#272 - AI-Native Is a Journey: From Coding Agents to Software Factories - Vinay Perneti
“AI-native is not a particular destination, it’s more of a journey.”
What does it really take for an engineering team to become AI-native? Vinay explains why AI-native is a journey rather than a destination, and how engineering leaders can keep their teams ahead as models improve.
In this episode, Vinay Perneti, VP of Engineering for Cosmos at Harness, shares findings from the 2026 State of AI-Native Engineering survey and what they mean for engineering leaders. He walks through how software development moved from code completions to coding agents, and why code review has become the new bottleneck. Vinay explains how teams can ship with confidence using agentic loops, such as a PR-to-merge loop, where specialized agents review the code and humans focus on intent. He also covers how to think about AI cost, model choice, and ROI, and what a software factory looks like in practice. Beyond the technical side, Vinay discusses the six skills his team now hires for, how to handle fear and uncertainty about AI, and advice for junior developers entering the industry.
Key topics discussed:
- Why AI-native is a journey, not a destination
- Why code review is the new bottleneck in AI-assisted teams
- How a PR-to-merge loop ships code with more confidence
- Choosing the right model for each task to improve AI ROI
- What a software factory looks like in practice
- The six skills to hire for in the age of AI agents
- Helping your team move past fear and uncertainty about AI
- How junior developers can stand out in the AI era
Timestamps:
- (02:20) What Did the 2026 State of AI-Native Engineering Survey Reveal?
- (06:58) What Does It Really Mean to Be AI-Native?
- (10:21) How Do You Ship with Confidence When Code Review Becomes the Bottleneck?
- (15:18) How Can Teams Tackle Cognitive and Intent Debt with Spec and Intent Reviews?
- (18:31) What Is Agentic Loop Engineering, and What Does a PR-to-Merge Loop Look Like?
- (28:08) How Should Leaders Think About AI Cost, ROI, and Model Choice?
- (33:09) What Is an AI Software Factory, and Will Humans Still Be Needed?
- (38:30) Can You Automate the Whole Cycle from Spec to Merge Today?
- (40:30) Are We Close to a “Dark Factory” for Professional Software Teams?
- (41:56) What Is Augment Code’s Cosmos Platform and How Does It Help Build Loops?
- (45:25) What Six Skills Should You Hire for in the Age of AI Agents?
- (49:44) How Do You Handle Fear and Uncertainty About AI as an Engineer or Leader?
- (55:55) How Can Junior Developers Break into Tech in the AI Era?
- (57:47) 3 Tech Lead Wisdom
_____
Vinay Perneti’s Bio
Vinay is an engineering leader for the Harness Cosmos Software Factory Agent. Prior to that, he was the VP of Engineering at Augment Code where he led Product, Research, and the Engineering teams building Cosmos. Cosmos is an AI platform that helps software teams build towards an Agentic SDLC. Before joining Augment, Vinay led product and platform organizations at Meta and Pure Storage. He’s drawn to problems that live at the intersection of technology and people.
Follow Vinay:
- LinkedIn – linkedin.com/in/vinayperneti
- Augment Code - https://www.augmentcode.com/
- 📄 State of AI-Native Engineering Report in 2026 - https://www.augmentcode.com/resources/state-of-ai-native-engineering-2026
Mentions & Links:
- 📝 The software factory needs a faster review loop - https://www.augmentcode.com/blog/optimizing-pr-to-merge-loop
- 📝 How we hire AI-native engineers now: our criteria - https://www.augmentcode.com/blog/how-we-hire-ai-native-engineers-now
- Spec-driven engineering - https://www.ibm.com/think/topics/spec-driven-development
- Forward-deploy engineer (FDE) - https://en.wikipedia.org/wiki/Forward_deployed_engineer
- Cosmos - https://cosmos.augmentcode.com/
- Mentimeter - https://en.wikipedia.org/wiki/Mentimeter
Tech Lead Journal now offers you some swags that you can purchase online. These swags are printed on-demand based on your preference, and will be delivered safely to you all over the world where shipping is available.
Check out all the cool swags available by visiting techleadjournal.dev/shop. And don't forget to brag yourself once you receive any of those swags.
[00:01:58] Introduction
Henry Suryawirawan: Hey, everyone, welcome back to another new episode of the Tech Lead Journal podcast. I’m very excited today to talk about AI-native stuff with someone who work at Augment Code, a company who is helping many companies to become AI native. Vinay Perneti here. He’s the VP of Engineering at Augment Code. So Vinay, welcome to the show. Looking forward for this conversation.
Vinay Perneti: Yeah, thanks, Henry.
[00:02:20] What Did the 2026 State of AI-Native Engineering Survey Reveal?
Henry Suryawirawan: So Vinay, first of all, I would like to open by, you know, chatting more about this report that your company has come up with recently, which is called the State of AI-Native Engineering Report in 2026. Maybe if you can, you know, highlight to us some of the key findings that you think will be interesting for us to start our conversation.
Vinay Perneti: Yeah, I think maybe before I go into the findings, I can tell you a bit of why we even ran the survey. I think the state of the world of software engineering is changing rapidly, right? And, you know, it’s hard to believe that as recently as 18 months ago, we didn’t have coding agents that people could trust and rely on. And yet now, I think a large part of the software community would say there’s no going back. There’s no going back to a world where you’re writing code by hand without using agents. So the thing that prompted us is to kind of get a snapshot of what’s actually happening in the community out there with, among software teams. How are they navigating change? That was kind of the motivation behind it.
And I’d say I think a couple of things that stood out for us from that survey. A lot of folks talk about being, about going AI-native. AI– Going AI-native as an important part of like the evolution of where they are. And one insight that we took away from that is AI-native AI native is not a particular destination, it’s more of a journey. And the reason why that’s the case is the models are improving at an exponential rate. It’s mind-blowing to sort of see what they’re capable today compared to what they were capable 12 months before. So if you had this destination in mind, then you’re kind of missing the point, I think is something that we learned. So that’s one concrete thing, which is the most forward-leaning engineering leaders are thinking about AI-native as a capacity building exercise within their team, and it’s more of a journey than a destination. I think that’s one.
