Latest / AI-First Engineering Podcast / Will AI Agents Redefine the Software Development Lifecycle?
Transcript
- speaker-0: Welcome to AI First Engineering, the podcast. speaker-1: Where engineering and technology leaders share how AI is transforming software delivery, quality, and the future of engineering. speaker-0: of engineering. Let's get into it. Welcome back to AI First Engineering, the show all about how AI is actually changing the way software gets built. My name is John Robinson. I'm your host from Test Grid. And my guest today has spent more than fifteen years building software and data platforms and a lot of time mentoring the next wave of engineers that are coming up behind him. Piyush Tawari, thank you for joining me today. Tell us about yourself. Who are you? What are you doing? Why are we talking? speaker-1: Thanks, thanks John for the invite. And ⁓ so I'm Pyush. I've been working in software industry for close to fifteen years. ⁓ throughout my career, I have been through different stages of software development, data infrastructure, testing. And ⁓ right now I'm an engineering leader with the e commerce company, ⁓ leading their platform infrastructure. ⁓ So yeah, why we are here is I think where we have talked a bit about is like how the AI is changing the landscape of software development, how the evolution is working and how we are looking at the next generation of things, how the market will move. So it is a exciting time. speaker-0: Yeah, I think that next generation's the one that I'm most interested in. But before we kind of get there, I wanna ask you a question. You you you've used both of the systems, both co pilots and and agents. What what's the actual difference there? Like give me the the beginner's version of of why one is different than the other. speaker-1: So I think what I will do is I've been seeing this ⁓ both the versions in real life in the last two years. And how the things have moved is when the AI boom started, ⁓ it was more of a code assist. So let me pull it an example, and I think that way we can talk more about it. Is earlier when the code assist or co pilots were coming in. There would be a story or a ticket which will land on your agile port or ⁓ your input, how will you take the input? And then you ⁓ engineer on the team will go work with the co-pilot, write the code, and then take it all the way through the deployment ⁓ and op sh operationalizing it. So at that point of time it was a function writing better code. But how the AI agents are evolving is now look at the same scenario. What is happening on my team is. As soon as a news story hits the Jira board, the agents pick it up, run it end to end, and tag any active engineer on the team for a review. So how it has moved is ⁓ moving you're gone from basically an actor with codes to being a director on the on the show with the agents. speaker-0: Okay. So that that opens up an interesting question, right? Where if those functions that you're describing are are kind of how people have historically learned good software development practices. you know, I ⁓ I I think you even wrote about this, you know, using clawed code with Opus can can replace a junior engineer. What does that look like though in say five to seven years when we're looking for that next wave of what I would call senior engineers or even even mid tier, right? What is that? Where are those skills coming from now then? speaker-1: I I think you hit the nail on the question. I wish I had like Yeah no, I I think that's a good question to talk or I think good topic to talk about. ⁓ and I wish I had a very good answer for it and I had been thinking about it a lot. So I think like to just set some context, how did we learn like who are in software industry? You do the grunt work. speaker-0: I like to ask the hard questions here. speaker-1: ⁓ that 2 a.m. in the night when something broke and you are under pressure to fix the issues, you learn that you learn by doing it for by like by failures. You learn through that or learning from your senior engineers. So I think that is how that is where the shift into the the shift into the industry is happening. Like with the AI agents coming into picture, it will reduce that skill set or it that how we train our engineers till now will change. So I think where the skill the skills which which will become more relevant, right? We need to take a step back and think about what skills are becoming more relevant with the new AI agents. Is the domain knowledge or how I call it where the skeletons are hidden. speaker-0: Right. speaker-1: ⁓ so nobody knows why we had like some repo standing up and just running one tandem job. So I think speaker-0: Don't don't turn that job off because everything breaks if you do. speaker-1: Or th why there has been basically some egg some sheet with some ⁓ essays or some information. So I think that is where the things the judgment factor or anybody who is ⁓ who is ⁓ next six years looking into it, first don't worry about the things which agents can execute. So writing that super awesome code. ⁓ That is you need to understand the code, but that is not the skills which you need to build on. What you need to build on is the domain knowledge and the judgment factor. ⁓ so