---
title: "If You Wake Up Hoping the Model Got Dumber, Get Out of That Business - [Dev]olution Podcast - Coder"
description: "Listen to the [Dev]olution Podcast from the team at Coder."
image: "https://i.ytimg.com/vi/lx7fAw1EBAA/hqdefault.jpg"
canonical: "https://coder.com/podcasts/devolution/if-you-wake-up-hoping-the-model-got-dumber-get-out-of-that-business"
---

[Back to podcast list](https://coder.com/podcasts/devolution)

# If You Wake Up Hoping the Model Got Dumber, Get Out of That Business

![CEO, Coder](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1694538377-rob-squared.png%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D400%26w%3D400&w=2048&q=75)Rob WhiteleyAug 5 2026

[YouTube](https://www.youtube.com/watch?v=lx7fAw1EBAA)[Spotify](https://open.spotify.com/episode/0qOhnbjHvMC1VeU9blszS4)[Apple Podcasts](https://podcasts.apple.com/td/podcast/if-you-wake-up-hoping-the-model-got-dumber-get-out/id1839475248?i=1000780142780)

![](https://coder.com/_next/image?url=%2F_next%2Fstatic%2Fmedia%2Fdevolution-masthead-bg.af392453.jpg&w=2048&q=75&dpl=dpl_6wsypvurgW3Wca4tyrJ6aVYnLrcb)

![](https://i.ytimg.com/vi/lx7fAw1EBAA/maxresdefault.jpg)

## Transcript

_This has been generated by AI and optimized by a human._

**Rob Whiteley (00:00:00):** This entrepreneur said, "If you wake up every day hoping the models got smarter because it makes your business better, you're in a great place. If you wake up hoping the model got dumber, you should get out of that business. You will be commoditized. That model will eat your business and you will become a feature of it.

**Nicky Pike (00:00:20):** This is \[Dev\]olution, bringing development back to speed, back to focus, back to freedom. I'm Nicky Pike. Okay, so everyone's talking about how AI is going to let anyone build software now. Prompt it, shift it, the whole world to developer, party it in the streets. Cool. Now the first bills are starting to show up. Folks are opening their token invoice and they're finding a number with a lot more zeros than they planned for. The prototypes the product team vibe coded turns out a bunch of them are held together with duct tape and security holes. The thing that nobody wants to say out loud is that we've all been here before. This is a cloud all over again. We ran to the new thing, we lifted and shifted our old mess onto it, and then we got the bill and the breach and we had to learn how to actually build for the new world.

So here's what I want to know today. If wave one was everyone grabbing a shovel and digging, what does wave two actually look like? And is there a map for us getting through it or are we just going to speed run the same mistakes that we made with cloud except faster and more expensive this time? So let's get into it. My guest today is Rob Whiteley, CEO of Coder, where his team builds the governed AI infrastructure, used to run AI agents without handing over the keys to everything. Quick disclosure so that we're all square. Coder sponsors this show and Rob is my boss. I want to make sure that everybody knows I'm not going to go easy on thing. I owe you guys the harder questions because I know where the bodies are buried on this one. The guy I'm asking, he's watched this movie before.

He ran Nginx through the last platform shift. So when he tells us what wave two of AI actually demands, it's worth leaning in. Rob, welcome to The \[Dev\]olution.

**Rob Whiteley (00:01:58):** Yeah, thanks so much for having me. I don't think anyone can be your boss, so I'm not worried about you taking it easy on me.

**Nicky Pike (00:02:04):** I want to go ahead and jump in. We're going to let everybody know who you are. So you've spent 25 years in this industry. You were at Nginx first, the analyst world before that. Now you're the CEO of Coder. You've called this the craziest moment that you've ever seen for the sheer velocity of change. You've also lived through the cloud shift as an operator. So when did it first click for you that this AI wave wasn't just new tooling, that this was another architecture shift just like we saw cloud was?

**Rob Whiteley (00:02:29):** Yeah, fair enough. And you probably can't tell from my super useful looks, but I was also in the dot-com and bomb. So these all come every 10, 15, 20 years, depends on sort of the order of magnitude. But you're right, I do believe the one we're in now feels materially different in two different dimensions. The first is I think the amplitude of the change feels bigger. And what I mean by that is if you think through internet and e-commerce and you think through cloud and mobile and social, with the exception of social, I think each one of those, we kind of always did the thing we did before. It was just better in some way, shape or form or cheaper or it unlocked a new business model. But now what I'm seeing is we as humans didn't really change. What we could do maybe got better, how easily we could do it.

What's so different about AI is it meaningfully changes our role in our own jobs and our role in our own companies. And you can deny that or fight that, but I truly believe modern companies will have an AI contingency in them. So to me, it's the first of the spikes I've lived through that meaningfully affects how and what we do for work, not just how. The other ones, things got easier, things get easier and I might have to think of a new job. So to me, that's sort of one dimension. And then the other dimension is just the sheer volume of change and how quickly they come. A new model every four to six weeks, models getting pulled, models getting deployed, models getting released, new companies, new companies that do the same thing using different words, do different things using the same words. It's just really hard to feel like you can keep your head above water.

So it feels like the amplitude is bigger and the frequency is faster. And so it just feels a little more chaotic. I personally am loving it. I think it's fun, but I can understand why if you're trying to pioneer through a Fortune 500 company, you probably don't like disruption and change management. And this is sort of both of those in speeds.

**Nicky Pike (00:04:45):** Well, and you brought up the dot-com. We both lived through the dot-com boom and a lot of people are comparing this to dotcom and the fact that there's a bubble coming. And I believe that. I think that there is a bubble coming just because of the growth, but that doesn't change what it did when we had the dot-com boom. It gave us information at our fingertips, allowed us to do new things. But to your point, this is different. One, not only in the speed that it's taking, but the way it's going to change the way that we do our jobs, the way that our jobs are going to function, the infrastructure that it runs off, even the abilities that people have and things that they used to not be able to do, now they can do fairly easily with that.

**Rob Whiteley (00:05:18):** That's what makes this so fascinating. And again, I do think there is an interesting parallel. I'm a huge fan that you can learn a lot from patterns of the past and then just adapt it for whatever is happening in the current one. So I do think there's a lot around the dot-com into. Bomb and the bubble. Are we in a bubble personally? And I have no inside knowledge. How can we not be? The amount of money that is getting pumped into this exceeds anything I've ever seen before. I mean, we could have three trillion dollar IPOs all in rapid succession. I don't even know if there's enough liquid wealth in the world to actually fund that phenomenon. But the characteristics that are similar is massive infrastructure build out. The internet gave us an internet infrastructure that led into mobile that has led to the connected world I can't imagine we would be in.

Now we're sort of back to good old-fashioned data center build-outs and all of that. And so to me, you have to have these huge build-outs to go pave roads that we have no idea what we're going to do on those roads in the future. And if you were to try to prognosticate now, probably not worth your time, but history has proven it's far better. You can go all the way back to, and we're here in the US, literal infrastructure build out, like actual roads and the Great New Deal. All of this stuff leads to fundamentally better outcomes over time. It just is a matter of how far out you have to stretch that time. Is it 10 years? Is it one year? Or is it a hundred? Don't know. Feels like it's going to be closer to single digits before we have the big impact where I'll leave the pundits to do the guessing.