The second thing, I think this was, the survey was earlier in this year. We… this is self-reported, right? These are about 219 engineering leaders that self-reported. The metric that people reported was actually surprising to me. The, you know, the team mentioned that 48% of the code that the teams are producing is now AI-generated. That was quite surprising to me. I knew, you know, we are in the Valley and sometimes I think that in the Valley we are much more forward-leaning into AI. But the fact that of the survey that we did, the team saying that 48% of the code is AI-generated was something that’s very surprising. This–
The third thing I’ll call out is the people side of things. I think, as this transition is happening, a number of software engineers are actively thinking about what does this mean for their career. And I think on the engineering leadership front, not enough leaders are putting in enough energy to describe what the new roles are and expectations are. And I think that’s leaving a number of people, like 84% of the people that reported, for example… Actually, let me– I think I’ll read you that one. So specifically on the one– I guess two metrics I’ll point out in terms of patterns that are coming up is 63% of the engineering team members, right, have raised that they don’t know what the role definition is. They don’t know what the expectations are. And there’s– generally, there’s a little bit of fear about what AI means for folks. And from an engineering leadership point of view, 84% of them have said, “Hey, are we shipping with confidence? Like how do I know we’re shipping with confidence? What does this mean in terms of codebase understanding?” Those are some of the patterns I would say have emerged from the report.
Henry Suryawirawan: Yeah, thanks for sharing some of these insights. I find it quite interesting, right? Although when you say 48% is AI code-generated, right? I would assume it’s actually much higher, especially in the Silicon Valley, because what I can see, especially in some of the developers that I work with, you know, once you use agents to generate code, I think there’s no looking back. There’s just no one who is willing to write code by hand anymore. So but it’s quite interesting, right? Because everyone is thinking about this, on top of mind, especially in software, you know, engineering companies, right?
[00:06:58] What Does It Really Mean to Be AI-Native?
Henry Suryawirawan: So one thing that I would like to call out is like, people talk about being AI-native, but in your survey you also found that there’s no one true definition of AI-native, right? And you mentioned it’s like a journey instead of destination. But is there a way for us to describe what does it mean to become AI-native? Is there some kind of like definition in your probably point of view?
Vinay Perneti: Yeah, great question. You know, if I sort of chart out, the kind of evolution that’s happened in software development with agents, maybe I can break it down into sort of five episodes. I think the first when Copilot is the first product that came into the market that sort of made it clear what is possible with AI, and a lot of what they were doing was code completions, right? Like code completions, leveraging LLMs. That’s the first episode where people sort of imagine what’s possible.
The second one I would say is chat and agents or like early agents, where they could do a little bit more, they could respond to questions about the code base. That was the second episode.
The third episode I would say is fully leaning into agents in the IDE and then pretty quickly CLI agents. And then what’s starting to emerge now is more around teams building agentic loops. You’ve heard, you might have heard people describe them as graphs, and then ultimately sort of going towards a software factory.
Now at any given point in time, if you– if an engineering leader defined being AI-native as my entire team is using code completions, I would have been like, “Great,” but what about when chat and agents came along? Okay, then that’s no longer the thing, right? Okay, then what about when IDE agents were there, and then everybody was talking about CLI agents being the thing to do parallel agents. And now I think we’re very much on the cusp of what is loop engineering and what is software factories, right?
So going back to the definition of what does it mean to be AI-native? To me, I think it comes down to, is your team operating from the beginner’s mindset where you’re constantly thinking about what is the state-of-the-art? As an engineering leader, are you creating the space within your organization for experimentation and learning? And are you constantly figuring out what’s the best way to get the most out of the current technology and then pushing the boundaries continuously? So that’s why I point out that it’s more of a journey as you look through. And by the way, what I’m describing has all happened in the last two to three years, right? It’s kind of how I see it.
Henry Suryawirawan: Yeah, it’s pretty, how should I say, very surprising in terms of advancement that really happened especially in the last one year or so, I would say, right? Some of these models, the frontier models are also competing against each other. And now also like we can also see some of the areas in the open model, you know, open weights, those kind of thing, are also changing the landscape, right? So I think definitely very exciting and I– we can see everyone is, you know, creating much more capable model. And I think engineering, engineers, we are kind of like being pampered by a lot of options.
[00:10:21] How Do You Ship with Confidence When Code Review Becomes the Bottleneck?
Henry Suryawirawan: But first of all, I think all these advancements are good, but there are also a lot of concerns that many people are experiencing, right? So for example, I think in– it’s also pointed out in your report, people are saying like they don’t have full confidence about the code base. They probably don’t understand fully what code are being generated by AI now. How do you see all these concerns? Because as we advance more, and especially if we go to software factory episode that you mentioned, right, we’ll have this problem a lot more. So maybe, let’s hear from you what’s your take on this?
Vinay Perneti: Yeah, you know, that’s a great point and that’s something that we think about a lot. Here is kind of my mental model around this, right? So I would say for the longest time, maybe twenty, thirty years since software engineering has become a practice, producing high quality code has been the biggest bottleneck, right? And code review was almost a thing on the side that was meant to actually serve a couple of different purposes. If you look at code review from a first principles point of view, it is actually serving two needs. One is maintaining the code quality. That’s definitely one. But the second one is actually, that was a mechanism by which people understood the code base and they were coaching and mentoring the other newer members of the organization, right? So those are the two things.
With agents increasingly being able to produce really high quality code, like in fact right now, I think if you have a very good spec, an agent actually is quite good at producing, like implementing the whole spec, right? So the first problem that I see all the– a lot of the teams running into is they now have agents generating a lot of the code, and everything is piling up at the code review phase. And I’ll tell you an anti-pattern that doesn’t work, which is an eng leader says, “We cannot afford to slow down. We need to make the code review go faster. Go, go, go.” And pretty quickly, what you would see people do is, “Okay, I guess I’m being asked to push the code review quickly, so I’m just gonna rubber stamp them.” And that’s when that then leads to a problem of, wait a minute, like how confident are we as an organization, with the code that’s being shipped? And more importantly, when incidents happen, how do we know like what’s happening in the system to comprehend it?
