Prakash Chandran on How AI Coding Tools and Agents Are Transforming Software Development

All right, here we go. Hey guys. Thanks Ro.

We're here with Prakesh Chandran, who's CEO for Zano, and we're having a little chat about the state of application development and software engineering in the age of AI because, well, it seems like we're all over the place. Prakesh, welcome to Shaw. Thank you so much for having me, Mike.

People are trying to figure out what's going on here because on the one hand you'll see some CEO OS stand up and say that they are getting rid of all kinds of developers because of AI and they're much more efficient. And on the other end we'll look at surveys and people will say, well, I think it saved me a couple hours. And this doesn't seem like really moves the GDP needle as much as people might think or expect.

And I think we're also somewhere in a spectrum here because, well, there's AI in the era of the copilot and then there's gonna be AI in the era of the agent and they might be fundamentally different. But yeah, your assessment of where we are and what's going on, Uh, it's a good question and you're right to kind of define it as kind of a spectrum. And I'm seeing different things in different organizations.

Um, I can maybe talk about it in the context of our company and our engineers. What we find is the most value comes from like the most senior software developers, uh, and engineers that are leveraging some of these tools, but co-creating and collaborating with it because they're actually able to read everything that is outputting and really can enforce the prompt in such a way to where the output is valuable and tested. What we don't see is like someone that's maybe a little bit more junior vibe, coat something and then just pass it off to be checked in.

And I think that this is relatively common in terms of when I talk to my colleagues where I think there's a lot that you can ideate alongside with the ai, but it, when it comes to efficiencies and just re replacing your engineering staff, uh, I think we're a ways away from that because, um, software engineering itself, not just developing, but the principles around architecture, around performance, around understanding what it makes to build a scalable and secure system aren't going to go away anytime soon. So while there's productivity gains, I think that that cannot be, uh, underestimated. And so that's kind of what I'm seeing today and uh, internally how we're using as well.

And that's interesting 'cause there has been this debate about whether it would benefit senior developers more than junior developers. And it kind of feels like right now the senior developers are kind of at least winning out this argument because they are able to assign a lot of tasks to AI that previously they may have given to a junior developer and then had to correct. Anyway.

Yeah, I think that's true. And I think it's kind of this interesting space that we're in around like the age of learning and literacy as well, because I think what you see is the more seasoned developers they've been learning and programming their entire life, they've kind of like explored all the edges they've got, they've come across all of the gotchas, but for the intermediate to even beginner developers that haven't seen everything, a lot of them right now are sacrificing speed, um, overkill development. And so I think that sometimes even if they do generate like viable code, they're not necessarily able to kind of, um, interrogate it and assess it and check it in the same way as the senior developer.

So it's not that the productivity can't be had maybe further down the development, uh, or the skill kind of spectrum, it's just that that context and that understanding will be different and therefore the productivity of a senior developer that fully brace embraces these tools can be a thousand x. Mm-hmm. One of the developers I talked to said it's been an interesting change in the sense that at least he felt that he was reading more code these days than actually writing it and then getting the prompt to kind of fix that.

And, um, he wasn't quite sure if he enjoyed that, but that was where it was headed. Yeah, I think so. And I think as the models get better, as the prompts get better, um, you'll probably end up trusting things more.

So you'll kind of have an agent, uh, you know, kind of build a certain sub sec, uh, section of the code base and you're like, okay, does this look good? I don't, I think we're a long ways away from like generate this entire thing or this entire microservice. For me at least, I have not seen that I, what I see is like, okay, this small section that I'm able to, uh, read and control, that is probably what, uh, I feel comfortable with.

But you're right, I think it's, uh, it's kind of a split. Some people like that, hey, it's like having a junior developer working with me that I can spot check the work. It frees me up to focus on the higher order bit.

Others are like, well, I don't really quite trust this thing, so it feels like I'm supervising more than I'm co-creating alongside it. Mm-hmm. How will agents change this whole conversation?

Because it seems like to your point, AI is getting smarter and these agents, well, maybe not fully autonomous, they seem to be able to take on a task and I can coordinate them and orchestrate something that feels like a DevOps workflow. So is that kind of the next phase of where we're going and what might that look like to you? Yeah, I think so.

I mean, it's easy for me to just paint a picture where agents are doing everything, but I think there's a lot of prerequisite work that hap has to happen in order for the agents to be successful. And more importantly for the humans that are leveraging these agents to feel, uh, like they trust what the agents are doing. I think we've kind of heard about this whole notion of like spectrum driven development spectrum and design.

And I think really what that means is like going back to the fundamentals of how you wanna operate as a business. So whether you're having the agent, uh, help assist or augment a certain workflow for you, or you're having it help build alongside of you, the spec will become the source of truth, the most important thing that is actually defined by your business, uh, processes and your personnel. And if that is very clearly articulated and written in a way where the machine can read it and work alongside of you, I think that's where you get the best out of agents.