**Nicky Pike (00:07:04):** Well, and on the infrastructure side, you've described Coder both on stage and in talks and webinars that you've done as a picks and shovels company for the AI Gold Rush. But one of the things here is Coder's whole thesis is that the tools are interchangeable. So it's not even really the shovels that you're selling, Rob. It's the dirt, it's the ground itself. So walk us through how you landed there. What did you see within customer environments that told you the money was in the infrastructure, not the tools and all the things that we're seeing coming out of the AI boom right now?

**Rob Whiteley (00:07:33):** Yeah. Well, if there's one thing I've learned the hard way, especially at Nginx is developer tooling, developer software of any kind is a bit of a fashion show. Whatever is hot is hot and it can go cold real fast. It can stay hot forever. VS Code has been bucking the trend for a long time, but then eventually variants came along like cursor and then you can see massive adoption shifts. And there's typically low barriers to entry. As a developer, I can port my workflow to pretty much anything as long as I still have GitHub, as long as I still have some of my basics. So to me, it's a hard business if your sole job is to stay in the developer tooling layer because that is where next season's fashion can just outpace you. And so part of it was we are infrastructure software, so it didn't feel like an identity crisis, but also resisting trying to become the hypey part of the gold rush and be part of the core, as we said, picks and shovels, dirt.

I like that. It's the fundamentals. So that's almost answering the inverse question of why didn't we go up to say developer tooling and try to get more into that workflow? But then the other things I think about with infrastructure are if you go back to major platform shifts, there's always some desire to have a vertically integrated stack where everything is made easier. And we've seen this arguably the mainframe was the first vertically integrated stack, but we have seen converged infrastructure, we've seen hyper-converged infrastructure, we've seen cloud, we've seen now what the frontier labs are doing where they'll give you all the things. And that always makes sense in the beginning. And then over time, customers don't want to get locked in either because of compliance reasons, cost reasons, or security reasons. And so they end up having to build an infrastructure layer that accommodates a multi, multi, multi kind of thing.

So I'm never going to have one cloud moving forward. Even if I think I'm all in on AWS, a department will use some other cloud. A department will use some other major SaaS will acquire someone that does something different. And so for me, whenever you're sort of thinking through these major platform changes, the infrastructure layer is typically the common sense place that it eventually sediments into to say, okay, here's how we're going to make it work. And so that made me want to lead into our ability to support all the gold rush phenomenon, but have an enduring place after the gold rush ends. Now we're baked into the stack and supporting all these different clouds because I do think we're going to end up in a world where you will have two to three major AI platforms at all times. One of them may be the batteries included Gemini that comes with your Google work suite.

One will probably be Cloud or OpenAI as the one where you do a lot of your work and then one will probably just be built into whatever platform runs your company. How do I operate all three of those at scale without having completely siloed policies, governance, infrastructure?

**Nicky Pike (00:11:03):** As much as we talk about infrastructure, a lot of what we're hearing right now on AI is really focused around the developer. And you and I, not to name our age like you said, but we can take this personal for a little bit. When we came up through this industry, a developer was a person who wrote code full stop. Now we're out there saying that anybody with an idea can build. That's what AI is bringing out to the table. So when you look back at your own career and you look at what you've done in this industry, have you ever seen another moment like this where the definition of a job just was immediately taken out and moved under your feet? And what did you think about it when it happened at the time?

**Rob Whiteley (00:11:38):** Yeah, I think that the best that I have heard that resonates with me and hopefully those in our audience that are of our era is what happened to finance professionals. Believe it or not, we used to actually do things in real spreadsheets and spreadsheets weren't cells. They were actual paper. They were sheets of paper and you would spread them out. And so a lot of that was done in bookkeeping, how ledgers were kept, how accounts were reconciled. All of that sort of got a little boost of productivity when computers came around, but it's when Excel came around that things just were wholesale shifted. And now I don't need this massive army of finance people to literally create, collate, and store paper. I just need a small finance team that knows how to use Excel. By the way, the interesting thing is that more or less, in my opinion, gave birth to software development because there had to be that displaced talent goes and learns something or it creates a vacuum.

And so what happened is, ooh, let's lean into this Excel thing and more and more software development came about as a result. If I compare that to today, unfortunately our friend, the developer is one of the main actors in this play where do I need as many software developers, and I should be careful with my words, software engineers as I used to? Probably not. The Excel sheets of this is going to be agentic coding or agents in general. Now, there will still be a need for a group of development, just like we still have finance departments. There will be software developers that are engineers that know how to manage this stuff because they're now empowered to do what used to take three or four or however many humans. So I'm not trying to paint a picture where software engineering shrinks. I just think we don't need as much of that skillset.

That displaced skillset will have to go reinvent themselves. But I think what'll happen is just like junior accountants were empowered to do all sorts of amazing things for their business now that they had all the data at their fingertips, we are going to end up in a similar world where, as you were mentioning, people that can now just build software because they have an idea are going to have a huge impact on their business. And those just like Excel now powers way more workflows than it was ever designed to empower. I think agents and AI will do the same, but it will be this sort of Renaissance period, this golden era where everyone's vibe coding, which whether you like that term or not, seems to capture the essence of creating a disposable lightweight project that may only last a couple of weeks to months, just like I would do a bunch of analysis in a spreadsheet and then it would probably stay on my laptop until I got audited within seven years.

We'll end up with all of these similar capabilities. And so I'll end my little monologue on this. I do think there will be as many if not more developers in the world. They just won't be software engineers. They will be people with an idea and an agent. And that's extremely powerful in that it makes us all net better. It is painful to reconcile while we're going through the transition. If in your world, developer equaled equaled software engineer. Then there will be, I don't want to say winners and losers, but there will be pluses and minuses in those columns. Well, you

**Nicky Pike (00:15:23):** See me smiling because I don't even think you know this, but it's fun to watch how great minds think alike because we interviewed Gene Kim of DevOps, The Phoenix Project fame. He actually gave the same analogy. He used VisiCalc instead of Excel, but the same analogy, how it impacted jobs. So it's fun to watch how the great minds in the industry are kind of thinking alike and how they're going on the same path here. Here's the challenge that I want to lay out on the table. So Rob, one of your largest financial service customers, they came into 2026 with 800 applications and their whole plan was to simplify. They wanted to get those applications down to 200, but then they rolled out coding agents across the firm and today you said they're sitting on 10,000 applications. To me, that's wave one and one number. Building got so cheap and so fast that the count of applications exploded past anything that anybody could govern.

And now the wave two shows up to collect. So that's the token bill on everything that's running that. The security holes in the half of it that got vibe coded by people who aren't engineers. And then it hits you. We've got this idea we can't run frontier models on all 10,000 of those things without lighting a great big pile of money on fire. That leads us to the building was never the hard part. The hard part was actually everything that comes up after the building. And that's really where I want to spend our time for the next hour on this, Rob, is I want to start with the bill because that's the part that's hitting inboxes right now. You put out this number on adoption. You said that 61% of engineering teams are already running autonomous coding agents and nearly three quarters want to go straight from experiment to production.

So we've got agents in the building now, but now that the token invoice is starting to show up and people are realizing that they can't run the premium frontier models on every task, you said that the answer is a tiered strategy. Break that down for the audience and what that means. This

**Rob Whiteley (00:17:09):** Very much feels like AWS where I lifted and shifted my application and I ran a small scale. It suddenly felt magical and awesome. And so I green lit the thing to go all the way up and then the meter went up with it and then the bills started coming in. And the beauty of consumption-based pricing like that is it has various elements that may surprise you, like the network portion scales disproportionately, and that's where a lot of the cost comes from. We're seeing that all over again with tokens. I built a little ap, it worked cool. It went viral within my company or maybe out in the real world. And next thing I know, I'm running it at scale, maintaining it at scale and the cost is way up. Not to mention, I'm now confident rolling out these agents to more and more employees. And so I'm getting more and more of these little sprouts that are becoming big oak trees over time.

