know, one of the ways that our, even our internal organization has been adopting AI is posting to each other about different stuff that we build because we are so wonderfully competitive. They're like, this guy doesn't even know the difference between Java and Python. his app can't be better than mine, or his agent can't be better than I. And so then we start playing with it. And I find that there was this early trend where our most senior people, you know, maybe they're older, maybe they're stuck in their ways or whatever,more, more reluctant to use these tools than, you know, somebody who's early career. And I think that is finally, breaking up into like people that are really curious and, want to test the upper limits of the technology come more so than early career versus later career. yeah, there was some of that that, if you're really good at what you do, you can do it faster than AI. And that's a true thing. I can do some things way faster than AI. But then I started experimenting with doing some stuff that I was like, I know I can do this faster, but I just, I only have so many hours in the day. I'm gonna I'm gonna outsource and see what I can get, to see how far I can get. And it's just easier, you know, editing than starting from a blank page. For a lot of this stuff, and that includes analysis, it includes marketing campaigns. It can be even competitive analysis. You know, when I'm looking left and right and I use AI to get me started, but man, do I not ship it. It's crazy how quickly you can get 80% of the way done with something. Just know it's That's the thing is, people will forget about that last 20.

48:16 That still requires, you know, some some scrutiny. Yes. And, what's interesting to you is I think with AI, you can you can build something so quickly now that projects, small tasks that used to not even really be worth the time to write a Python script to do, it's able to clean up a lot of that miscellaneous work that would kind of just sit around and not get done in the first place. Which is insane to think about. No, it's totally true. I mean, I was talking to one of our customers last week that what he wanted most in life was to not update angler anymore. Angular is just a booger bear to update. That's the technical term.

48:57 Booger bear. I like So he calculated it was taking them like 1600 engineering hours to just upgrade angular. Wow. He's got it down to 200 now. Wow. That's just a really clean ROI. And it's not a lot of transformation executions. Like, he didn't even flip in any of my dashboards that I look at. But he was so passionate and so grateful because that was real hours that he got back from his engineering team to go do other stuff. it's stuff like that that I really enjoy, like how that's changed. But I also think, like the syntax of these individual languages are getting dusty quick. Yeah. I mean, I've been trying to teach my kid to code since he was four. I can't get that time back because by the time he's the job market. What's the role of the syntax of Python? It doesn't matter. I mean, I'm not saying computer science is dead. I'm not saying logic or frameworks or any of that is dead. I'm just saying, like the syntax of Python. There's no point. It's going to get abstracted away. You know. Nobody's going to be able to whiteboard code. Remember, that was an interview thing. It was like, hey, can you write this Whiteboarding pseudocode and all that? Yeah. Well, I'm getting I'm getting flashbacks, but it's it's insane. It's gone. You don't need to do it anymore. I mean, like I said, the sum of all human knowledge is in their back pocket. They would just, you know, look at their phone for a little bit. Go. Yep. Shipped it. What repository? Great. Thanks. Yeah. Just merge. We're good.

50:32 Yeah. Hang on. What? It's. It's a breakneck pace. I had a customer say. Oh, well, agents are the new microservice, and it was like, oh. Oh. They are. Oh! Oh, that makes MCPs, MCP servers the new API Gateway. Oh. Oh, that's scary, because, I mean, MCP servers aren't as hardened as an API Gateway. And so it's like, we can't really work on those. Yeah. We need some structure to this quick. Yes. I mean, MCP servers are really flexible and it's great, but you still have to harden them.

51:06 and a lot of them, you can still sneak behind and get to the underlying. They're just not as, fortified as a good API Gateway. Right? That has to change, man. Yeah. And it's like we're still very much in the in the early days, which is crazy to think because so much has changed. But the fact that we're still really only in the first act of this play is like outrageous to think about, because so many I think it just underscores the importance of really verifying the work that AI is doing, having that human in the loop. Don't just YOLO mode, you know, dangerously skip permissions. Send it right. You need to have those checks and balances in place. Yeah. Yeah, definitely. Especially when you've got non-deterministic results. and it's so easy to, I don't know, tweak it to get something just not quite right. And then you're, you can go off in the wrong direction just really, really quickly. But the other trend that we've seen is the need for, you know, faster human interaction. and I mentioned this before with Project Mental. But I'll say it again, like being able to figure out when you've deviated and then sync it back up with the other humans or using AI to facilitate deeper human relationships.