So my, um, the, the summary really is that yes, I do see a world where we are moving to more, uh, autonomy in terms of letting agents in these machines do things for us. But I think we're still a long ways away from people understanding the importance of understanding the fundamentals of what they should be using agent for, whether it's like actually a workflow versus something that should be agentic and defining what success looks like for them. Mm-hmm.

Does the fundamental DevOps workflow change? I mean, we've kinda always had this infinite loop mindset and things are continuously being worked on and it seems like with AI agents will have the same, an AI agent will write code, another one will have to review it. 'cause you can't have the same AI agent review the code that it wrote.

Well probably won't get the right answer. And then there's all kinds of other functions and tasks that need to be done, but the workflow itself is pretty much the same. It's just gonna be operating at a higher level of scale or will the way we work today just fundamentally change somehow?

Yeah, I feel there's, there's two pieces. There's, I feel like the, um, the actual business logic that's kind of domain specific to the business that's creating the software to then come, like software is created for a business outcome. So the software that's being created by those teams, whether it be done with an agent or not, like those domain specific things need to be understood, controlled, and governed.

So that's like the first thing. I think the second thing in terms of like tactically building and the code and the checks and the balances around that, if that spec and if those things are actually, um, set and controlled and defined in the right way, I do think that everything from creation to DevOps can be handled, um, by, by agents. And I, you know, we see this, uh, now more agentic uh, power being able to spin up and work with Terraform and other types of languages.

Even we have a DSL where you can completely spin up everything with an agent, including unit test and mocking. So it is happening, but I do think that that definition, that understanding and that supervision of what the agent is doing is critical. Do you think at the moment we're a little too obsessed about using agents and AI just to write code and we're not thinking through all the other functions, which, you know, as far as I can tell, make up 80% of the workload anyway.

I, I, I a hundred percent agree. I think that, um, a lot of people are saying and using agents even saying like, we, we even use the word, okay, egen era, but what does, what does that really mean? And what, what does that mean today?

What are the actual use cases? And what we find in our customer base is a lot of people are building more AI powered things, not like agents that are going autonomously and doing uh, things or that's, that's not what we're seeing. I think there's some experimentation, but right now what we see is kind of like digital twin, um, knowledge base.

We are seeing AI powered things where, for example, it might take a corpus or media synthesizing it and pushing it forward. And then when it comes to true age agentic, like their sentiment analysis where it might like scrape the web use kind of a rubric internally to like get sentiment based on what a company is looking for. But outside of that, there hasn't been a lot that I've seen that I'm like, oh my gosh, true business value is happening there.

I think probably where it starts, and we even leverage this today, are kind of more dedicated point solutions. So certainly within customer success and customer support where it does have access to your knowledge base and it can serve as frontline support, that's probably where you're going to see the change. Um, the biggest change happen the most right now today and eventually other industries, uh, will kind of be, uh, affected by agents.

But right now that's not what we're seeing. So what is your best advice to folks? 'cause I think you're pretty far down the path and a lot of folks are still kinda feeling their way through this whole thing.

And what would you tell 'em about, you know, how to set this all up in a way to maybe guarantee success because well, you know, you've been around the block before. Yeah. So I think it really starts with, um, actually a documentation process on your most important business processes.

So just even outside of software development, uh, alone, what are the things that you basically are spending a lot of time and resources on internally? What are the things that, uh, you might give to an intern to handle for you with clear direction? That's like a perfect use case for an agent.

Just documenting those out and being prescriptive around like, okay, well what are the steps that it takes to become successful here? And then starting to experiment. And I think that's my biggest piece of advice.

Like, I think yes, there's a lot of hype, but at the same time it's probably also under hyped and it really just comes from putting in the reps, being tactical and being, having a builder's mindset. So I think one thing that's important, like even internally within our organization, we're having internal hackathons and we, we happen to be a tool where you can build agents on, but you can use a lot of other great tools out there to experiment and just start building and try to tackle real problems within your organization. So I think making the space, dedicated space where you can play with AI to tackle defined processes within your company and doing that in a very intentional and frequent basis is my highest and best recommendation.

Don't, do not stand by the wayside and just expect things to happen. You've gotta be participating. Um, you bring up the intern and it's an interesting storyline because, uh, I was talking to one team and they were like, yeah, we assign things to the AI agent that we used to assign to the intern, but then the CO came down and started yelling at us about the total cost of using this thing.

And he said, this is more costly than the intern. Yeah. So how do we kinda reconcile or manage the cost of this thing?

I think that's, uh, I mean it's a really good point and I think this goes into like setting like the proper guardrail. So obviously that goes to the model that, uh, you're using, like the data processing workload that you're gonna put through it, it really is use case dependent. So maybe for, uh, simple business processes like getting a good handle around like what, like your TCO should be a very much a big part of like if we give this and we offload this entirely to an agent, what does that look like?

Do we have the proper guardrails to say like, and I will not spend more than this on this, uh, particular model. So I think that to your point, has to be a part of the calculus and I think not too much too soon baby steps so your costs don't get outta control. Mm-hmm.