The technology came, we got enamored with it, and it really does legitimately offer a lot of value, but we probably could have been a little more thoughtful in how we green lit it. So that's the moment in time we're in. And by the way, I put Coder in this. We have some pretty shocking token bills here for our size and shape. So you end up in a world where eventually you are having to say, okay, the period of irrational exuberance is over and we have to go back to managing this in some way, shape or form. You said, okay, how is this going to get tiered? Well, again, let's borrow from cloud. We ended up usually doing some sort of tiered strategy. We put something in front of AWS so that we didn't have to have all the network traffic there. We put something behind AWS, maybe cheaper clouds.

Maybe we repatriated some applications because we got to a point where because of Kubernetes and cloud native software, I can actually manage some of this stuff without having to have this skillset that only AWS had 20 years ago. We'll see that right now. I think what's going to happen is we'll end up putting something in front of the frontier models that are a bit cheaper and do a lot of that heavy lifting. It may come from the frontier model. If we just pick on our friends at Anthropic, there's Haiku, then there's Sonnet, then there's Opus, then there's Fable. Haiku's really good to be honest with you. Haiku's probably as good as Opus was several versions ago. So I think just having that be your sort of heavy hitter, thinking about what tasks really require frontier modeling. And then ultimately this concept of repatriating, I may bring a lot of this back in - house and run open source and open weight models.

Those right now are probably feel a little sluggish compared to the frontier. But again, if I'm asking it to do the right job, it can be the right tool. And I think it's going to be for a vast, vast majority of basic agentic work. Agents are basically just tools in a loop with some inference. If I'm not asking it to do a particularly difficult task, I don't have to give it a frontier model. Right now we're getting so much benefit from the frontier models being able to stay in a reasoning loop better and just keep calling tools. Whereas I think the open source open weights kind of crap out. And so you're like, oh, that wasn't a very good experience. But they're getting better and better and better. And so I think I'll end up with this tiered model where I'm not going to say $0 tokens because I have to pay for the infrastructure, but I'll have relatively cheap tokens doing a huge amount of work.

I'll have a cheaper model front ending a more expensive model for more difficult work. And then the key is just how do you roll that out? Do the tools even allow it? Can you enforce it? So there's a lot of questions, but that tiered strategy I'm already seeing in more compliance oriented companies that are doing it not for cost reasons, but for sovereignty reasons and then realizing actually this is a lot better than I thought. Certainly better than when I looked six months ago, which in AI was an eternity ago. That was probably three or four generations of model.

**Nicky Pike (00:21:18):** Well, okay, we've got engineering leaders that are listening to this. And I do think when we're looking at some of the open source models, they are improving at the same speed that we're seeing everything else and who knows where we're going to go, but we've got engineering leaders that are listening. They're like, okay, well, what do we put on the cheap stuff? What do we put on the cheap tier? What do we put on the expensive tier? How do we start figuring this stuff out? Because most enterprises, as you know, they're still on the one frontier model or one model to rule them all kind of concept.

**Rob Whiteley (00:21:44):** Yeah. I think you have to kind of break it down into the complexity of the task. And the best unit I can think of is compare each model to a certain tenured employee. To me, the open source open weight ones are about the quality of someone with one to two years work experience. So if you would trust someone fresh into your workforce to do a certain task, then you could route it to that model. So I think it's more assume they're up and running. Then the majority, if you think of this as a bell curve, that's probably standard deviation out here. Then there's the standard deviation off to the side of the really complex tasks where we want to funnel that to the R&D, to the innovation engines, to the folks that actually do need a thought partner on cutting edge stuff. And then I think there's the 60 to 80% that fall between those two poles where I'm probably going to go with a frontier model, a cheaper frontier model like a Haiku or a Sonnet.

So I can almost think of it as like, well, I'll give you an example because I'm a nerd, so this isn't my home lab, but Quinn is my daily driver that runs most of my agents. Sonnet is the one that I use to do any kind of output. And then I defer to Opus when I'm in planning mode and I really want to think through something. And so I could see a very similar distribution. The middle is sort of your knowledge workers of two to six to eight years of experience, kind of like the vast majority of your workforce. And then the frontier models are your fellows, your PhD quality folks that are really, let's be honest, in an engineering work, probably 10% of your engineering and they're probably creating 40 to 80% of your value and everyone else is a scaling function of them.

That's kind of how I think about it. And so again, we have all the patterns to match. We just haven't bothered to enforce any of it. I don't know if I would encourage people to enforce. I think we're getting to the point where people need to start enforcing this.

**Nicky Pike (00:23:51):** Well, I mean, to your point, we saw the cost aspect come up through the cloud era. The repatriation to talk about, we started getting the cloud bills, people started repatriation. So how did we not see some of this stuff coming? Why is this coming as a surprise where everybody's kind of, "Here's one model, it'll rule them all. And wow, don't need this model to do certain work and the cost is too high." How did we not see this coming?

**Rob Whiteley (00:24:16):** I think two things converged at once, which was made for a fairly tumultuous period. The first is the frontier model providers are really good, sorry to be crude, drug pushers. They know how to give you a hit of the good stuff and then keep you coming back. So you get a little bit of the frontier model and then those tokens run out real quick and you're like, "Well, I want more of that. " And so I think that classic almost just subsidizing the power users in the beginning so that they maintain a high level of spend really kept us on leaning into AI and wanting to use the more expensive models. And then creating tokens and sort of a funny math that combines compute as well as a quality of that compute made it so that it was just vague enough to know how much am I actually paying for that good stuff.

I'm addicted to the good stuff now and I'm looking and it's like, "All right, so that was more tokens. What is a token?" And so to me, that's a brilliant pricing packaging moment, but I wouldn't stop there. I think the reason why we didn't see it coming is because there really is a tremendous amount of value people are getting out of these tools. And I often say if you're not getting value, it's just because it's not configured correctly. You're either giving it too little context or you've locked it down too hard or whatever. But chances are once you dial it in, these things can be pretty transformative to workflows and you want to encourage that. This is a scary moment in time because we don't know exactly what it's going to do to the labor force. And so you really don't want to discourage people who are using it till you get really high bills.

So to me, there was a combination of a couple things maybe we didn't see it coming and then a couple of things we willfully turned a blind eye because we needed people to get through this moment in time. And it was a lot easier than trying to hit them over the head and say, "You got to use this tool or you're fired," which never goes well.

Nicky Pike (00:26:35): So Rob just compared the frontier model companies to drug pushers. Give you a hit of the good stuff, the really good stuff, then watch you come back for more while the meter spins. And honestly, he's not wrong. That's how half of you ended up with a token bill that made your stomach drop. But here's the thing, knowing why you're hooked doesn't unhook you. So when we come back, Rob gets specific. He's going to lay out the bet coder is making its whole future on, and it's not the one you'd expect from an infrastructure company. He's going to untangle the words everybody's mushing together right now, agents, harnesses, sandboxes, workspaces, because if you don't know the difference, you're about to spend money like you do. And he's going to tell you flat out whether your company should build this thing vertically or horizontally with an actual line in the sand on company size.