now agent can write a very awesome code, but it won't understand that ⁓ what impact it would have on your consumers. So I think that judgment skill you have to build. And the challenge is how you build it is through that two ⁓ grant works or like Getting reviews from your seniors, ⁓ working with them, right, talking to them day in, day out. So Yeah speaker-0: Yeah. Yeah, I I was just gonna say, I mean, one of the things that that's interesting about that is you're you're talking about the learning at the two AM failures, right? The things that's how we often learn in humans, right? We learn by screwing up. ⁓ you nobody learns and does it perfectly the first time. ⁓ the people that do better are the people that learn from the mistakes. What how do you build that skill when AI is is by all accounts incredibly confident that what it's done is correct. And that is not necessarily always true. I mean, how do you A balance that but also B, how do you recognize when it's done something incorrectly but super confidently? speaker-1: So I think there are two two phases to it, right? I think I'll spread your question two part. First is how do you learn that ⁓ how do you learn when the AI agents are really good at coding? And then the second part is how do you make sure that the agent which is hundred percent confident in what it has done does not go and break something or does not go and ⁓ basically brings your whole business down. which we have seen examples. You have we might have heard the news. So on yeah. On to the first question. I think how do you build that skill set is through understanding what models are good at and let them do what they're good at is writing the code, ⁓ taking the requirements and If you have something really documented with right guardrails, it will be like 90% be able to take you to the finish line. So that is one thing. How do you take an ambiguous problem, a fuzzy problem, and build it into like something which has clear defined goals and guardrails? I think that is one skill set. And then the next one is understanding the system, what you're working on, so that You can make the right judgment. So what I want to do, or like what I've been trying to do on my team at least, is pairing our junior engineers with the agents that let the agents write the code, but you are the first reviewer of the code. ⁓ review the code, and that's how you learn through it. ⁓ that's the grant work you have to do to make sure that the agents, what they're writing, is hitting the all the goals what we have set for them. They are following the right standards what we speaker-0: What what what's really funny is you're kind of describing the QA process on a lot of teams, right? It's it's somebody that's responsible for understanding the requirements and then validating that the system built what it said it was gonna build. ⁓ that it's it's kind of the first wave of testing. speaker-1: Yes, to a point, yes. ⁓ first variable testing or how I see is the judgment calls on ⁓ no, but yeah, I I think I I will agree to it's like the first line of testing. speaker-0: Yeah. So, you know, given that, given that agents are are becoming more and more ubiquitous, right? They're doing more things. It's not just code that they're writing anymore. I mean, I've seen some systems that are designing the entire platform, right? You give it a, hey, I want to do X, and then you've got a variety of agents that go out there and define it and build all the specs for it and build all the user stories and you know, pull start pulling it together. What is a function, okay, let's fast forward four or five years. What's something that that the agent still can't do? What's the one thing humans are gonna have to be there for? speaker-1: I think th final sign offs, ⁓ like basically building that I an agent can like at least how I see it when I build a system ⁓ or something which is going out, what is my confidence level that this like what is keeping me up at night is like, okay, mm an agent can build a code, but when I did that or somebody on my team reviewed it, deployed it. I can sleep peacefully at night because they know what the edge cases are and ⁓ what the domain like basically the domain knowledge they had built over years. But now if I have autonomous agent which is making a change in my infrastructure, which is making a change in my platform. Will I be sleeping at night if they I give them permission to deploy? I I'm not sure maybe the word can change in five years, but I think the final sign offs, the final judgment that ⁓ this can go into ⁓ this can go into production without breaking shit, ⁓ or without ⁓ impacting my consumers, my business, I think that is one thing which will become very relevant and at least ⁓ need to be b ⁓ need to be taught about how we gonna do that. speaker-0: It it's funny that you you put it like I mean, this this is a recurring theme in our episode so far has been the the shift in kind of what the human element is morphing into, which is more that governance and oversight and orchestration and validation and and the things that requires logic and processing that until we have AGI it just isn't gonna be there and then we're all screwed. Right. and so One of the things that that that kind