There’s a couple of ways of, I think, addressing it. One is you say don’t use agents, but we all know that’s not– that’s, there’s no going back, right? So then you then think about what is the path forward that allows us to build the confidence in what’s actually being shipped and maintain codebase comprehension. And I think I see a couple of– a lot of experimentation going on in this particular area.
What we’ve been doing at Augment, we’ve also been experimenting what it makes sense. We have come to the conclusion that, at this particular point in time, we want to leverage agents to produce really, really high-quality outcomes, right? And what that looks like in practice is the agents generate, let’s say a particular PR. We have a system of agents that actually look at the PR from multiple different angles, doing a thorough and deep code review, looking at the risk of that particular change, and then decide where the human should focus on that particular PR. And the human is looking at the intent of that particular PR, like how is it modifying the codebase? What are some assumptions that this particular PR is changing about my codebase? And by sort of engaging with the PR in that capacity, they’re actually understanding how their codebase is being evolved without sort of getting bogged down by all of the details of, you know, every single line of the code. They’re looking at it this way. Now, the beauty about this is there are exceptions where you may want to look at every single line, but you can do that without that being the norm.
So the higher level, I think, takeaway in this is the bottlenecks will keep shifting. The bottleneck used to be coding, but it’s no longer the case. I think it’s starting to shift to code review. But then how do you tackle that in a way that allows you to ship with confidence and it gives your team an understanding of what’s going on? I think it’s a very important thing that leaders need to be thinking about.
Henry Suryawirawan: Yeah, so I hear what you say. I think in terms of code review, I think everyone is kind of like tackling this challenge, right? Especially if you don’t have standard in terms of PR size as well, that could become really, really challenging very fast, right? Because AI agents, you know, they– it’s like people say producing code is cheap now, almost free. Like everyone can produce like massive tons of lines of code in, within just minutes, right? So I think definitely it’s very difficult to code review by line-by-line, I think, in this case.
[00:15:18] How Can Teams Tackle Cognitive and Intent Debt with Spec and Intent Reviews?
Henry Suryawirawan: So I think maybe you also quoted this research, right? I also had the conversation with Margaret-Anne Storey a few months ago, I guess, right? So she is coming up with this triple debt model, in which she also coined this term about cognitive debt, intent debt, right? Because like when you produce the code, right, sometimes, even for me, sometimes if it happens so fast, I don’t even remember what I produced, especially if I didn’t craft it by hand, so agents are doing it for me. Sometimes there’s this, kind of like loss of understanding, loss of reasoning why it was there in the first place. And the intent debt also is because like if there are things that you opted for but you didn’t actually put down why you choose that options, right? Sometimes you forgot like what are the trade-offs that you actually thought about when you produced that. So how do you tackle this as well? Because I’m sure every company that goes through this phase, AI-native companies, right, will have faced this challenge, right? It’s not just the technical debt portion. And I know you blog a couple of blog posts about this as well. So maybe what’s your take about cognitive debt and this intent debt as well?
Vinay Perneti: Yeah, I think to me, going back to first principles thinking and sort of looking at where the world is heading and how do we sort of align with that thinking, right? As I mentioned earlier, it’s now abundantly clear that there’s no going back. There’s no going back to a world where you’re not using agents to sort of in software engineering. So for us at Augment and a lot of the customers that are building on our Cosmos platform, what we focus a lot on, where is the most highest leverage thing a human can be doing in this particular way of engineering software. And the two areas that we found to be very, very high leverage is the spec review phase, wherein at the beginning of like working on a medium to large size project, you’re actually spending a lot of time getting the spec right, and that’s where I think all the choices are being made.
You talked about, you know, various decisions about should we pick this, should we pick that? All of that is at the spec layer, and then the agents translate that into code, right? But at the tail end, you look at the PR at the intent layer. So you’re looking at the spec layer saying, “Here is what I wanna produce.” And then you have a system of agents that are producing the code, that are validating it, that are doing code review, that are making sure that it, all the right things have been done. But then a human comes at the tail end and says, “Wait, is this actually achieving the intent of what I started with?” And that to me is, how I think we address the cognitive debt and maintain high quality outcomes.
Henry Suryawirawan: Yeah, so I think a lot of people are also doing this spec-driven engineering, right? So also like capturing intent I think is always difficult, right? Even before AI agents were there, because like people maybe were lazy to, you know, document stuff, right? And it also goes back to the practice within engineering teams. But I think this just exacerbate those problem, right? Because if lines of code are produced not by the person itself, I think the intent might be lost.
[00:18:31] What Is Agentic Loop Engineering, and What Does a PR-to-Merge Loop Look Like?
Henry Suryawirawan: So let’s go through all these five, maybe not all five chapters, but I think we can skip some of the chapters, right? So we are currently maybe in the agent in the IDE / CLI phase, right? Many companies have implemented this. I know maybe some who are still using probably the chat-based and things like that. But let’s assume everyone now is at the agentic CLI or IDE layer, right? So you mentioned the next couple of steps would be to have this loop engineering, agentic loops, and also software factories ultimately. So maybe some people may still don’t understand what does it mean to have agentic loops. Can you explain to us a little bit what do you mean by this agentic loop engineering?
Vinay Perneti: Yeah, maybe I can describe the kind of the patterns that I see in terms of what people run into, when they max– when they reach the, maybe the max throughput of individual agents, right? So in the five-episode sort of journey that I described, typically, I think, I would say the first three are very much individual developer tools, right? If you look at code completions was very much for the dev– for an individual developer. Chat agent, IDE, CLI. It’s all meant for an individual developer accelerating their coding. So what did most people do? They all used that to write code faster. So when that happened, you very quickly ran into this code review bottleneck, right? So the first pattern that I would say people run into is like, “Wait, I have this code review bottleneck, and I don’t know how to solve it. I can’t– Like, me giving my developers a better coding agent does not solve my code review problem.” So you start seeing other things break down in the software development life cycle. That’s one thing.
The second thing, I kind of alluded to this earlier, eng leaders are now starting to worry about, “Oh my God, there’s so much code being produced. Like, how do I know that I’m shipping with confidence?” That’s the second thing that people run into, right? The third thing, here is an interesting one where for almost like every organization that I’ve had a discussion with, there’s the top 5% developers that are producing at 10x more throughput than the remaining 95% because they have these incredible setups that they have figured out: how they prompt the agent, what skills they’re using, what MCPs they have set up, and they haven’t been able to translate that, right? So these are typically the problems that folks run into and see.