**(00:27:21):** The first half was the bill coming due. The second half is what you're going to do about it. Stay with us. All right. So Rob, we've been making the point throughout all this that what we're seeing right now kind of rhymes with what we saw with the move from data center to cloud to cloud native for the last 10 or 15 years, lift and shift and everything. What do you think was the mistake that we made with that cloud shift that leaders have a shot at not repeating this time with AI? Do you think we've learned anything here?

**Rob Whiteley (00:27:49):** For sure. And again, it's one of those things where if you stretch a story out long enough in the moment, it's almost impossible to figure out. To me, if we go back to our favorite analogy or metaphor, I never know the difference between the two with cloud eventually cloud native came out the other end and it was a re-imagining of how do we build applications and services to take advantage of this new thing rather than just lifting and shifting the old thing and trying to run it as a big monolith on this new distributed architecture. So we have yet to really uncover the AI narrative. We are still lifting and shifting our human workloads and asking an agent to do it. I think we're going to have to entirely rethink an application, a service, a business process in a way where it's not sort of like a human, but it could be thousands of concurrent agents that are breaking the task up and then reassembling it.

If you've used any of the new modern tools, they spawn agents and then they wait for those agents. And so the agent is already playing an orchestration role. Imagine that on steroids. And I think we haven't really thought through all the business implications of having to redesign all this stuff and truly go AI native. So to your point, did we learn anything? Could we have learned anything? Well, I think we probably took too long to just give up on doing things the old way and just start embracing a new way. We should have probably just scrapped a lot of applications or just completely firewalled them off and run them in a legacy mode and build the application. But instead we always try to bolt the new thing onto the old thing and build that. All that does is create cruft and a modern veneer over a crappy legacy candy core.

And so I think that's the biggest thing I think we should learn is it's probably going to be a lot faster to just rip the bandaid off and create some of this stuff from scratch than it will be to try to constantly bolt egentic things onto the old core business process. And I'm sort of using business process as the unit here as opposed to an application. But to me, I don't know if we're going to have a lot of applications in the future. It may just be an agent conducting a process and returning an answer. And so to me, that just makes sense as more the atomic unit. So anyway, I think to me learning that now the other thing we learned taking too long is then also not getting married to it or religious about it like OpenStack or Dai, Kubernetes or Dai. I know you come from this world, so you can get religion on a particular architecture and ultimately one of them will win.

And so if you picked that one, great. If not, you just have to be learning the principles as you go and applying those. So to me, it's the architecture and the infrastructure should not be hard coded to each other. You should be thinking about the architecture and then adapting the infrastructure to support that architecture. I think what we did in the cloud is sort of fall in love with an architecture, hard code the infrastructure to it, and then proceed to cram that down everybody's throat, whether it made sense or not for a given application.

**Nicky Pike (00:31:10):** Let's uncover that a little bit because I wonder, can we even really aply? Yes, it feels like the same cycle, but this is completely different. When we talked about moving to the cloud, it was still binary. Infrastructure's infrastructure. Now we're talking agents. Agents are non-deterministic. They're the multivers of software right now because we can ask the same thing and we're going to get different answers. Now, yes, there's prompt techniques and specs and stuff where you can lower that, but we don't truly know what agents are going to do. The idea of is this is completely different. We're now asking something that has an infinite number of ways to respond. How do you build for that? What are your thoughts on that?

**Rob Whiteley (00:31:48):** Yeah, I think my first thought is it's really funny. I mean, it's a great, great, great point. And sort of the magic of AI, at least generative AI is that non-deterministic. And it's funny how we're then trying to bolt as much terministic guardrails on it as possible to then kind of take it back. And so to me, that's that lift and shift. This is the way humans work. We're more predictable in our output because of human nature, so we better make our agents behave that way. I think we'll end up in a world where rather than saying, how do we make a non-deterministic thing deterministic? We end up thinking about new ways of working that benefit from all the non-deterministic outcomes. Maybe they build on each other and I end up in a totally new place that I wasn't expecting. Right now I feel like the way we use agents is, I already know the answer.

I just need it to do the work. Instead of I actually don't know what it's going to do and maybe it uncovers a new problem for us to go sol and essentially invents a new product or service for a customer. And by the way, that future is not around the corner. It's not too far off, but we're going to need a while to sit in that before we can really trust an agent to think on its own. We can't just wrap a giant deterministic wrapper around them and call it a day. So that's kind of the first thing I would think about. But the second thing is there's a difference between governance and rules. And I think what we need is to think about what are some basic principles that we want this non-deterministic thing to abide by rather than here's a set of deterministic rules so that it hopefully gets to the same answer every time I ask it a question.

And so I really think more so than even the cloud era, the governance of AI is going to be the big hairy opportunity/problem over the next couple of years. Because the more we can create guardrails versus gates, the better for actually beginning to harness the value from these things. So that to me is where I think most companies are getting it wrong today, to be honest with you.

**Nicky Pike (00:34:06):** I love that statements that governance is different than roles. And I think that's going to be the key takeaway from this because we are spending a lot of our time trying to figure out how do we make agents perform like humans? How do we make it look like what we're doing? And that's going to kill some of the value of agents to your point. Maybe we'll get answers we weren't expecting. It'll uncover things that we hadn't thought of. So governance is, in my view, not trying to necessarily put rules around it, but how do I uncover what the agent was thinking, how it got to that point? What was it that led to it making this decision? And being able to backtrack that I think that is a primary function of governance. Do you feel the same way?

**Rob Whiteley (00:34:44):** I think we've all experienced the AI or the agent that hallucinated, was overly sycophantic or wanted to be overly verbose. These are all things that can be governed, their behaviors versus I don't think it makes sense to say you can and cannot do these exact list of tasks. We do need to do that, but they need to be guarded around principles and what we want the behavior to be not necessarily an overly prescriptive. I mean, we've seen this with every security standard, not prescriptive enough and it's not valuable overly prescriptive and we just build debt. And so we have to get some kind of governance policy that's somewhere between some vague SOX compliance versus PCIE kind of like whatever the credit card pay. Those too prescriptive, not prescriptive enough. I think we've learned that we can be better and pick policy that goes down the fairway.

**Nicky Pike (00:35:46):** Man, I am so glad that you came back and said, "No, we actually do need to kind of restrict agents."Because I though, "Damn it, I'm going to have to go on with my CEO on air and disagree with him." But luckily -

**Rob Whiteley (00:35:57):** Yeah, do not let your agents go rogue people.

**Nicky Pike (00:35:59):** All right. I'm glad that came up. Well, and that was one of the anti-patterns that you kind of flagged recently is wave one, everyone said that the project managers, they could go out and prototype now and that engineers were going to be toast. But then we started seeing that a lot of the stuff that was vibe coded was riddled with performance, it was ridled with security holes, and now you're starting to see the flip. So make the case for those that are listening. Why do you think that an experienced engineer driving an agent beats a product manager that's driving the same one? What does that tell us about where the real skillset lives with agents and what we want to do with them?

**Rob Whiteley (00:36:35):** I've seen so many times, which is you took the extremely smart entrepreneurial product manager who give them a capable agent and they could take the customer requirements, prototype it and then say, "This is good enough. I'm just going to let it run." It's very tempting to do that and you don't really know the spaghetti code that's under it or how many vulnerabilities there are or rather inelegant agent probably created. Again, in a software role, I think product managers should be responsible for the what and the why. And then engineering does the how and then you kind of argue over the when. I think you have to have a similar relationship with AI. And so I would rather in this world treat AI as a thought partner that helps me figure out the what and the why using a frontier model. And then I'm responsible for the how with the agent as my copilot, not capital Copilot.

