How spec-driven development is changing the rules | AWS’ Amit Patel

0:05Welcome to Dev Interrupted. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd Pearson. This week, I'm sitting down with Amit Patel, Director of Software Development for Kiro at AWS. And we dive into how his team, operating like a startup inside of Amazon, created a new agentic IDE called Kiro that uses spec-based development to create persistent structured plans. This approach keeps context alive between sessions and your teammates and fills in a crucial part of the agentic handoff. Amit shares how his team used the tool to build itself, and all by creating the space for engineers to do their best work.

0:44But before we dive into it, in this week's news, we're covering a few topics, including something that came across our desk about AI being in its dial-up era, engineering managers who are paralyzed by fear, and how they act out, a hack that someone did for their Wi-Fi-connected robot vacuum that was just too good to read. So we're going to cover all of this stuff. And Ben, I think let's start at the top. Yeah. So I wanted to cover this story on AI's dial-up era. So this comes from Substack user Nowfall. And look, Andrew, it's 1995 all over again. You know, the AI boom feels like the early internet, like full of hype, destined for a bunch of shakeups.

1:25But while all the while along, we're building these foundations for what the future is going to look like. And, you know, so back in 1995, like, let's all think back 30 years, you know, the internet had just started to go mainstream. Many of us were just starting to get connected to it and imagining the possibilities that it would bring to us. And, you know, looking back 30 years on, in a way, both the optimists and the pessimists were right about many things over the years. But ultimately, it was incredibly hard to imagine what life would have looked like here in 2025, if you were sitting there back in 1995, just getting connected to the internet.

2:04Let's just think about the fact that the internet search or the internet giants now started as a search engine in an online bookstore. Would you have imagined that those would be the titans 30 years later? Truly. And this article brought up everyone's favorite philosophical concept in 2025. That's Jevon's paradox. But approached it in a way that I think is nuanced in a way that not a lot of people have covered this paradox, you know, and that is that, you know, more supply, yes, it can lead to more demand. Like if you build more efficient steam engines, you may just increase demand for steam engines, but there's a, there's a limit to that actually.

2:42And many industries, they grow in demand for a while, but then productivity and demand for that service or that object, uh, reach some sort of equilibrium that sort of causes employment to drop over the longterm. And, you know, we're all like navigating this space of AI with its excessive hype, its inflated valuations, these inevitable failures that are certainly around the corner. But there's a lot of infrastructure that's being built in the space right now that will outlive a lot of that hype and bubble popping. I feel often that we're kind of at like the peak of it right now, especially with like what we're hearing around like open AI coming out and, you know, they've got 1.4 trillion in spending obligations over the next few years and off revenue of like less than$2 billion revenue per month.

3:33You know, like do the math on how long it's going to take them to pay for that without extreme growth. And, you know, but I've seen a lot of these parallels sort of between the internet bubble popping and how it gave us all this internet or inexpensive infrastructure that was then used by all the web 2.0 companies to sort of build the next generation of the web. And, you know, this article really just sort of makes the point that, you know, we can predict change, but we, it's really challenging to know the specifics of that change. You know, very few in 1995 for Saul dating apps or Uber or influencers being a career option.

4:12And similarly, AI is going to unlock a lot of new domains and job categories yet that, that we just don't really have a concept of because they haven't been invented. And, you know, a great example is there's a lot of job transformation that happens during these sorts of changes. Journalism as a function has expanded over the internet era, but the journalist job has declined all along the way. You know, just as a great example, like I would say you and I, Andrew, we are certainly not journalists, but we do engage in a lot of acts of journalism, as do many professionals in our space. Yes. But I wanted to cover this because this is one of the best articles that I have read in quite a while about the state of the AI industry.