And now the question is: how do you go solve them? And, you know, the whole concept around loops and loop engineering, there’s I think multiple different definitions of it because it’s still not something that’s standardized on. One definition of loop engineering is still very much at the individual level. I think for an individual, you would say set up a loop that then goes, pulls your, you know, PR on GitHub, and it responds to your PR comments if something happens. That’s one individual loop. But the more forward-leaning AI teams are setting up what I call team-oriented loops, right? And these are meant to tackle specific parts of the software development life cycle towards very specific outcomes. And I’ll give you, maybe illustrate two examples of the most common loops that I see people set up with Cosmos, our platform.
The first one is what I call the PR-to-merge loop. And the PR-to-merge loop is basically one where once you’ve created a PR with whatever coding agent of your choice, how do you get, go from there to merging it with high confidence? I think that’s the problem it’s tackling, and the outcome is a high-quality merge that happened, right? And the way you do that is there’s a couple of different components. With agents, I think what we are finding is rather than having one agent do many, many, many tasks, you have well-scoped agents that are doing specific tasks that are working together that’s actually producing a much higher quality outcome. So what does it look like in practice? In PR-to-merge – we actually wrote a blog about it – but in PR-to-merge, when a particular PR is created in GitHub, have a deep code review agent that’s doing a very thorough review of that particular PR. Some of our customers actually have a security review agent and a reg– and a generic code review agent, wherein each of this agent is using that perspective to do a much, much more thorough review of that particular PR.
And once you have that, you can transition that into an end-to-end validator, right? You can have an agent that actually validates the change and then feeds that feedback back to the coding agent and then says, “Hey, I actually found something, so why don’t you go make that change?” So now you have a deep code reviewer and end-to-end validator. And there’s a risk analyzer or risk score agent is something that you might have seen a lot of people talk about, where this is basically saying, “Hey, this particular change that’s being made, what’s the risk of this change? Do I need a human to be involved? And if a human is involved, in what capacity? Are they– Like, what– Where should I guide their focus? Should they be looking at the architecture of this particular change about maybe an assumption that’s changing? Or no, I actually want them to look at the details line by line because it’s actually important for this particular change that I want that to happen.” Now, and then you close the loop by saying, you go to a human intent review, and if the human is making some comments, take that learning and go give this back to the deep code reviewer, because the next time the deep code reviewer should be able to say, “Hey, these are some things that the humans have identifying as important. How do we then tighten that loop so that the number of things that the human’s looking at is sort of highest leverage over time, right?”
Now, the loop is the fact that it’s continuously improving from human input, and it– the loop itself can reflect and then saying, okay, where can I improve and what patterns am I seeing that I can catch earlier in the cycle, right? That’s what makes it a loop. And the other interesting part here is where does the human spend their time? Because I think as we go towards this world where agents are doing a lot of the work, the human time becomes the highest leverage thing that somebody might be doing. So how do you continuously optimize that? So that example that I described is a PR-to-merge loop. And, you know, the loop engineering part of it is there’s actually a lot here, and I think people often underestimate that. What is the loop engineering part? There’s– Let me describe it in a couple of different dimensions.
I talked about the deep code review and the security review. Is that the right set of agents for you? Maybe for your team, something else might be more important. Maybe governance is super important, and so you have an agent that’s thinking about data and PII and governance in a much, much more stricter fashion. That is loop engineering to me, right? So what are the agents that you’re curating that gets you the highest quality outcome for your team? That’s one dimension. The second dimension is cost. Cost is a huge factor for people, for organizations because they don’t want infinite– they don’t have infinite budgets. So in cost, there’s actually multiple components of engineering when you design this loop. One is, what model should I choose for these particular agents? That’s one thing. How many agents do I have, right? I don’t want ten agents doing ten things all the time. That’s– I need to engineer towards my a particular outcome saying, what are the specific agents that I would want in this PR-to-merge loop, and what models do I wanna be choosing, right? That’s the other thing. Then the third lever is where does the human sit? Does it make sense for the human to be responding to all of the code review comments from an agent? I don’t think so, right? Like a Sonnet 5 or a, you know, a GPT-5.6, Terra or a Luna are perfectly fine fixing the comments coming from a code review. But you notice that majority of organizations, humans are the ones that are still fixing up comments in a PR, right? So that’s the loop engineering part and how you get to very, very high quality outcomes.
Henry Suryawirawan: Wow, very exciting, you know, the way you describe all these loops, right? Because for some, maybe this is still kind of like not there yet for some teams, right? But I’ve been seeing a lot of people also posting, you know, over the internet, right? There’s so many definitions of this loop engineering. But first of all, I mean, loop is kind of like you give it input, there’s a process, you get output, you probably retain the feedback, and you continue on and on, right? So it could be taken from a small size, you know, to a larger size, right? But what some people, I think many people, where they are at now is actually to create these individual loops, right? Maybe running multiple agents that could run things in parallel and, you know, get back to one if the agent asks back the human, right, and like optimize from the individual point of view. But I think what you mentioned is something like that is, kind of like beyond that, the team-oriented loops. I find it quite fascinating, if people can start implementing this, right? And I think that kind of like makes sense in terms of putting, you know, agents as part of the PR cycle, right?
[00:28:08] How Should Leaders Think About AI Cost, ROI, and Model Choice?
Henry Suryawirawan: But one challenge obviously you mentioned is about cost, right? And I think the trend is gonna be the cost is getting more and more expensive. Although now we have these, open models as well that probably kind of like compensate that in a way. How do you see the trend of this, you know, AI cost? Because I’m sure many engineering leaders have this in top of their mind as well.
