Platform Engineering Paves the Way From Vibe Code to Production
Platform engineering was designed for a world before agents. Now platform teams must serve human developers, citizen developers and autonomous agents through the same golden paths.
Two industry voices examine that shift with Alastair Cooke at KubeCon + CloudNativeCon in Salt Lake City. The panel features Thoughtworks’ Rickey Zachary and Weave Intelligence’s Luca Galante.
More Users, Same Principles
Galante says the core principles still apply. Treat the platform as a product and understand its users. What is changing is who those users are.
His research finds that application developers still make up most platform users. Agents and bots are a fast-growing share. Data engineers, ML engineers and business users are arriving too.
Zachary sees mature clients move beyond buying AI tools. Handing out coding assistant licenses without a platform pathway usually stalls adoption.
From Prototype to Product
Cooke warns that a vibe-coded app is a prototype, not a product. Piling technical debt on top of it is not a sustainable business.
Zachary wants platform teams to lean into citizen developers rather than act as the Department of No. Prototypes can validate business intent. The platform then provides the guardrails and the path to production.
Galante describes probabilistic, deterministic and hybrid path specifications. The hybrid path should still be mostly deterministic. Codified in the platform, it builds governance and security into every workflow.
Start With Pain and Business Value
Agents are reshaping platforms too, from headless front doors to pipelines generated for each change. Galante says few teams are that far along. Most begin with toil, such as writing YAML.
Both panelists advise CIOs to start with use cases that have measurable outcomes. Token spend does not equal performance. Galante notes that companies using the same models see very different returns. The difference is the maturity of the platform harness around them.
For Zachary, that makes platform engineering more business focused than ever. Platform teams need to tie every AI pathway to real ROI.
Transcript
I'm Alistair Cook. I'm a event lead for Tech Field Day. So we'll be making a bit of video and streaming that this week from KubeCon.
It's a delight to be back here, and I have some of my delegates out here in the audience. I'm also a research analyst with the Futurum Group. So, those are two of my split personalities.
I'm sure there are more of them hiding somewhere. Joining me for the panel, Ricky Zachary from Thoughtworks. Would you introduce yourself?
Yeah. Hello, everyone. Been with Thoughtworks for about six years now.
Lots of different roles. Very difficult to describe what I'm doing at Thoughtworks right now, but I'm somewhere in the space of tech advisory, also doing a lot of work on AI platform engineering, but more recently have been leading our agentic PDLC practice as well as our newly kind of formed forward deployed enabling engineering practice on the AI side. Yep, and I'm Luca Garante.
org community for a while. We recently also launched Weave Intelligence, so we have a small team of analysts, very, very far from the greatness of Futurum. But we're very focused on our little niche that is platform engineering, and hope that we can share some interesting numbers and from our latest research and from what we see on the field or from the community.
And the basic premise that we want to work with is the idea that platform engineering's been around for a while. It's in some circles extremely well understood. Other places, maybe not so much.
There's been some challenges on adoption. " But there's some challenges, as we've seen with observability, there's some challenges as we start moving from deterministic applications towards AI-driven applications, and particularly agentic applications, that are really making platform engineering a bit more of a challenge. And I want Luca to start on this, because once we start, Ricky, he's not going to stop.
I'll take the first two minutes of this, and then you can take the other 28. No, but yeah, I think it was said on the previous panel, too, I think, a lot of the best practices, the principles of platform engineering still apply. For me, it's been tricky, I think, the last year or so because of course, the guy that's been saying, "Platform engineering, platform engineering" will tell you that platform engineering is a solution to AI.
But I think there is a lot of truth in that, if you want to move past this kind of cute, sort of demo stage, a lot of enterprise are familiar with across different business units, different teams. If you want something that's truly enterprise grade, production ready, and ready to scale, you need some sort of platform, right, to go from A to B. And so all the concepts, all the best practices that we've been talking about, right, like minimum viable platforms and platform as a product, and really making sure that you understand your users, your customers internally, all of those things apply.
I think one of the interesting things is they're evolving, of course. So for example, the number of users and the type of users of platforms is exploding. In one of our latest reports, the "State of AI and Platform Engineering Volume 2" that just recently came out, we're actually showing that something like 80% of users of the platforms are application developers, so the good old internal developer platform users.
9% of users. But also, we're seeing a lot of data engineers, ML engineers, as well as a lot of non-technical business users. What's now being starting to be referred to as citizen developers.
And the interesting thing is what we see on the field is all these requests, right, from everybody that's playing with agents and it's like, okay, well this is cute, but how do I actually create something that's prod ready that I can use in an enterprise setting? I need help for that. And the help request falls onto the shoulders of the platform team, right?
And so what we are seeing is the platform team starting to support more and more use cases, more and more users, as platforms become kind of the underlying substrate for the agentic enterprise. I'm going to echo that, Luca. Great.