4:52And it really closely matches my own perception on where everything stands. And I just love the closing line. Our AI future is loading. We're all still determining that today. I love how it ended. Ben, you phrased that all so well. This article for me was also really amazing to read, really poignant of a time when transformation was happening really rapidly. And none of us could have really predicted how it was going to shake out to be the world that we live in today. Like you said, it ultimately called us this transformation of so many things that we took for granted or just thought were going to be around forever, both in good and bad ways, right?

5:25And so ultimately, it's about understanding where that change is taking you now and really living in the moment. Because I think the people who got the most out of all of that transformation and change were the ones that were looking not at what could be in the far, far potential ahead, all of the hypotheticals, but looking at what's possible right now. Like you hit on how journalism is something that became more approachable for everybody. That's because the freedom of dissemination of information over the internet became so powerful and became so easy for everyone to share their perspectives, their experiences in real time.

5:57And on top of that, it really brought everybody into the world of journalism. It helps everybody think like a journalist. And the same thing right now is happening in engineering. We're inviting the entire world into how the engineering world works with AI. Because AI is making it something that anybody who can conquer language, which everybody can, can use and can do. So we're in this powerful new era where our roles and jobs are transforming. But we're going to see a lot of things of people playing with these new parts that they never had exposure to before. And that's where bloggers and influencers come from.

6:32Like, I'm really excited to see what the people of the world do with their accessibility to technology. Yeah. And one great point that this article makes is that software engineering as a function could or as a role, as a job, could be plateauing either today or in the near future. But software engineering as a function or as a capability is probably just going to continue to increase in demand. It may just not be the role anymore. It may only be like a part of a role. That's right. That's right. Let's jump from the environment that we live in and how it's rapidly evolving into the real life work environments that we live in now and how they're populated by people.

7:10And we have to navigate people problems sometimes. And this really amazing article by Stefan Moreau came across our desk as well. And it's about a study from Work Life that's looked at how workplace environments from managers can create psychologically unsafe environments. We're talking about managers who control power with fear and who act out because of their own insecurities. And these kinds of behaviors, they cause an estimated$36 billion in lost productivity worldwide. I think, you know, many of us have been in toxic work scenarios before. I touch on this because this article in particular frames it in the engineering manager world.

7:49If you've ever been in that kind of environment in a technical one, where there's also maybe, you know, some bad people chemistry going on in terms of a manager who doesn't feel secure in their role or they are lashing out for other reasons. You know, ultimately, we're all human. And unfortunately, sometimes we get the worst of humans as well. But Ben, you want to walk us through some of the ways that Stefan framed how this fear really operates and how our listeners could really recognize it and navigate it. Yeah. So I really like how Stefan framed this through two different lenses. So there's almost like an internal focus to fear where you exist in an environment like you describe where it's psychologically unsafe.

8:30And this can lead to like a vicious loop of things like micromanagement that just destroy trust. There's also a version of fear that's almost more external focused. It's fear that keeps you aware of competitive pressures, technical debt, the ability to optimize for excellence. And the latter type of fear is completely normal for humans or for any life form really for that matter to engage in because it's a healthy way of just promoting continuous improvement, responding to external stimuli, things like that. But there's a lot of great tips in this article. First and foremost, psychological safety is the top predictor of team effectiveness.

9:10Everyone on your team should feel free to voice their concerns. It should be an environment of being able to share opinions in a trusted space. You also should understand the things that you fear to analyze whether they're actually worth fearing or not. You know, I personally always think about worst case scenarios for practically every project I engage in. And it's not to be afraid of them or to get fixated on them, but it's to understand the level of risk that I'm taking on. You know, and usually the worst case isn't so bad, you know, and I'm willing to live with whatever that level of risk is.

9:43But that's a perfectly healthy response to fear. And then if you do feel like you're in an environment where you are maybe stuck in a fear-based leadership loop, just pick one to work on. Whether that's a communication challenge, micromanagement, decision paralysis, conflict avoidance. There's a lot of ways you can break down your fears to determine which one might be causing the most disruption. Literally telling you to pick a struggle. Yeah. Yeah, but there's some great quotes at the end of this too. Like I love fear, fear mediocrity more than mistakes, fear stagnation more than experimentation.