Vinay Perneti: Yeah, you know, it’s a great question because a number of people are like, “I’m spending so much. What do I– what am I getting out of it?” I would say this in a couple of different ways. I think for the folks that are in the journey of just in the CLI agents and the IDE agents, I think the frustrating part for most engineering leaders is, you know, they are spending a lot on AI, but their whole team velocity and throughput did not like double or triple, right? They were all expecting 2-3x improvement productivity, but they all got 20-30%. That goes back to systems thinking, right? Your engineering organization is a system. In fact, there’s multiple systems that are at play, so you need to be looking at holistically as to how do you get the best throughput out of your system.
So going back to cost, you can optimize what models you’re using and things like that, but if you’re not optimizing the system, you’re not gonna get the ROI. So the first thing I would, encourage the, you know, the listeners and the engineering leaders out there to be thinking about is, “Hey, where is the bottleneck in my system, and how do I best tackle that bottleneck?” I think that’s one.
Now, in tackling that bottleneck, now of course the cost also comes into play. How do I approach this and what do I do? I think, when I think about what’s happening in the model landscape, it’s fascinating at the rate of progress. So there’s a chart by the METR, the organization that’s been plotting out model intelligence all the way going back to 2022 in terms of like how they’ve been improving in terms of their ability to do long-running tasks. And if you look at that plot, it’s a lin- it looks linear, but it’s a log scale. We are very much on an exponential. And most recently, in the last six months or so, I would actually say we are at a super exponential. Like recently they’ve been growing at an even higher pace, right? And here is something interesting I think we are observing. If you look at Fable, Fable launched a couple of weeks ago now. At the time it was the best model there was, and it was priced 2x of Opus. And GPT-5.6 Sol came in literally a couple of weeks after Fable, almost the same level of intelligence as Fable at half the price. And then right after that, Opus 5 came, almost similar intelligence of Fable and half the price. So I’m– I would actually expect the intelligence curve to continue in that way and the cost I think will become highly competitive.
But the more interesting thing is, do you need the most intelligent model to be writing code for you? I think the answer is already no, right? Like in fact, Sonnet does a really good job at writing most of the code, whereas some of the Opus models are quite good at the spec writing level, right? And even model, I think, diversity matters a lot. Like we found internally in our own research that GPT-5.6, the Sol class of models, do a really good job with code review more so than some of the Anthropic models, right? And for an engineering leader that’s thinking about cost, I would encourage them to think about both cost and outcomes and ROI.
So there’s a question of what is the right model for the task that I’m using? And with model intelligence increasingly exponential, when do we reach a threshold where a more intelligent model doesn’t actually matter for you? Like you’re fine with whatever it is. And then the open source models that are right behind that are catching up, maybe you can get most of your coding done with an open source model, and you’re using the most frontier and intelligent models for really hard thinking tasks like spec review, architecture design, and things like that, with most of the coding being done by potentially open source models. I wouldn’t say they’re quite there yet. I think they’re still a little bit behind, but that could be a possibility a couple of months down the road.
Henry Suryawirawan: Yeah, so I think you brought up a very important point, right? So it all comes back to the ROI, right? So in terms of AI spend, what kind of outcome can you produce probably, right? And is the cost justified or not? And I think the other thing is about, you know, improving the systems, right? So see where the bottleneck, the biggest bottleneck you are at the moment, and try to optimize that rather than optimize everything. So definitely that makes sense.
[00:33:09] What Is an AI Software Factory, and Will Humans Still Be Needed?
Henry Suryawirawan: So the next episode that you mentioned is about software factories. Again, like this is an up-and-coming term as well. Like many people think it’s kinda like you know, AI produce all the software by itself, right? We just give them the spec and it will run everything by themselves. So is that the correct definition of software factories? What do you think, you know, the best way to describe it?
Vinay Perneti: Yeah, you know, I’ll share. Eric Schmidt is one of our investors in Augment, and we did a fireside chat with Eric a couple of months ago, back in February actually. And I asked him this question around, “Hey Eric, where do you think engineering organizations are going? Like what org structure makes sense?” And, you know, I don’t know if he was joking or not, but Eric responded by saying, “Do you need more than like one person?” And I was like, “I don’t understand. Like, can you elaborate?” And Eric was like, “Okay, you have a founder, and the founder has a thesis for what is possible with this idea. They hand this over to a system of agents. The agents go implement the code, instrument the code, ship the code, look at the data, and then keep cycling it until a put– a certain budget is, you know, sort of met. And then you sort of figure out whether that idea is worth it or not, and then you move on, right?” When he said that, you know, this is back in February, it felt like so far out in the distant future that I didn’t know what to make of it.
But I think there is some early patterns of what a system like that would look at– look like, right? And I think that’s g- probably where some people talk about in terms of what a software factory is. But let me share my perspective on how I think about where we are headed. I don’t– I personally don’t think we’ll get into a world where everything is just done by system of agents because there’s so much context that is missing that the agents are– it’s almost impossible for them to get in terms of business context, how the reality is changing, what the world looks like and all of that. So I believe the future is teams of humans working with teams of agents to produce really high-quality outcomes.
Now, what is the role of the humans and what is the role of the agents is what’s rapidly changing. If you rewind back six months before, the role of humans was interactively prompting an agent, and agents were writing the code. If you look at it right now, I described the PR-to-merge loop. The role of the human is looking at the spec level and then getting the spec right, and then the agents are doing a bulk of the work, and then the humans get involved in the intent review to make sure that’s happening. So the roles have shifted. And when I talk about software factories, it’s basically engineering teams building multiple loops and connecting them together and engineering where the humans and the agents are able to produce the highest quality outcomes. So it’s not that you start from scratch and you build something massive. To me, it looks like you start with a simple loop, and you build more loops, and you connect the loops, and then you keep going, right?
As a simple example, if you have a ticket-to-PR loop or a project-builder loop where a product leader or a business owner is describing like what outcome they want, and they feed it to a project-builder loop or a ticket to a PR loop, the engineers come in and they are doing the spec review, they’re doing all of the judgment calls, they’re understanding what trade-offs make sense for what we’re trying to do given the context that they have. That loop executes, kicks it into a PR-to-merge loop, and then the humans get involved in the intent review, that basically looks like a factory, right, to me. Even though I think folks may not be using that term, but that’s kind of, I think, what the pattern looks like.