Some folks will also say that in addition to the code being a little more on the verbose side, that it's also rife with more vulnerabilities and then their, their security tools are sending more alerts than ever and developers don't have the expertise to go fix all that. So do we need to kind of figure out how we're gonna apply AI to DevSecOps a little bit before we maybe start running faster with all this stuff? I, I mean, I think that's absolutely critical.

I mean, one of the things that we focus on and are looking at internally is, uh, a security agent. And this starts at the business logic level. Like what are the areas when you're building your application that are prone to vulnerability?

Where are there areas where you have middleware or authentication or role level security or any areas that might expose the data internally, externally? And just helping define that along with your security team. And if you don't have one of those just, uh, basically going through a simple business logic security checklist, which even chatt PT is, uh, is very good at helping to, uh, assess if you don't have a security team in place.

Once you have those in place, this again goes back to the spectrum and design, you can then have the foundation to build an agent that works alongside you, that's purpose driven, to like help based on your domain and your business logic. So I think there's the holistic agreement with what you're saying around let's not move too fast until we have kind of the proper guardrails in governance in place, but also let's make sure that the agent itself, even if it's given that autonomy is governed. So for example, if an agent should across the board not have access to a certain data repository, how do you enforce that at the highest level?

You know, and it's important, these tools that you're experimenting with, do they have that awareness or that capability? And that's something that we think about a lot and we talk about with our customers. They are trying to balance, like we're talking, uh, one of our customers is a big bank.

Yes, of course they want their organization to move fast, but they have to industrialize and enforce the governance on top of it. So like I mentioned, what data does the agent have access to, right? What, um, what authentication, uh, modalities are they going to enforce across every new agent being created?

Those are the types of controls that you need to think about. Even if you're not a big bank, like I think that's really important because Asia has so much autonomy, just put in the proper guardrail so you can feel more comfortable about setting it loose. So you mentioned vibe coding earlier.

What is your take on vibe coding? 'cause some folks will say, this is the reinvention of low code, no code, and citizen developers will create more applications than ever. And others are saying, well this is just gonna maybe accelerate, uh, the early stages of application developments and we'll rapidly build prototypes, but all that stuff is just gonna get dumped on professional developers to fix.

Yeah, I think, I mean what's interesting is like both, uh, both of those perspectives are accurate, right? Like I, this is my, my take. You know, I think that right now Vibe coding is associated with a growing number of new developers that for the first time are able to create software.

Why? Because they're using the English language or language to simply prompt and then boom, something is ready to go. So I think in terms of like articulating the art of the possible, uh, for prototyping, for product managers, for designers, it's an amazing boon to accelerate their workflow.

However, over time that, uh, prototyping phase needs to meet production and at that production phase, it's exactly what you're saying. A proper developer software engineer needs to ensure that the scaffolding that was set up or even the business requirements that were defined are actually built properly when you go to production. So I look at vibe coding today as associated with hey vibe coders don't know what they're doing.

It's like the experiment, uh, uh, experimentation group. Over time it's going to, uh, become less associated with this group and more of a modality. That modality is I type something in and I co-create scaffolding and a foundation alongside the ai.

So we think about it with three different modalities. There's the high code, which some people will always prefer. There's the visual development low code, which we, the space that we play in.

And then there's the vibe code, prompt response that will just be three different modalities to create software. And I think that the, uh, input response or prompt response is great for the beginning, the early innings, but not so much when you wanna get into the nuances. Alright.

So is there one thing that you know now that you kinda wish you knew before you started this whole AI adventure? And as you look back and go, wow, if I had thought about that earlier, things would've been gone a lot smoother. I don't think anyone could have anticipated what has happened in the last like, you know, 24 months in terms of AI and it basically turning everything upside down.

I mean, I can't remember. I've been, you know, in, uh, the technology space for a long time. You obviously longer have you, do you remember an event that has changed things so dramatically, even search behavior?

Like, you know, most people I talk to now get the answers from like a chat GPT or a perplexity, and we've seen search volume top of funnel go down across the board. The intent is higher, right? But that shift has happened in the last couple months and who knows what's going to change in the ne uh, next couple months.

So I guess the first thing is just a recognition that this feels like, it's kind of like technology's pandemic. It's not necessarily negative, but it's the Black Swan type of event that has just turned everything on its head and has given way for like, so much innovation. So I think that, um, I feel grateful that I've kind of come up through the ranks for like before the internet existed, the days of Geo cities, and then kind of having some time to work at Google to have the experience that I do to meet this moment where I can now kind of leverage a lot of my knowledge to use these AI tools in the best possible way.

So I don't know if it's like I do something different, um, but I am happy that, again, my statement around sacrificing speed for seal development, a lot of people are building and they don't know what's happening. It's just like the output and okay, I guess this works. So I guess we're good to ship it, right?

But like you, people, like you and I know a little bit better, right? We're a little bit more reserved. And it's not that we don't wanna move fast, but we wanna move fast in a way that is safe and that is secured and that is governed.

All right, folks. You're heard it here. Hey, even in the age of ai, if you don't know how it works, you're gonna be in trouble.

Hey, Prakesh, thanks for being on the show. Thank you so much, Mike. All right, back to you guys in the studio.