10:23There are healthy things to fear. Just use them to channel yourself in productive and positive ways. Here's the deal. We all end up in these scenarios from time to time. I myself have been in a toxic workplace and the first step is acknowledging it and understanding it for yourself. And then like this said, you know, carving out that safety for yourself, picking a struggle as comical as it might sound on the surface is actually the best thing you could probably do for your psyche, as well as evaluating the things that you are worried about and have fears about. Like, is it worth worrying about even in the worst case scenario?

10:56Literally walk it all the way through to the end. Is it really that bad? A lot of times it's not. So don't make your own boogeyman as well. And if you're in these kinds of situations, you know, we wish you the best in navigating it. And we hope that this article can help you do so. All right, now let's move on to fear of IoT. We have a really interesting story about vacuum robots. Angie, what is this? Okay, I love this one because I'm always really skeptical about the Internet of Things. I'm not that person who has like all of their stuff on their router. I do have a few smart devices around my house, including though a smart vacuum against my own wishes.

11:30It's unfortunately just something that other people in the household really want to have. So it lives with us. I love my smart vacuum. I do not. Me and my smart vacuum, we have an understanding. But ultimately, this story is about how invasive these things can really be in your home. And it hits at how the technology inside one of these particular vacuums works, including transmitting telemetry information about it going around your house and mapping out your home. You know, so you have to send that stuff off to a third-party server to get processed. Of course, there's not enough processing power on the robot vacuum.

12:03But there's a lot of other security concerns in there as well, transmitting really sensitive information about your house, but then also being connected to your local network through systems that may not be very secure. This was an article about how he really got into a war with his vacuum, disabling it, changing how it worked, and ultimately it was issued a kill command by the remote manufacturer. And he was actually able to, or they were actually able to revive it, opening it up, connecting it with Python and creating some custom scripts with custom hardware to actually get it to run again. And this was a really fun article about how you can take the Internet of Things back into your own hands.

12:39And I myself really felt justified about all my hatred for robot vacuums by reading this one. What did you think, Ben? Well, first of all, I mean, it really lends credence to the narrative of how bad centering your life around services that have to exist on the Internet for like core functionality. You know, like if a manufacturer can just turn off your vacuum robot because you disabled the setting on it. Like that's, you know, very anti-consumer practices, unfortunately. But I mean, I love the ingenuity, you know, software engineers are certainly the types of people that will figure this stuff out and make it work no matter what.

13:17But yeah, I mean, there's a lot of, it's just a great journey on like understanding this vacuum and following the journey of this person. You know, there's some great advice on not using your primary Wi-Fi network for IoT devices and sort of training them as strangers in your home. So, you know, I did say I do love my robot vacuum, but it also is not allowed to go on the Internet. And I have actually had a lot of smart devices in my home in the past, but they are always on their own separated network with a lot of restrictions around how they can access the Internet. And many of them are just refused access.

13:51And I'm also a big fan of if you are doing IoT stuff, the more that you can control within the confines of your house. So using local protocols and control devices, the better because then you kind of cut out the needs of internet services. But yeah, it's a great story. So go read it if you just want to learn how to fight off these anti-consumer practices with engineering. And just take this hardware back into your own hands. I think it's getting easier than ever to really like poke at these devices. is we've covered this in a few different stories now recently of people taking everyday pieces of hardware, like a receipt printer or an actual printer, or in this case, a vacuum cleaner, and changing it to modify for how they want it to work.

14:34I think hardware modifications are really expanding right now because you can use LLMs to do a lot of probing information about how the technology works and actually get that architecture up and going. But I'm over here right now staring at my vacuum cleaner like, yeah, I'm talking about you. And it's sitting over there looking back at me. So before it starts moving, I think we should move on to the next story here. Yeah, absolutely. Yeah, so I wanted to cover this one. So Have I Been Pwned out with a really great article written by Troy Hunt. Two billion email addresses were exposed, and the Have I Been Pwned team had to go through the arduous effort of indexing all of these.