But to me, that's a better division of where this technology sort of maps. And so to your point, it leads to the inverse of I'd actually rather an agentic PM and a human engineer than a human PM and an agentic engineer. Because as an engineer, I have fluency on the how. I know if what it created is good or not. If you can't tell if your agent created something that is good, then you can't be accountable for its output and therefore you are putting your company at risk if there are problems. And so I think we just need to be more thoughtful about just getting away from the dopamine hit of I went from prototype to production and it was awesome and more into how do I shorten a loop, but it's probably the requirements gathering loop more than it is the building engineering loop.

**Nicky Pike (00:38:32):** I do agree with that. I do think, and we've said it on this show several times, explosion of software that's going to be coming from citizen developers and the vibe coders. And a lot of that's been focused on LinkedIn. Yo get on Twitter. Well, this is the death of the engineer. I don't believe that to be true. I think vibe coders are actually going to bring the engineers back in because somebody's going to have to go in and polish that thing up. They're going to have to make sure they close the holes. You can't rely on necessarily the agents to do that. They may use agents to do those things, but they'll do it in a knowing fashion.

**Rob Whiteley (00:39:01):** Yeah. And if you, again, not to keep constantly torturing it with analogies and metaphors, but when technology first started showing up in the form of PCs and eventually into the software side with things like spreadsheets, we gave birth to IT departments to be the center of excellence that does that and makes it happen. We already have the nucleus of that with platform engineering. A centralized team that makes this stuff safe for people to use and it paves a golden path to use it. They were doing that for developers, but to me, the platform team becomes so critical in this because they have to do the golden path for those vibe coders. So if you ask your average platform engineer, who's your customer? They'll say the developer. Well, if the definition of a developer has changed, doesn't mean your job definition has really changed. It just means you have a different constituency, maybe a wider constituency now.

So to me, I think we're going to see a huge Renaissance and platform as a center of excellence to make all of this stuff safe. And then the software engineers will be like how security eventually became a specialized function that it kind of sits alongside IT. Those are your software engineers. They're the ones who are actually out there making sure you don't do stupid things that put your business at risk.

**Nicky Pike (00:40:16):** I couldn't agree with you more. I think platform engineering is going to become a much bigger discipline because of vibe coders and citizen developers. And I do think that that's the perfect segue because when we're out there talking to enterprises and we're asking them what their AI structure is, I do think that we have a vocabulary problem out there, Rob. We're seeing people, they're mushing agents and harnesses and sandboxes together like they're all one thing. Now this is the place where you're different. And I want you to lay this out for the audience here. What is a harness? What is a sandbox or a workspace? What's an agent? And why do leaders need to know the differences between these before they go out and start spending dollars on this?

**Rob Whiteley (00:40:52):** Yeah. And this is one of those things where history will prove us right or wrong. I believe we have to be precise in describing things. The exact labels matter less over time. Eventually we'll have a common vernacular. We just don't at the moment. And the language that I hear the most in the industry from customers, from vendors is a sandbox. And it's the idea that I need my agent to be in this isolated environment. Part of that is for performance reasons, but mostly for security, compliance. And so you can run a sandbox on a laptop, you can run it in the cloud, you can run it anywhere. It's sort of the bubble that I wrap the agent in. Absolutely critical. And that is a necessary component of all of this. The problem I have with it is the majority of sandboxes are quite dumb. They're very fast.

They're very ephemeral. They're meant to live for seconds, maybe even down to microseconds for an agent to do a very quick task. Read and write a file, execute a web search, not something where the agent is going to sit in an inference loop and do a lot of tool calls. I think the more you do that, the more that sandbox becomes a workspace. And the difference between the two in my mind is a workspace can be just as ephemeral. It doesn't have to be something that you set up. So it's not the pets to cattle analogy, but it is something where I will create that environment and pre-plummet with necessary tools, necessary policies, necessary context so that it's not just this sort of dumb sandbox where the agent completes something, but really it's like imagine asking an employee to go and sit at a blank desk with zero on it versus go work at a workstation where there's a laptop, docking stations you can plug right in and there's probably some basic amenities around.

That's a workspace. It is the sort of pre-configured desk that the agent gets to sit down at and do its work versus just dropping them off at a blank bench and saying, "Good luck." If all I'm doing is thinking for a second, a bench is fine. The second I actually do real work, I want to give it real tools. And so to me, this workspaces will be where real agents will do real work. Sandboxes are where they're real agents, but they're doing very disposable work. So another way I think about it is all workspaces will have a sandbox, but not all sandboxes will have a workspace. I think you need to think about when do I give it the more rugged version of this, the more built out version? And I think people will be surprised. The more you want to run agents and longer lived, which by the way, a long lived agent could be 10 minutes.

You could also do 25 hours or three days. And then I would really encourage you to think about workspaces. But I think we'll end up breaking things into sort of the unit of work and not the ephemerality, but really the amount of tools it needs. Whereas right now I think we're just overly indexed on ephemerality. That is a byproduct of the infrastructure you choose and both can be ephemeral.

**Nicky Pike (00:44:13):** I don't know if this will hit or not. I always tell people I can put a boat in a desert and I can still do some boat things, but that boat's going to be a lot better when I put it in an environment that was made for it if I put it in the ocean. And I do think ephemerality is a need I think, but most people are, because of the non-deterministic nature of agents, what's my blast radius? How can I control that? I think that's why ephemeral is so important. But to your point, we may be overextending on the ephemeral nature. If you build the environment for the agent, give it what it needs. I think you have less of a worry. I'm not going to say no worry, but les of a worry. Thought there?

**Rob Whiteley (00:44:49):** Yeah. First of all, I love the literal sandbox that you're putting your boat in. That's amazing. You know me and my

**Nicky Pike (00:44:55):** Analogies, Rob.

**Rob Whiteley (00:44:57):** I'm going to steal that one for sure. But yeah, I think you're right. I think to me, the other problem I have with sandboxes is kind of what you're touching on is there's sort of this lightweight and disposable. So I end up not putting a lot of controls into them. And this is where you can put yourself at some real risk is your granularity of control and the blast radius are too mushy. And so an improperly segmented agent can still do a lot of damage because I don't want to bolt a lot of stuff onto the sandbox. It makes them slow or clunky. Well, again, at that point you have a workspace. And so to me, I have to think about what I want the agent to do, how sensitive is its work? And again, we have all these concepts now. We have role-based access controls for humans.

The work to be done in job definitions and trust levels, a lot of that has existed as tribal knowledge and we police it when you break it. Now you have to figure out a way to do that as code and do it as policy as code, governance as code, infrastructure as code. But I'm a big fan that we can take our mental models of how we've been disrupted in the past, how work gets done, how we want to create a new native version of that work and apply it to this world rather than just throwing our hands up in the air and saying, "Well, this is new. I've never seen this before."

**Nicky Pike (00:46:20):** Well, okay. So I do want to kind of bring this back. I want to bring this back to what we were talking about at the start of the show, which is kind of the bet that Coder has made. And you told us that coder is planting its flag on one of the things and it's really that choice is freedom. And that really revolves around model agnosticism. The harness, the tools, the models are all things that you kind of think are temporary, that's just the devs being picky for right now, whatever's current. But the model layer is one of the places that you say you refuse to get locked in on. So explain that. Why is the model the one thing that you won't bet on? And what happens to the companies that pick this wrong?