Henry Suryawirawan: Wow. yeah, I, cannot still wrap my hand, my head around this concept, right? How can you run these software factories? But I think when you break it down into like, one simple loop, multiple loops connected with each other, producing an outcome, probably that is much more, like, makes sense to me. But definitely, yeah, we will see this pattern coming up a lot more, right, when people share how they do it within their organizations. Definitely something…
Vinay Perneti: That’s actually very well said. I think the way I, you know, when I talk to folks, there’s almost a strong reaction against something like a software factory. But when I describe it as, look, you’re setting up these loops, which it makes sense as to why you’d wanna set up these loops because they are very high value in terms of the outcomes that they’re producing, you connect them together, that is a software factory. Like you basically have a system where you have these multiple loops, humans are doing their things, agents are doing their things. That is a software factory, right?
Henry Suryawirawan: Yeah. Although some people think the software factory is like dark factory, you know, the dark room factory, like in the, you know, photo thing, right? Like where you just go there, submit something, and then everything is produced in the dark. And somehow it comes out, something out of it. So definitely, will be interesting this way forward, right?
[00:38:30] Can You Automate the Whole Cycle from Spec to Merge Today?
Henry Suryawirawan: So I wanna ask about this, you know, you mentioned just now the, the first loop is about PR-to-merge loop. Some people are actually also thinking about from, like specification or business requirements point of view until the code gets merged or deployed even, like spanning the software development life cycle even further. Is this something that is achievable? Do you see it in some organizations that you help with, they are also implementing it this way?
Vinay Perneti: Yeah, I think teams are already building these. To me, going– I think there’s a couple of principles worth keeping in mind. One is by having specialized agents do specific tasks, you get much higher quality outcomes. So the job then becomes where do you draw that specialization, and then how do you connect them together to produce a high-quality outcome, right? And that’s kind of what it is. And to your point, if you say spec-to-merge, there’s multiple steps involved. I think the spec review is one where we still think a human needs to be very much involved because that’s the most highest leverage thing you could be doing. And then you could have, if it’s a complex spec, you could actually have multiple agents building parts of the spec and then, like, getting to a point where they’re testing it, validating it, and then you post it to GitHub as a PR, and then that can do a bunch of the next loop kicks in. So then you basically– once you connect these two loops, you get there.
But you don’t get there on day one, I think. I think the way you get there is you start by picking your most painful bottleneck. For almost everybody, it’s a code review bottleneck. Why? Because everybody’s been using CLI agents and IDE. So tackle that first, then you get the most highest ROI from that. And once you’ve tackled and you have high confidence in your PR-to-merge loop, then produce more code. Because if you produce more code before you have high confidence in your PR-to-merge loop, then you’re in this spiral that’s very scary, right? So get the confidence there, and then you move upstream, and then that whole thing is operating much more independently.
[00:40:30] Are We Close to a “Dark Factory” for Professional Software Teams?
Henry Suryawirawan: Yeah. I think definitely makes sense, right? Start small, right? Don’t try to achieve everything. Although these days so many different products out there that try to sell their kind of like AI-native capabilities. How you see all this proliferating of tools, right? Because again, like sometimes it’s also very difficult for us to choose, what matters for us at this point in time.
Vinay Perneti: Yeah, great question. So the– what I’m describing is actually most of it is working with professional software teams, right? I’m not talking about small projects where maybe one engineer can keep everything in their head and like for them, maybe a doc factory is perfectly fine. I’m talking about professional software teams that have paying customers that have certain expectations for uptime and quality and outcomes. In that world, I’m still finding that maybe we are little– still quite far away from a completely dark factory. I think we are still at a world where you need… you talked about codebase comprehension and like understanding what’s happening and making sure the intent is the right place. I think for– in those environments, what I’m seeing is it’s still very much around teams of humans working with teams of agents, but increasingly intentionally thinking about where is the highest leverage activity for the humans, so that as a system, it’s producing very high quality outcomes.
[00:41:56] What Is Augment Code’s Cosmos Platform and How Does It Help Build Loops?
Henry Suryawirawan: Right. So maybe also now is a good segue to introduce a little bit more about Augment Code to us. Like what the, what do your company do, right? What kind of tools that people can explore?
Vinay Perneti: Yeah, great question. So we started back in 2022 as a research lab, and the focus at the time was, we saw what Copilot could do, but how do we bring that to large complex codebases where the quality of agentic coding is very high? That was– That’s what started us. For the first 18 months, we invested in our context engine is what I call it, which is our own custom embedding models, along with a backend that’s highly optimized for retrieval. And we– Our focus continued to remain with enterprise teams, enterprise teams or professional software teams. And our majority of our customer base was using our IDE and CLI all the way back in 2025, right, like in their large complex codebases. But something for us shifted towards the end of last year, I would say, where we pivoted pretty heavily into what I now call Cosmos, which is observing the two trends.
One is Opus 4.5 was the defining model, I think, going back to November 2025, where a number of software engineers came, like came to the conclusion that software engineering is changing, right? And it’s changing fundamentally. And then for us, we started running into the three problems that I described, wherein everybody was running into code review bottlenecks, individual developers, individual– Like, teams would say, “I have 5% that are incredibly productive, but 95% are not.” So that’s when we shifted our focus to what is a team-oriented system. So right now, Augment, the main platform that we develop and then we help our customers with is it’s an agentic platform to build loops or orchestrating agents, and both of them are going towards building software factories. And we built it in a very agent-forward way where you don’t need FDEs. If you look at some of the other customers, like they talk about FDE, and we’re all hiring FDEs. We actually took a different approach where you don’t need an FDE to set up your loops and to set up your factories. We made the system self-aware, so the system can set itself up.
So Henry, so if you want to set up a PR-to-merge loop, you’d go to Cosmos, and then you would say in natural language, “Hey, here is what’s important for me in my organization. I care about security review, I care about governance, I care about these things. So I want you to set up a PR-to-merge loop that gets me these outcomes.” And then Cosmos would just go build it. And then the loop engineering part of it is, “Hey, I’m noticing that we are spending a lot more money in this agent. Does it need to be that model? No, maybe we need to change it to a different model. Let’s run an experiment and then see how the quality compares between now and before.” So that aspect is where you’re spending your time and you’re offloading all of the connecting integrations and like all of the building blocks of how to set up this loop to Cosmos.