So one of the things that I'm seeing anecdotally is our more mature clients that we're working with at Thoughtworks have moved from this idea that AI across the enterprise, AI adoption across the enterprise, can be solved with some specific tool, right? That was kind of our less mature clients. So if I go and I go talk to Client X and they are telling me, "Hey, we've just made a big purchase of Cloud Code," or, "We made a big purchase of GitHub Copilot and no one's using it," my initial thought is, well, how is your platform team involved within the adoption curve?
Right? Are you talking about the deterministic pathways that enable those developers to use it? And if those enterprises are coming back to me and saying, "No, we've got a ticketing system that we're doing and we're just provisioning out licenses," I immediately go, "Oh, well, you're not going to make it," right?
Your AI adoption is going to stall. So this idea of moving from, hey, we're just going to give a bunch of people some tools and that's going to solve our AI problem, that's going to solve our adoption problem, to kind of the classical, hey, how do I improve my engineering kind of productivity throughout the entire enterprise? That's going to have to be a systemic approach.
Something that looks like a platform that allows people to come to it, provide feedback, and then grow that over time. And you would have to have some sort of team to own, operate, and then improve that. We classically have talked about one of the big things at Thoughtworks is treating the platform like a product.
You have to evolve that to include your AI use cases, your AI pathways, both for people building on top of that, but then also talking about how do we deploy those things and apply the governance pieces to it. So it's a holistic portion of things that we have to do. And I think there's a real challenge in the fact that we are seeing the citizen developer, the vibe-coded marketing engineer who knows marketing, who can then actually build themselves an application.
And this kind of plays into that death of SaaS conversation that we were having a few months ago. That, yeah, you can vibe code something, you can build it, and it stands up now. But that's a prototype.
That's not a product. That's not a sustainable way of running a business. And there's a real challenge in the perception that actually all I need is to vibe code it and continue to vibe code it and continue to pile technical debt on technical debt and mortgage ourselves on the really inefficient ways that we end up with these applications.
I mean, we've seen these patterns. We've seen the projects that say, well, there's this vibe-coded thing. I'm going to handcraft something that uses a 10th of the resources to achieve the same outcome.
Getting that buy-in to put the effort in to do things well and to do them sustainably is a real challenge. Luca, before you go, I'm dealing with this right now with a client, so I'm going to jump in really quickly. That buy-in is important.
I think historically, as an engineering community, we've been very focused on, oh, those people are not doing things the right way, so we're going to be very hands-off. Am I holding the phone the wrong way? No, no.
We've been very hands-off on those types of prototypes and those things. I think what, as an engineering community and as a platform engineering community, there's two things that we have to do. Because ultimately, a CIO or a CEO or a CTO is going to say, "Well, if I can just have a marketing engineer just live code something out, that's going to be way more valuable for the business because I know it's going to get used.
" I think it's our job to lean into those use cases and say, "That is great to validate business intention. Let's use AI to rapidly identify what the value proposition is, identify what the measures of intent are, all of those things from an ROI perspective, and then we're going to provide the guardrails to get that into production. " So I think it's a team sport, to be a bit cliché, and we need to be leaning into those citizen developers way more and providing them with, "Hey, you can go prototype something, POC something.
" We don't have to restart that and lose that. That's, again, I'm an old school infrastructure guy more than a developer guy, and we were the Department of No who said, "You can't innovate. " And that's a big problem.
And Luka, you go. Yeah. No, and I think to your question earlier as well, this finding the right balance of this new probabilistic sort of where we're going with the good old deterministic/no department.
And this is actually something that we tried to unpack in one of the last reference architectures that we published where we kind of differentiate path specifications. You talk about the AI pathways. Path specifications between probabilistic, deterministic, and hybrid.
And probabilistic is the skills-based stuff that a lot of people are now familiar with. Deterministic is the good old CI/CD, whatever, different gates. And then the hybrid's a combination of those two things, which I think is also what a lot of people refer to as loop engineering in some cases, for example, because essentially you're looping between these two states.
That is ultimately what you want. What we're seeing is a high correlation with high-performing teams that are already trying to think about, okay, how do I combine these two things? Because to your point, if I only rely on probabilistic, I think everybody has had the experience of running a skill and then getting a different result every single time.
But then to your point, we still need to have a setup that embraces the new speed that people want to move at, that embraces non-technical users that have no idea what CI/CD means. And so how do you bridge the gap? And I think it is in this hybrid path specification that then have to be codified, to your point, as a standard in the platform so that it can drive governance security by design across the different workflows, across the different teams.
And we're at KubeCon, and I think this is an area where CNCF can be leveraged quite highly from my standpoint. What I see a lot of organizations doing is taking a very tools-based approach to that. They'll go and say-- I'll pick on them since they're not here and weren't in the last panel, but they'll go and say something like, "Oh, I have to go and use something like Chronosphere for observability, and that's going to solve all of my problems from an observability perspective.