**Rob Whiteley (00:46:56):** I'm about a hundred billion shy in funding to go do the model one. I don't think the world needs another frontier model provider. We have several. The competition is fierce. They're keeping each other honest. The downstream economics will get made better while incumbents like Google continue upstarts like OpenAI and Anthropic, interesting SpaceX, Meta. There's no shortage of I think folks that are capitalized. And I think it's going to be critical that they just keep that innovation going rather than we try to out - innovate them. So to me, that is the area, that's the economic curve I'm trying to ride, is that models are constantly getting better. I was at this event just last week and this entrepreneur said, "If you wake up every day hoping the models got smarter because it makes your business better, you're in a great place. If you wake up hoping the model got dumber, you should get out of that business.

You will be commoditized. That model will eat your business and you will become a feature of it. " And so I think in my mind, I want the solution to... I want the coder experience to be that every single time a new model comes out, if it adds value, I can use that value. And as they get smarter, my infrastructure gets better. And I might be a rev or two behind or I might starting to have completely totally different incentives like trying to cut my own costs. And so now I'm not going to give you the really innovative thing. And so to your point, choice is freedom. The more we bake that into the architecture and the infrastructure, the more locked in, the less valuable this will be. And we're way too, we're in the first at bat of the first inning. Even though it feels like we've been in this for a decade, it's been like 18 months into what is probably going to be a decade long transition.

And so I think we just have to not hard code too much of that now because then you don't know how the game will unfold. It is essentially how do I serve up that intelligence in the fastest, securest way possible versus how do I be that intelligence? I'd love to be at the helm of any one of those companies, but that's not coder.

**Nicky Pike (00:49:31):** Do you think that the macro you just described makes agnosticism a survival move or is it just a preference? Oh,

**Rob Whiteley (00:49:38):** That's a great question. No, I think it's a survival move. I think if you are trying to build a business and you're going to hard code a bunch of assumptions today at this point in time in the first of nine innings in what many of us think is the largest disruption ever, you're either the smartest person in the world, you're shortcutting. And so I think agnosticism is going to be a business imperative so that you don't lock yourself in. When we're in the seventh inning stretch and we're all going to go get another popcorn, then we can maybe start making some hard coded choices. The outcome is much more understood at that point in time, but we are so far from that. So to me, agnosticism is going to be an imperative because they're going to be geopolitical, there's going to be environmental, macro. I mean, pick your favorite whatever external factor that I can't control is going to affect me.

And my job is to make sure that that doesn't sink the boat in the desert rowing in the sand.

**Nicky Pike (00:50:46):** Yep. Well, I mean that's one of the things that coders... That's our whole reputation, to be honest with you, is being neutral ground.That agnosticism, choices, freedom, whether it be in the runtime that you use, whether you're cloud, on - prem, the models, tooling that you use. We are Switzerland. We don't tell people what they have to use. We just want the customers to be able to run all those safely. But betting on that, right betting on that agnosticism no matter which way you're looking at is still a bet, Rob. It's still an opinion. So how do you square that up when people ask? Are you still neutral if you're telling people which dimensions they need to stay flexible on?

**Rob Whiteley (00:51:25):** It's fair. And I think if you go back to Coder's roots, our roots were always to be as little prescription as possible. World's your oyster. And I think what the founders here unlocked was how do you afford that flexibility without getting a ton of complexity? Because those have always been at odds with each other and then probably security as a third leg of that stool. And so to me, we're now in another world where I don't want a trade-off in flexibility versus complexity. I want to give you as much flexibility as possible. And hopefully our magic sauce is how do we reign the complexity in so that you can continue? So in this world, I think the flexibility comes from the models. So again, that's the one variable I don't think we should lock down. I think a little bit of insight we are going to be releasing something called coder agents.

And this isn't a commercial for coder, I'm just using this as an example, where we will, for all intents and purposes, that is a harness. It is the thing you wrap around the model to make it behave like an agent. And so have we prescribed? No, you can still apply whatever tools to that loop. You can still run it with any model. You can still give it any skill file, any agent's definition. You can apply different governance. So I think it's sticking to our roots of being agnostic, but it is moving a little bit away from, well, what is your goal? Is your goal to make every model available across every tool? Or do I want a golden path where you can swap those out, but I will make some of those choices as a starting point for you? And that's kind of how I think of agents is it'll be what we've seen as some of the best tools configured to run in a safe way, but you can always change it.

I think what will keep us true to our roots is not that it's hard coded to the point where you can't make a choice. We're just in this kind of first or second inning. We have to make it a little bit easier for customers to get value from this. And so let me make not a decision, but let me include some batteries for you and then you can swap them out for better batteries or different batteries or a different energy source or whatever tortured metaphor over time. But to me, I think that's going to be necessary for sort of a time to value argument. But I would keep us honest. If we ever end up hard coding this, then I don't think we will be the coder that got us here. That would be a slippery slope for me. That would be one I'd want to watch out for.

**Nicky Pike (00:54:06):** I do agree. And let's take that and let's kind of pivot into the implementation side of this. I do think one of the strengths of coder was the CDE portion. I think sandboxes got validated as core AI infrastructure fast and hard this year. You had Anthropic, they went and shipped them. OpenAI, went and bought ONA to get at them. And you've even pointed to Databricks putting out a meta harness as well. So for the leaders, they were now convinced, yes, we definitely need this layer. What's the first thing that they should put in place and what do you think is the thing that they'll be tempted to skip that they absolutely shouldn't?

**Rob Whiteley (00:54:41):** I think the temptation is to think of this top down instead of bottom up. If the modern AI stack is a chat interface that is sort of connected to the agent layer, that's the harness, that's running the model as the core brains and all of that is isolated in a sandbox. So think of that as three or four areas. If I'm making my choices top down where I'm picking my interface, a lot of what's below that will get vertically integrated and dictated for me. I think you need to think about it a little bit more bottom up in terms of primitives. I'm going to want isolated environments on which I can run different models on which I can then run a tight agentic loop and I can connect that to any interface. It's really hard to create that any, any, any, any architecture because it tends to be more complex.

I have to make more choices, have to think through principles. But this was sort of the monolithic to microservices to do that thought is more important now. So I would encourage, think a little more bottoms up. Don't just pick your favorite interface. That's the fashion trend that's going to probably fall out of fashion. But now I've hard coded my entire stack to that fashion trend. I mean a year ago you would've been insane if you were deploying anything other than ChatGPT. And out of nowhere, Anthropic became the most valuable company in the world and is arguably lapping them in terms of their capability. I think it would be naive to think that that's not going to happen two or three more times in just the next year. And so I think giving yourself the most optionality possible regardless of it. But to me, it's breaking it down into those kind of layers and thinking of more horizontal, like this is my core AI infrastructure layer, which to your point, sandboxes are a validated part.

Here's my core model layer and I have to think about how am I going to support and route multiple models and here's my harness layer and then here's my chat interface or whatever. Those are all just different layers rather than one big stack.

**Nicky Pike (00:56:50):** Well, agreed because I do want you to make that a little bit more concrete for us because we've talked about the vertical side of this anthropic and co-work and some of those policies, cursor. We've also talked about the coder side of this, which is more horizontal. And you do have those choices. You can build the vertically integrated stack. It's going to be fast, you're locked to that one stack. Or you could do the horizontal one, which is going to be more work, but you get the governance and the policy once at the infralayer per stack. Do you think there are companies out there that should go vertical and other companies that should go horizontal? And what would be the difference on what choice they should make there?