Henry Suryawirawan: Wow, sounds really cool. So I think people who would like to actually experiment with this loop engineering definitely can check it out, right? So go to Augment Code and try this Cosmos product.
[00:45:25] What Six Skills Should You Hire for in the Age of AI Agents?
Henry Suryawirawan: So the next thing I would like to discuss with you is the people aspect. You mentioned in your report the, one of the biggest finding as well is about people. The first is to ask you, how does it change for you to hire people these days, right? Because if we go AI-native, more AI agents gonna be in the loop. What kind of skill set do you see for engineers now?
Vinay Perneti: Yeah, this is a great question, and I think in the survey that we talked about, I think a number of ICs reported that there’s– there not being enough clarity within the organization. So back in April of this year, we actually took half a day. I took– I basically took our recruiting, head of recruiting, and a couple of tech leads and a couple of EMs, and we basically went back to the drawing board and then said, “Hey, given what we know about where software engineering is going and what agents are going to be capable of, what are some of the dimensions that become way more important as we go about hiring?” So we blogged about it. I’ll highlight the six dimensions that we called out, right?
One is product and outcome taste. The second one is system and architectural judgment. So this is basically understanding what are the right trade-offs you need to be making when you’re building things and which ones are– where do– where should we take shortcuts and hack solutions together? Where should we slow down and build something that’s much more robust? Agent leverage, this is something that’s interesting in the sense that I noticed that for the type of problems that you want to solve, there’s a muscle memory that you need to overcome to think about, can an agent do it, right? And that’s a little bit of agent leverage. How are you leveraging agents to try to solve some of the problems?
Communication and collaboration is a huge one, and I say that because If the bulk of the doing is done by agents, I think the human aspect of the job becomes having the difficult conversations, having heated debates in a good way that allows you to get to really high-quality ideas as opposed to just agreeing just because you want to, you know, avoid a confrontation, right? So communication, collaboration become super important.
Ownership and leadership, can you own outcomes? I think gone are the days where you can just deliver tasks. I don’t think you can expect in a job anymore to say, “Just tell me what you want me to do.” It’s like, no, what is the outcome that you’re trying to own? And you are then figuring out what are the tasks that need to get done to that outcome. And the tasks might change if the outcome is not achieved by when you execute a certain set of things, right?
And then finally the sixth one I would say is learning velocity and experimentation mindset. I talked about AI-native is not a destination, it’s a journey. But the– what it makes, the journey aspect requires you to embrace a learning mindset. You have to be constantly– You can’t be like, “Oh man, I perfected how to use CLI agents.” But then before you know it, it’s like, wait, what? Now I’m– That’s not the thing anymore. Like, how do I, what do I do, right? So there’s this constant experimentation I think is required.
So these are the six things that we called out. If there’s two things that I would really, maybe surface, I would say it’s around communication and collaboration. I think there’s something there where it becomes super important. And the next one is learning velocity and experimentation mindset. Those are the two things I’d probably highlight.
Henry Suryawirawan: Yeah. I find these six dimensions really interesting, right? Especially when you are thinking about, how are you going to change the practice for hiring new engineers these days, right? So I highly recommend people to check out the blog post. Although when you mention about communication and collaboration, I’m sure in the past when we opted, I mean, not we all of us, but some software engineers when they opted to become engineers, they wanted to work on the code rather than the human aspect, you know, having all these difficult conversations. But now that is changing. I think we all have to kind of like assess again are we going to improve in terms of communication.
[00:49:44] How Do You Handle Fear and Uncertainty About AI as an Engineer or Leader?
Henry Suryawirawan: So speaking about the human aspect as well, one part of it is definitely, you know, the emotion part of every software engineers who get affected by all this advancement, right? So I think, again, in the report, you call out quite interesting, kind of like a perspective where people feeling anxious, although excited at the same time, right? I think that kind of like represent almost all of us software engineers in the industry. How do you then foresee, like for individuals now, what can we do to adapt, right? What can we do to tackle all these advancements? Sometimes we feel, “Oh, it’s great. AI can help me to be more productive.” But at the same time, we’re also thinking, “Oh, AI is going to replace me one day.” Right? So definitely this kind of a thinking is there for everyone. So what do you do you advise us to actually do in terms of individuals now?
Vinay Perneti: Yeah, I think there’s actually two lenses to that. There’s the individual lens and there’s the engineering leadership lens, right? As to what’s the right thing to be doing. Let me start with the leaders because I think it matters more, because that’s the environment that they’re creating. To me, I think, I mean, I can tell you, even at Augment, where we were obviously fully AI forward, I could tell that the pace at which these models are improving, it was a difficult transition for some of the folks on the team, right? In terms of like just coming to terms with what does this mean? What does this mean for software engineering? Where are we heading? What role am I playing? How does it look like? And the best thing, I think to do as an engineering leader is acknowledge it, right? Like if your team is feeling it, acknowledge it, talk about it. This is early in this year, early in– Because if you don’t allow them to get through that, they’re gonna stay there.
So what can you do to help them get over the hump? So at the beginning of the year, I actually took, in Q1 – we had a very intense quarter – but I took two days off and said, “We just– We’re gonna do an engineering offsite.” And we intentionally shaped it in a way that the first day was all is a hackathon where everybody was like build and build with agents only. This is at the beginning of the year, like fully all in on agents, like no writing code by hand. And then to sort of see what’s possible and what role they would be playing and what role the agents are playing. And then the second day is when we actually took some time to talk about how are people feeling, right? And I actually had a Mentimeter on the screen. Like I didn’t know what would show up because it’s a live Menti, and I had the whole team say, “What are the emotions that you’re sitting with?” And we– And what you saw in the report is ac- is pretty much very similar to what we saw in my own organization where people are describing, again, fear and uncertainty at the same time as like excitement and curiosity and like, you know, for what the future might look like. And there’s something very liberating about knowing that you’re not alone.