52:25 Yeah, that's going to happen. And I honestly was very surprised by that trend because I don't know, a lot of technologists are a little, introverted and isolationist by nature. And so this extroverted we've got to collaborate. Better trend than I'm seeing now. Kind of surprised me a little. just because the the previous generation how we related to technology now, we were relating to it in an entirely different way. Yeah. It's it's interesting to think. Right. Like, I think the, the stereotype of the, the basement dweller who, who's solo, you know, operates in their own domain. I think that's, you know, going to I think, you know, we have to change. Yeah. The, the like the foundation of Linux basically. Yes, exactly. Yeah. No insult to, to Linus But no, I mean but Linus didn't create all of the There were a lot of people in their basement correcting things and making it better. Of course. Because of the nature of, you know, open source. And what is the future of open source when you have so many people that are going to be able to contribute in isolation, or how do how is open source going to evolve with and how people are going to collaborate in these tools. I don't know yet. I think it's fascinating.

53:40 Yeah, no, it's exciting to see what the future holds. Scary but exciting. Scary but exciting So do you have any advice for, be it a hobbyist, developer or an SRE in an enterprise organization or business leaders looking to make a decision on AI? Do you have any advice on how they can best get started with agentic AI? I still think a build versus buy is the most important decision over the next 6 to 12 months because, the advancements are happening so quickly and so cheaply. Like, I'm of the opinion that you don't have to learn how to rebuild carburetors. Not really. so using these managed services for what they're really good for and as, and moving as fast as you can go is great. I do think you, right now, especially some of our verification methods that are not humans, aren't as good as they need to be in order to not have a human in the loop. So there we're in this transition point where we're just not at the point where I would say, I trust any of these tools 100%. Or maybe I'm just a cynic, but I keep thinking about how to raise my kids, you know, because I've been having this internal debate in my head whether or not I'm in this in my 16 year old college. I'm not kidding. Like, I've been thinking about this way too much. Like, what is the ROI of top down learning? When Harvard posted all of their classes online very affordably, and we're talking about a generation that went through Covid and online learning, they are intrinsic learners. They're self-motivated learners. If you figure out what goal that they want, they can work backwards from it. And I'm looking at my kids going, I don't know if top down education is where it's at for them. I do like how universities, you know, sort of push you to think differently or more deeply than you can on your own I think that push is necessary. I'm just not entirely sure it's a classroom anymore. it's a very radical thing for me to think about, but I'm also figuring out ways to teach my kid intrinsically. Like, I give my my eldest and unlimited tool budget, unlimited. You can build he can buy any tool that he wants. And so I mean, at 16, he's rebuilt the engine of a 2012 Subaru mainly off of YouTube. I'm not in the garage with him. He's doing it all himself. That's a that's a different generation than not than my generation where we had, you know, books and tests and university. Yeah.

56:13 Everyone can learn almost anything they want just through the power of AI and the internet. Now, you don't necessarily need that that structured university course format. No, I mean, but it the only thing that I think is really missing out of that way of learning is friction. You've got to have enough friction with really smart people to be smarter yourself. I mean, I don't know. I've always strived to be the dumbest person in the room, because then I'm in the right room and I have plenty of something to learn from everybody in that room. Right? and to be able to do that, you know, university opened those doors for me to get into the room to be the dumbest person. And so I'm looking at this next generation. It's like, how how do we open the doors for them?

56:55 And then in order to be able to challenge them to be in those rooms and get enough adversity, to develop grit and hustle? I don't know yet, but I'm still figuring out. Yeah. Exciting times. Yeah. For sure. It's been a pleasure having you on the show, Rory. Your insight on AI is just absolutely great. You're one of my favorite personalities in the AWS sphere, so thank you for taking the time of day to join me on the show. thank you for having me. I don't get to do this stuff enough because it's kind of outside my day job, but I really do enjoy it. I absolutely would love to have you on the show again sometime, but thank you very much, Rory. Thank you.