15:11I'm a huge fan of this organization. If you're not aware, they're an organization that keeps track of all the passwords and email combinations that have been published to the Internet by malicious actors. This data dump, the most recent one that was covered in this article, is three times the size of their previous record. So 2 billion emails, 1.3 billion passwords, and about 625 million of which were unique. So just an extraordinary amount of data. It brings the total accounts that are on the website to just over$17 billion, about$17.2 billion. Wow. This article has a lot of really interesting technical details about the challenges they faced while handling this data.

15:54The mere act of hashing the email addresses was enough to just crash their infrastructure. So they ended up having to build this fairly complex process of batching things in chunks of 1 million records. And then on top of that, they had to notify 2.9 million subscribers by email that their account was in this dump. So, you know, just a great PSA. You know, make sure to check your accounts if you haven't already. It's a good practice to check this regularly. Yeah, really fascinating story. What did you think, Andrew? This one was really interesting. The size of this breach is huge, y 'all. I know it's like it's easy to kind of roll your eyes.

16:31You're like, oh, another data breach story. That's kind of how I initially kind of glanced at it, too. But, you know, 625 million unique passwords in this dump is really significant. And it's worth going in and punching in your info to see how it works. And speaking of how it works, you go to the website and counterintuitively it asks for your password that you want to test, which is really fascinating to me as like an engineering challenge. Because like Ben said, you know, they had to go through all of the batching to ingest all of this and to actually do stuff with it. But then they have to manage and actually let you check against this data in a really safe way.

17:05So the engineering architectural challenges of making Have I Been Pwned possible are magnificent. And the service that they offer to, you know, the entire world is really impactful. Because you have to think about this in downstream consequences. If have I been pwned didn't exist and people couldn't verify, then people's accounts would further get compromised, which would then cause more things to get into the breach, which would ultimately cause a lot of internet security problems around the globe. So by having this kind of step in place, it kind of like stops the bleeding. It's like a tourniquet, right, on these password breaches.

17:40I mean, it lets people kind of get a control on what they could do with it. And working with this huge amount of sensitive data that in most cases is hash, you can't even see it. You don't even can't verify what's going on underneath. It actually reminds me of when we had Min on the show recently from Transcend talking about the engineering challenges of working with highly private data, data that you can't see, the security practices behind working on that kind of scale of really sensitive data. It's a huge challenge. And it's something that Herner engineering team carries itself as a first class concern within their engineering work.

18:14So same thing is happening here. This is like a highly secure engineering process to make this possible. Really fascinating to me and kudos to that Troy Hunt and the whole team there for continuing to operate the service for everybody. Yeah. And if you haven't listened to that episode that Andrew just mentioned, Make sure you go back and check it out. Yeah, it's a pretty good one. Yeah, so speaking of guests, who do we have this week, Angie? Yeah, so like I mentioned at the top, after the break, I'm sitting down with Amit. Amit is from AWS. He's talking all about Kiro. So you don't want to miss this one.

18:43Stick around.

18:47AI has changed how we build software. But faster code doesn't always mean faster delivery. That's why Linear B is launching The Essentials Plan, your toolkit for measuring and improving AI productivity. You get the new Linear B MCP server, AI Insights dashboard, and developer surveys, all designed to reveal how AI tools impact delivery speed, code quality, and developer experience. Stop guessing which AI tools work and start leading with data. Visit LinearB.io to learn more and unlock the next chapter of AI productivity.

19:23You know, my guest today for those listening is Amit Patel. He's the director of software development for Kiro at AWS. And today we're diving into the new agentic IDE his team built called Kiro. And picture this. It's a tiny team inside of Amazon that starts from scratch. In under a year, they build a vibrant community with hundreds of developers testing nightly builds. They use the tool to build the tool itself. And they even start contributing code back to a foundational AWS service, S3. And what made all of that possible? Well, it's a new approach called spec-driven development. And today we're learning what it takes to reinvent the agentic coding experience.