So when sort of people saw that, right, and saw the two things bold side by side, they’re like, “Wait a minute. I’m not the only one that’s feeling this way. So everybody here seems to be feeling this way.” And that’s the exciting part because we can go through this together, right, I think, is the thing. So then for the second half of the, of that day, we basically went and talked about what do we need to do? What do we need to change? What like what– Where do we need to focus in? And coming out of that is where we invested in the role definitions and like what does it mean for engineering to be in this way? That’s where the six dimensions came in. And then the team now feels much more, comfortable navigating what’s in front of us, right? Because the expectations are clear. They now know what the future holds and like how they can contribute to that future. So that’s on the individual side.
On the leadership side, I highly encourage, if in your one-on-ones, if you’re an organization where some people are expressing this fear and doubt, don’t ignore it or don’t avoid it, right? I think, in fact, that’s an invitation for you to acknowledge it front and center and create the space for your team to talk through that, and then figure out what you need to do to get your team over the hump. And for the individual specifically, by now, I think it is abundantly clear there’s no going back. If you’re thinking we are going back to a world where software engineering is gonna be done without agents, I would really encourage you to like talk to peers, talk to others in the space that are getting a lot of high-quality outcomes from using AI and figure out why you’re not getting those outcomes. And then once you sort of accept the fact that there’s no going back, there’s no… software engineering is fundamentally changing, then I think it liberates you to lean in a lot more, right? How can you leverage agents a lot more?
And in terms of people that are thinking, “What is my role?” Let me make this very concrete, right? Today, I can give the same highly capable agent to an experienced software engineer and to a non-engineer, and just look at the kind of outcomes that they both can produce. Very different. Very, very different, right? So the learnings, the mistakes that you’ve made, all of those are incredibly valuable to software engineering and your job. So the more you lean in and embrace AI and figure out where you can provide the biggest bang for buck, I would say will set you up really well for what’s ahead.
Henry Suryawirawan: Thank you for sharing your offsite story. I find it very, how should I say? Touching, right? Like when you mention people are, you know, kind of like experiencing the same thing, seeing people experiencing the same thing, definitely something that we need to do more often, I guess, like in the organization, in teams, rather than just chasing the, you know, the trends, right? Looking at the news, you know, like some companies have introduced AI, layoffs happening. So obviously all these don’t really help in terms of human emotion. But talking to each other, you know, sharing the, you know, like group experience, I feel is much better for the organization.
[00:55:55] How Can Junior Developers Break into Tech in the AI Era?
Henry Suryawirawan: So speaking about software engineers who are either already in the industry, we know that there are so many others software engineers, the juniors, who are trying really hard to get into the industry. They don’t have the same experience. They don’t have the same know-how. What would you advise to them? Because I think in some parts of the world I know that they are struggling, like they couldn’t find even a job to start with.
Vinay Perneti: So I think for the folks that are trying to break in, I think we’re entering a world where actions speak louder than words. And by that, what I mean is I think there’s no excuse not to build. So build what you can with agents. Showcase what what you can bring to an organization and to a table. That’s the new resume, right? And the other interesting– I think we talked about the six dimensions: communication, collaboration, ownership, learning mindset. How can you demonstrate that in the way you are pitching yourself to an organization? Gone are the days, I think, where the resume is, you know, I won a coding competition. Like that may only– If you’re competing with an agent in a coding competition, I think we know who’s gonna win, right? So I think there’s, yeah, build, I would– is probably the best advice that I would have.
Henry Suryawirawan: Right. Yeah, I think I fully agree as well, right? You need to showcase what you are capable of, right? Not just in terms of resume, your education, or maybe your competition won, but like build something, maybe share on your GitHub, deploy something that people can browse and try to use, right? I think that’s always the strongest indicator that you actually try and you actually learn something from those experience.
So Vinay, we have discussed a lot. Is there something else that you think you would want to call out as well before we go into our last question?
Vinay Perneti: No, I think we covered a lot of good topics.
[00:57:47] 3 Tech Lead Wisdom
Henry Suryawirawan: Right. So in that case, I would like to bring my last question to you which is like a custom in the podcast. I call this the three technical leadership wisdom. So think of it just like advice you wanna give to the listeners before we wrap up.
Vinay Perneti: Yeah, we covered I think a lot of really good topics. Maybe I’ll start by, the first one that comes to mind is, everyone wants to be AI-native. And I think the– if there’s one thing I want you to take away from this conversation is don’t think of AI-native as a destination. I think it’s, the world is changing way too quickly. So how can you think about AI-native as a journey, and how can you prepare yourself for the journey? I think that’s one.
The second thing I think is, for most of the software community today, they’re thinking that adopting a CLI agent is the destination, right? And when you hear things like loop engineering, it feels so futuristic, but I would just say set it up, go try it out. Like it could be Augment’s Cosmos product, but pick your favorite product. It doesn’t actually matter. Go set up a loop, and then understand the difference of what it means to set up an individual loop and a team loop, and how does it actually unblock your team, I think is the, is something super important for where you are, I would say. That’s the second one.
And the third thing, I think this is more maybe for engineering leaders, maybe now more than ever the human element is more important than ever before. I think there’s a lot of talk about AI and agents, but the folks that are bringing that to life are the humans. So think about how can you stay more connected to the team. How can you, if your team for whatever reason is not leaning in, lead with curiosity and figure out how you can help them get over the hump. Because ultimately, I personally believe the future is teams of humans working with teams of agents, and you want the teams of humans to be energized about that future, right? So you wanna make sure that that’s– they are– they feel that way.
Henry Suryawirawan: Yeah, very– last advice, very good one I think for leaders to think about how to feel more connected with your people, right? Talk about it rather than, you know, just thinking about how to optimize AI agents, be more productive, you know, achieve more outcomes. I think that’s also something worth to think about.
So Vinay, if people like this conversation, they would like to connect with you, ask you more questions maybe, is there a place to find you online?
Vinay Perneti: Yeah, LinkedIn is the best place. Happy to, yeah, happy to connect and, yeah, feel free to reach out there.
Henry Suryawirawan: All right. So thank you so much for your time. I’m sure many people have a lot of thoughts thinking about their AI-native journey. Again, it’s not a destination, right? It’s a journey. So thanks for giving us all these thoughts and insights.
Vinay Perneti: Yeah, thanks, Henry. Really enjoyed the conversation.
– End –