of leads me back to is is kind of your earlier question then, ⁓ or or statement around, you know, mentoring people and developing people and you know, you've got new engineers coming up, AI's changing that. What is how what is that conversation looking like now with these junior engineers and like where are you telling them to develop skills? speaker-1: So, I think that's I one thing is like this is something which we all are learning as we go how to build Netflix Junior. There is no one playbook for it. ⁓ it's a runtime playbook. We all are in an incident and trying to fix a problem. I think if you break it down into what AI is good at. Earlier we used to have speaker-0: Right. Yeah. speaker-1: mm, what colleges or graduates were learning in college, how to write a better code. and how to write like a good test. ⁓ I don't think like a lot of folks ⁓ enjoy writing tests into ⁓ speaker-0: It it it it is it is a there's there's a reason they're they're a a development skill set but they're different than developers. speaker-1: Yes. So I think two things where like I will I think the shift needs to happen. One is f moving more from moving away from going deep into building that writing those cool codes, then understanding the system. Now I think it will become more relevant you understand what an object oriented programming is and how basically those systems work actually. And then all the analytical skill set where like you understand how to break ambiguous problem into clear goals. And then the ability to look for like the judgment calls as in like why if the change is done, why it is done and what is the right whether the right checkbox is hitting or not. So building that domain knowledge, ⁓ and if you're in school, building that system understanding. So we do a lot of system design, we do a lot of app architecture interviews. I think for me, those interviews are becoming more relevant or those things are becoming more relevant to build a skill than writing, like understanding Two programming languages or writing a really awesome code because models are getting really good at it. So I think that is where the next evolution needs to happen of the skill set that you can go deep into your system. speaker-0: So it kind of sounds like it's moving more away from the how you write the code to what you're writing and why you're writing it and where you're writing it and all of the other aspects of the code outside of the code itself. Right. Like you said, system design and and why we went with this system over this system or this architectural approach over that one, right? What it was the end goal here. and that that's that's seemingly a big shift, right? From the the last several years of software development, right? Is is it used to be, you know, we we kind of make fun of it, you know, script kitties and all all the people that have come through the ranks, they they learned how to program really quickly, you know, whether it's doing things like Minecraft or, you know, some of the other teaching mechanisms as kids grew up, it was easy to learn how to write code, right? But what you were writing and what you were building And why you were doing it the way you were doing it. Those are skills that I don't know that I I kinda think we lost that a little bit over the last ten, fifteen years. ⁓ have you seen that in some of the people coming through ⁓ right now? speaker-1: Yes, absolutely. ⁓ I I I think I hundred percent agree with it because our focus has been so much on it the execution that we forgot about like how ⁓ and there were layers, right? I think for good reasons there were layers who were with speaker-0: Absolutely. Not not that there was a nefarious reason for it, just it's just the way it played out, right? speaker-1: So I think that is where like those layers will blend together a bit. and now worrying more about what you're building. How it will work rather than like what to write, how it will work and ⁓ how it will look like in next two years, or how the system will scale in next two years, building in that fashion, those things will become more important. ⁓ and I think you basically pointed out exactly how the shift is happening and how I think our education system will need to evolve with it and as speaker-0: It sounds like theory crafting is gonna become more of a skill we're gonna have to teach as opposed to just the this is how you write this code, right? Or or this is how you solve this problem. Why are you solving the problem? W why is it a problem to begin with? Right. Those things it feels like we've kind of lost that skill a little bit. speaker-1: Yes. So technical skill and functional skill. Like I think when I joined the industry, one of my mentors ⁓ said that hey, a successful engineer who is a techno functional. They understand the business as well as they understand the technology. So I think both things are will become important and I think n understanding the functionality, how things will work, what we are building towards will become more important. And Technology is not just coding. ⁓ it is the whole ecosystem what you're building around it ⁓ is important. speaker-0: I I I couldn't agree more. And I I think it's I your point about the education system needing to