20:05Amit, welcome to Dev Interrupted. Thank you. Great to be with you. Great. Well, we're really excited to dive into this with you. I'm a Kiro user and I have so much success I've had with the tools. I'm really excited to learn more about where it came from. And I think going back to my tool adoption this year, I found that early AI coding tools, they were great at quick wins, but it struggled on a team level to get code across the finish line to production. And it became a real pain to build complex stuff that was shared across lots of workspaces with project-specific details that, you know, live outside of the chat window itself.

20:41I think it's what we've all learned this year. So Hiro at the top addressed this with a structured spec-driven flow that bakes in all of that context. So what signals did you see that showed you that prompting isn't enough for building real-world software that led you to creating Kuro? Yeah, when we started out with Kuro, it's probably more close to a year ago now, one of the things that we were looking at was what are the problems that developers are able to solve with AI as it was at that time? and what are the things that they struggle to solve. And we talked to a lot of developers inside the company.

21:25Obviously, we have a very active internal development infrastructure and community. And we talked to a lot of those developers. We talked to a lot of customers as to what they love about using AI in the development lifecycle, what kind of frustrations they were hitting, and what they would like to see improved. And at the time, the industry was also improving things, for example, introducing things like steering and rules and other ways to guide the LLM to do what they needed it to do. So the industry as a whole recognized that there are limitations to what can be done in a simple chat window with just Vibesoting, especially when the problem to be solved is a reasonable size problem, a meaningful problem.

22:13it becomes quite difficult to have enough context and enough memory as an individual developer to remember what you did yesterday or two days ago and what you told the agent. Even three hours ago, sometimes I forget what I told the agent to do. So one of the things that we wanted to do was to solve that problem in a more structured way and not simply provide you know snippets of memory or some kind of steering or you know we have those things in Kiro and those things are important but we didn't feel it was enough to just kind of provide those things we felt that in order to build more complex projects in order to tackle harder problems which required multi-day initiative we needed more structure so that's that's the the sort of genesis of the product idea.

23:09And how does Kiro keep that context alive between days and teammates? Because I completely agree with you about that ephemeral nature that happens in the chat. Maybe you don't remember what happened three hours ago. So how does Kiro transform that ephemeralness into something that lasts? The concept we introduce is what we're calling spectra-driven development. And rather than remembering what the prompts were three days ago, So what we're saying is, Kira will allow you to essentially build a plan, a structure for your project, or whatever work you want to do. And it has three stages. So first stage is, and again, this is typical of any software development project, which is, hey, let's define what the requirements are.

23:58That's the first step. Then let's define what the design is going to be for those requirements. and then let's figure out how we're going to implement it in a step-by-step way and make sure that at every step we have made progress towards the ultimate goal and that we are not introducing issues or regressions or whatever else. So those are the three stages of what we're calling spectrum development. And what we've built into the tool is the ability for people to plan complex things and then to have a very structured task list. And the task list that comes out of it is really the way that the um the user of kiro the developer can keep context on what was done you know i completed tasks one to five yesterday now i'm going to go on to task six and seven and eight kiro actually can look at all the tasks that were completed in a previous session and say okay these things are done kiro can also see what the requirements are kiro can also see what the design is.

24:58And so all of that context is available on task six and seven and eight and 20 as you go down that list. So that's how it maintains the knowledge of what it has done, what it's supposed to do, and what the ultimate goal is. Yeah, and I want to call out the great part about Curo 2 and how it captures that spec is, you know, this becomes like a really a first-class experience inside of the IDE in a way that maybe is new to some developers. The spec-driven development in some cases will really push you back into doing the designing, planning, making sure you have a really clear structure because of how critical it is for it to work.