**Rob Whiteley (00:57:28):** Yeah, I think the most logical is probably along a size dimension. If I'm a startup today, it is so much easier for me to go and stay cloud native. I've got dozens of employees and everything I want can come out of Google and AWS and I don't need to go buy all sorts of complicated tooling. But as you grow, your needs become more sophisticated. And the cloud native version from AWS of that or the Google native version for analytics, I end up needing that specialized tool. And then as you grow bigger and bigger, unfortunately, you become a Snowflake, lowercase S, not the company, and everything is kind of very bespoke to you. If you're small and your needs are not complex, go vertically integrated. And I think that's up to, if I had to hazard a guess, medium-sized companies, think a thousand employees. I think once you're over a thousand employees, your needs are going to be complex enough.

You're going to want to be able to start subbing in some best of breed components for certain tools, tool calls, agent layers. Maybe I want to diversify my models at this point because I am getting the token bills. And so that's where I'm going to want the horizontally integrated. And so today I think it's sort of a small up to medium and then sort of a medium up to all the way to the largest should probably do vertically versus horizontally minded architectures. The one exception would be you can be in that smaller camp, but if software or technology is your core competency, you're probably going to go horizontal sooner because you need that ability to stay on top of the innovation curve, which is why we see a lot of tech companies that even though they may not be huge from employee perspective, their needs are that flexibility.

**Nicky Pike (00:59:23):** And that makes a lot of sense, especially from the CEO level where you're coming from. But what about for the person who's actually running this, not the one who buys it? What do you think's going to break first if they pick wrong? Do they have to worry about things like the cost or the data sovereignty issues that may come up? What do you think's going to break first if they get that wrong in their first choice?

**Rob Whiteley (00:59:42):** I think the first thing that will break is probably cost because we know how to measure that and we right, wrong or indifferent are getting billed monthly and so we can see trends. So that one will feel the pain for First, it's going to be the security, the sovereignty of it that will, if it breaks, and I think it will, it'll be existential. That can land companies in a lot of trouble on the wrong side of laws. And so that will probably be a little more behind the scenes, but it's a ticking time bomb for a lot of companies. So I think you should solve for the bigger problems and try to optimize along the way rather than try to make only cost oriented decisions because then you might end up pushing yourself into an architecture that's too prescriptive or standardizing on something that ends up later was highly vulnerable.

Now you have no way of switching it out. So I don't know if that's what you're trying to get at, but those are two of the dimensions that I think matter. Cost is upfront and the bean counters help us with that. And by the way, my sovereignty is you have to have control of dot, dot, dot, your data, your infrastructure, your applications, something. You have to draw the boundaries. Something has to be sovereign to you, otherwise you have no competitive advantage. And so I think you have to think about where do I want... In the past, it was really just data and that was the only thing that was really dictated to us. Where I'm seeing a lot of companies say, well, then it became cloud sovereignty. Where I run the application and where I connect to its data matters. And now we're seeing the era of AI sovereignty.

Where the agent runs, what data it connects to, what applications and tools I give it has to be something I have a lot more control over, especially as I go up that pyramid into sort of larger, more complex environments. I just can't rely on...

We've all had the world where AWS went out and half the internet goes off. Well, imagine when we have that same risk sitting inside of a single frontier lab. So we just have to be thoughtful about that.

**Nicky Pike (01:02:02):** Yeah. Or you get another regulation that takes a model that you've built your whole infrastructure on off the table. There are lots of choices out there. So Rob, now we're getting to the fun portion of this. Rapid fire questions. First thing comes off your mind. This is not something that I want you to put a lot of though on. This is probably the one that's probably going to get you in the most trouble, but we're going to see what you think on this one.

**Rob Whiteley (01:02:23):** Okay.

**Nicky Pike (01:02:24):** Hottest take. Is prompt engineering a real skill or is it just a temporary one?

**Rob Whiteley (01:02:28):** It's a temporary one.

**Nicky Pike (01:02:30):** Okay. One model company you think is not going to exist in the next three years?

**Rob Whiteley (01:02:37):** Meta.

**Nicky Pike (01:02:38):** Meta?

Rob Whiteley (01:02:38): I'm cheating. No one even thinks of them. If you want my honest answer, OpenAI is the most in trouble.

**Nicky Pike (01:02:44):** That's a whole other conversation. I actually agree with you on that one as well. The vibe coded prototype that made it to production, funniest disaster people have ever seen or is it going to be too painful to even laugh at?

**Rob Whiteley (01:02:53):** Funniest disaster. Look to me, you'll get a lot of business value so you have to be okay with the messy middle. But right now I wouldn't want to limit all the fun disasters that are going to occur and let them happen. Limit their blast radius.

**Nicky Pike (01:03:05):** Yep. Open weight models running on - prem. This a real strategy or is this just a phase in a fad?

**Rob Whiteley (01:03:11):** Oh, real strategy and I think it's coming this year. I mean, we already see it in intelligence agencies and it's only a matter of time before I think large enterprises can do it economically.

**Nicky Pike (01:03:21):** I believe the same thing. We're getting a lot of people asking about open weights and how they can bring that in. You're a brand new grad in 2026. Do you still learn to code by hand or do agents make it not even worth bothering?

**Rob Whiteley (01:03:33):** I think it'll be important to take a comp sci class to have an apreciation for the engineering principles. I would honestly learn softer skills. I think all ICs become managers. You need to start learning to think around how do I keep my fleet of agents happy?

**Nicky Pike (01:03:48):** All right. Last rapid fire. The company that ignores its token bill is what? Fill in the blank.

**Rob Whiteley (01:03:54):** Going to have no CEO because the board will fire them.

**Nicky Pike (01:03:59):** I think we're starting to... I don't know. Real quick on the same rapid fire, the companies that are now doing the token spend leaderboards, what are your thoughts on those leaderboards?

**Rob Whiteley (01:04:08):** Oh, it depends. If it's there to inform and observe, I'm okay with it. If it's there to stack rank and judge and do performance, very, very bad. We're too early. I'm a big fan of you can't manage what you can't measure. So I think you should be measuring all of this stuff. But I think even at Coder, we're at a point where for six months we just wanted to see what the heck did people do with AI if we gave them completely and totally unfettered access. And so I think for us, you have to allow the experiment to unfold and then put rational controls in place. So to me, a lot of the controls are actually just for observing. It's just to see where the business value comes from.

**Nicky Pike (01:04:48):** Gotcha.

**Rob Whiteley (01:04:49):** Sorry, that wasn't very rapid.

**Nicky Pike (01:04:50):** No, that was a great answer by the way, because that one brings up a lot of visceral reactions with different guests as well. All right, prediction time, put a number on this one. You've called parallelization and maintaining agent state one of the top five enterprise problems of 2026. And you said that AI native companies will out - compete the market the way that cloud native ones did. So Rob, we're halfway through the year, bud. Where does this problem rank by December? And with the macros that we mentioned, the IPOs, the mandates, the things of that nature, who's the category company that bets wrong on the model layer and doesn't see it coming?