change. I I'm really curious to see what happens with kind of the the the new generation that's coming up, the kids that are in grade school these days. Like what does the world look like for them in five, ten, fifteen years as they start to hit the workforce? ⁓ i because I think it's just it's gonna be completely different than it is today. ⁓ through no fault of their own, the world has changed. And, you know, AI is good today, it's gonna be even better tomorrow, right? It seems like every week there's a new L L ⁓ model that's come out that is the new hotness and you know, everybody can't wait to see what all it can build and do. we're only three years into this process, right? Like how far we've come in three years is crazy. speaker-1: No, mm, absolutely. speaker-0: Yeah. So give me give me a hot take, right? You know, if you look at the industry as a whole right now, you know, what what is a a hot take on where things are at and where things are going? speaker-1: So I think my I the answer would be a newness. I think I wish I could give one thing. I think AI is going to evolve. It is going to get better and better. ⁓ how we use AI to our advantage or how do we best deploy it or execute with it is big will become more relevant. One thing I think we have not talked about is ⁓ I think we are in a space where right now the Cost of running AI is not huge, but you see the shift hap happening with Fable and with all the top models. So you have to pay the price for the intelligence of the models. And at that point of time, the ROI of using those models to do something will become nuanced. So my hot take is. basically like the execution with AI will become costly and that is where we are moving. So I think a lot of things where we are trying to force to use AI everywhere, we need to get smart at ⁓ mm finding the use cases where AI has a role. And then how I how I talk to my team or how I think about is can a reject solve it or do I need to Mikael ⁓ Cal. speaker-0: Well, I and I think that I I couldn't agree more. And I I saw something I think just this week where Salesforce was spending something like three hundred million dollars a year with Claude. That's just one of the providers, right? While we may not as individuals see it, it is becoming incredibly expensive for companies to focus on AI. And so it does seem like we're gonna have to really zero in on where does it make the most sense and How do we get the best value out of it? Because at a certain point, it's I mean it's pricing things like even game consoles are are no longer a viable product anymore simply because the components are too expensive because they're all going to AI. speaker-1: Absolutely. Like I think ⁓ right now we are in a phase where we are getting used to it, but then it will become costly. ⁓ I think we went through a similar journey through cloud, right? We all moved to cloud, ⁓ speaker-0: You're you're the second person that's drawn that analogy. speaker-1: Yeah. ⁓ we had unlimited budgets, we had a lot of discounts. Then I I think cloud is there and cloud is relevant. Like I I think it has changed how overall industry works, but now we are more conscious about our optimization or we are conscious about how to do it right rather than just doing it. So I think that is we have to take the same journey with AI to do it right rather than blindly doing it everywhere. speaker-0: I I mean, yeah, i it again, you're the second person that's drawn that parallel to to the shift to the cloud and what that looked like and how we had to go through this this learning period, I guess, is the best way to put it. And at least with cloud, the technology was largely understood, even if the the mechanism was different. ⁓ with AI, we're learning about the technology as we're going. So we're literally building the car as we're trying to drive it down the freeway. speaker-1: No, no. I'm gonna pick up ⁓ pick up on that. speaker-0: Well glad glad I could throw one up there for you. ⁓ Pius, this has been great, man. ⁓ tell us how we can learn more about you. Where can we find you online? Where can people get in contact? speaker-1: So reach out to me on LinkedIn, Pyush Tavari. ⁓ happy to chat. Like I think I've been talking a lot about AI. So ⁓ let's connect it. We all are learning in real time. ⁓ or driving the car as as we build it. Yeah, probably let's learn from each other, right? I and I think we have talked about like a lot of things about like how there is some ambiguity around it, but I think the future is with AI. speaker-0: Belder. speaker-1: So we need to embrace it and embrace it the right way. speaker-0: No, I I totally agree. So ⁓ thank you again. I'll put all those links in the show notes. ⁓ get get people to contact you. ⁓ but you know, I g if this has you thinking about how to build or work with AI, be sure to check in, keep keep coming around for the the conversation. ⁓ we're gonna keep bringing on people like Piush to talk about this. In the meantime, I'm John Robinson and we will see you next time. Thanks everyone. speaker-1: If you enjoyed this episode, follow TestGrid for more conversations with engineering and technology leaders building the future of AI-driven software delivery. Thanks for listening to AI First Engineering.