**Rob Whiteley (01:05:25):** The companies that are in the most trouble are the big hyperscalers. There's Google, Microsoft and AWS. And I think if you look at what Google's done, even though their AI strategy is bananas because it's so fractured, they're at least have all the raw ingredients to do something. I think Microsoft made too hard a choice upfront on which AI model and AWS isn't making a hard enough choice and they're just sort of sitting on the sidelines. So for me, I worry that one of the three cloud providers isn't going to go under by any means, but they're going to get usurped by one of these frontier model providers that are essentially becoming infrastructure providers anyway. And so I think for the first time we're seeing that kind of oligarchy of companies get seriously challenged. I expect some real cracks in that armor by the end of the year, but I think my real winner is going to be Apple.

I think they're the ones who are going to come out on top because whether you believe it's just marketing or not, they've leaned hard into privacy, which is sovereignty, which is on device models. They're partnering with Google. I think they're going to be the people that really bring AI to the masses and that's going to really kickstart the next wave of demand for AI in the business.

**Nicky Pike (01:06:52):** Yep. See, I didn't even think about Apple. I am thinking more along the lines of strange bedfellows. And what got me started was the whole SpaceX, IPO, their orbiting data centers concept, their bid to partner with Anthropic. I mean, we don't know where the next one's going to come from because we are seeing some strange combinations out of this.

**Rob Whiteley (01:07:12):** I completely agree. And yeah, it's going to be crazy, especially when we are supply chain constrained. And so eventually just sitting on assets becomes a competitive advantage.

**Nicky Pike (01:07:23):** Yep. Well, Rob, I mean, you came up as a developer where the person was the one writing the code. You spent the whole hour describing what wave two is going to look like where the building gets handed off to agents and the human becomes the one architecting it. We've talked about pricing, steering agents, checking it. Coder sells the tools to the world that is betting on how this is all going to shake out, that infrastructure is the core aspect of what we see with AI. So after all of it, the token bills, the cloud native parallels that we talked about, the choices, freedom bets and the shovels and the excavators, what does it mean to Rob Whiteley to be a coder?

**Rob Whiteley (01:07:59):** I think to me, we are literally in a moment where the way software is written is changing. And to be a part of that, however small, is exciting. Very few times in your career do you get to be front row to a major transformation. So to me, that's what it is. We're changing the way software gets built, our small part of that, and that's exciting. I think the era of vibe coding will be amazing. And trust me, one day we'll go back to artisanal handcrafted code that was written on a paper, just like we do that in all industries when you eventually get over processed. So it's a pendulum, but right now it's swinging hard towards the disruptive side.

**Nicky Pike (01:08:42):** I can't wait for the era of artisanal coding. That should be fun to watch. And see, and I don't even know how to end this show this time, Rob, because I usually ask people, can we consider you a member of the \[Dev\]olution? But hell man, you're the one financing this whole thing. So I don't even know how to ask this question, but I'm going to. I'm going to assume that we can consider you a full foot. Well, hell, you're one of the leaders of the \[Dev\]olution.

**Rob Whiteley (01:09:06):** I'm honored. I'm a litle offended. It took you so long to invite me back because this thing's blowing up. So I'm hoping I can stay a card carrying \[Dev\]olutionist and that you don't outgrow me.

**Nicky Pike (01:09:17):** Oh no, I don't think that's the case. But yeah, for the audience, we took this long. One of the reasons is because if you ever get a chance to meet with Rob, he is one of the funnest people to sit down and nerd out with his stuff on. He can explain these things in ways that make sense to me and as well as others. So if you ever get the chance, that's why this episode went long because it's just so damn fun to talk to you, buy.

**Rob Whiteley (01:09:41):** Likewise, Nicky, always a pleasure.

**Nicky Pike (01:09:43):** All right. Well, thank you very much.

**Rob Whiteley (01:09:44):** All right. Thanks everyone.

**Nicky Pike (01:09:47):** Thank you for listening to \[Dev\]olution. If you've got something for us to decode, let me know. You can message me, Nicky Pike on LinkedIn or join our Discord community and drop it there. And seriously, don't forget to subscribe. You do not want to miss what's next.

Show more

## Featured speakers

![CEO, Coder](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1694538377-rob-squared.png%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D400%26w%3D400&w=2048&q=75)

Rob Whiteley

CEO, Coder

![Field CTO, Coder](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1746230577-nicky-pike.jpg%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D400%26w%3D400&w=2048&q=75)

Nicky Pike

Field CTO, Coder

## Watch more

[![ 96% of Companies Run AI Agents. Only 21% Can Control Them](https://coder.com/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FqTRwwcx4QQI%2Fhqdefault.jpg&w=2048&q=75)](https://coder.com/podcasts/devolution/96percent-of-companies-run-ai-agents-only-21percent-can-control-them)[![ Open Weight AI Isn't the Risk You Think It Is](https://coder.com/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2FrFFbzoKGMtc%2Fhqdefault.jpg&w=2048&q=75)](https://coder.com/podcasts/devolution/open-weight-ai-isnt-the-risk-you-think-it-is)[![Why Developers Are Ditching AI and Talking to Humans Again](https://coder.com/_next/image?url=https%3A%2F%2Fi.ytimg.com%2Fvi%2F-DYZyPLovjc%2Fhqdefault.jpg&w=2048&q=75)](https://coder.com/podcasts/devolution/why-developers-are-ditching-ai-and-talking-to-humans-again)

## Additional content

Read more in these blog posts.

## Additional content

Read more in these blog posts.

[![A blog post thumbnail image with the title and small purple icon](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1784830943-blog_when-anyone-can-build.png%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D324%26w%3D604&w=2048&q=75)
AI • Jul 23 2026 • 3 min read

### When Anyone Can Build: Gene Kim on What Vibe Coding Breaks First

![Nicky Pike](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1748473544-nicky-pike-linkedin-image.jpeg%3Ffit%3Dcrop%26fm%3Djpg%26h%3D100%26w%3D100&w=2048&q=75)Nicky Pike

Read more](https://coder.com/blog/when-anyone-can-build-gene-kim-on-what-vibe-coding-breaks-first)[![A blog post thumbnail with the title over a light background](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1782244428-blog_five-months-to-patch-one-day-to-weaponize.png%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D324%26w%3D604&w=2048&q=75)
Jun 23 2026 • 3 min read

### Five Months to Patch, One Day to Weaponize: CVE Remediation Has a Math Problem

![Nicky Pike](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1748473544-nicky-pike-linkedin-image.jpeg%3Ffit%3Dcrop%26fm%3Djpg%26h%3D100%26w%3D100&w=2048&q=75)Nicky Pike

Read more](https://coder.com/blog/cve-remediation-dashaun-carter)[![Blog thumbnail image with the title over a dark background with a purple icon.](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1781801403-blog_everyone-said-ai-would-kill-developer-community.png%3Ffit%3Dcrop%26fm%3Dwebp%26h%3D324%26w%3D604&w=2048&q=75)
AI • Jun 18 2026 • 4 min read

### Everyone Said AI Would Kill Developer Community. The Opposite is Happening.

![Nicky Pike](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1748473544-nicky-pike-linkedin-image.jpeg%3Ffit%3Dcrop%26fm%3Djpg%26h%3D100%26w%3D100&w=2048&q=75)Nicky Pike

Read more](https://coder.com/blog/ai-and-developer-community-pauline-narvas)

## Subscribe

[YouTube](https://www.youtube.com/watch?v=lx7fAw1EBAA)[Spotify](https://open.spotify.com/episode/0qOhnbjHvMC1VeU9blszS4)[Apple Podcasts](https://podcasts.apple.com/td/podcast/if-you-wake-up-hoping-the-model-got-dumber-get-out/id1839475248?i=1000780142780)