Oh, I'm going to go and we're going to do Kubernetes because that's what we've traditionally done from an infrastructure perspective," instead of thinking about it on the use cases that I have to do and whether or not they're probabilistic. And I think the approach that we're looking at is that from a probabilistic standpoint, there's a set of things, capabilities that I need to ensure that those probabilistic use cases are achievable. And that we have the right telemetry to be able to validate those things.
" It's now widened the gap, so our platform teams need to be looking at those patterns and those reference architectures and saying, "Here's actual probabilistic, deterministic, and hybrid pathways that developers are going to use, citizen developers are going to use. " I think the deterministic is, yep, we're going to keep doing the same things that we were doing before. The probabilistic become more difficult because I think it's a layered approach.
There's a bunch of different techniques that we could use. Some things with Arize, Phoenix, I know Braintrust is out there, Langfuse, to continue to name tools and defeat the point that I was trying to make. But the capabilities that are there are really important.
If you're going to do evals, observability, telemetry, all of these things, you should have that tied back to, what am I actually trying to do, not, okay, I know I have to use this tool because it's on a CNCF reference architecture. Well, and I think that said, I think when we talk about this hybrid thing, I think maybe the first instinct is like, oh, so it's going to end up being mostly probabilistic with a little bit of deterministic around it. I actually think it's absolutely the other way around.
Yeah, I like that too. It's still 98% deterministic. In fact, there was I think an interesting paper from Anthropic where they were showing under the hood how a lot of the stuff is routed, and it's completely deterministic, a lot of it.
And so I still think that's the vast majority of the answer that we're looking for here. At least it's the starting point. It may become more probabilistic if the models improve, but that should be the starting point.
One of the things that strikes me is we're talking about generative AI being used as part of building applications and products, but our platform's a product. To what extent are we seeing people actually using generative AI in the process of building their internal developer platforms, continuing to operate them? The AI SRE was all over KubeCon last year.
I imagine it's probably still all here with my AI platform engineer. Me? Okay, good.
Okay. I think what we're seeing is this huge proliferation in the AIOps and AI SRE and AI platform space. I don't know if one of our Thoughtworkers talks was submitted, but one of the areas that we're seeing a number of talk about the agentification of SaaS applications.
We're seeing a lot of our clients essentially say, "I don't need to go and use a port or cortex or any of these backstage for my developer front door. " It was very similar to someone was mentioning on the last panel about dashboards being generated at the point of context. We're seeing that start to proliferate all parts of the platform engineering space.
So if it's, hey, I need to do a CI/CD pipeline, we're seeing some of our most advanced clients working with us on automatic generation of CI/CD pipelines at the point of an agent building code. Because now they can do very specific levels of verification, auditability, compliance, not on the entire code base, but on a specific change set. So we're starting to see that proliferate across-- I think it started in AIOps and in AI SRE.
We're starting to see that move further upstream, where more and more of the use cases are becoming dynamic, are becoming driven by data, are being driven by just-in-time use cases. I think that that's going to continue to proliferate. For whether or not it's good or bad, I think it's going to continue to proliferate, and we're going to see more and more platform teams, if they're more mature, start to be informed by product changes within the platform, by feedback from citizen developers, from feedback from traditional engineers, and from feedback from the agents, which is why I think that the prior conversation about observability within the agent is going to be extremely important.
That feedback has to have some mechanism to inform the product strategy of the platform to some degree. Yeah. I think that's true.
1% of teams are- I said it was a mature customer. I caveat-ed it And obviously, Thoughtworks works with ultimately mature organizations out there. Because obviously, I think the risk is you end up with this vibe coded apps running on top of a vibe coded platform, and it's just a compounding problem.
But I think also what's interesting on the role, what we are seeing is definitely organizations trying to figure out from even an org chart perspective, what does it mean, right? So we have this platform team that has this Cloud Ops, IT, cloud native kind of background. And then what we are seeing is a convergence of that with data and ML teams into what is being called in some cases, AI platform teams.
Everybody has a different name for it now, but roughly that's the umbrella is AI platform team. And I think that's also a really interesting trend to observe because you can see just organizations sort of scrambling to figure out, okay, what does it mean, and who do we put there? And I think we need to be careful that we don't fall into one of the anti-patterns that we've seen before, which is essentially taking a bunch of MLOps engineers that are mostly infrastructure focused.
And hey, now you're supposed to be building this platform as a product, but you've actually never built a product. And then you end up in the same problems that we had before with infra, DevOps, platforms, and so on. So I think that's just one thing that's worth observing.
But I totally agree with your vision. I think that would be ideal. I think right now the first order of business is making sure that we're very careful.
And I think this is the approach of most platform teams right now. " I think everybody's being very patient with it, yeah. Measured.
And in that measured approach, I think that they're trying to figure out what use cases to start off with. And primarily, I think it's been about toil, cognitive load. I have been an advocate for, I do not want to write any more YAML ever again in my entire life.
If I have to write another YAML file and have Visual Studio Code tell me that I missed some sort of indentation, it's going to be all over. " And once you start doing that a bit more and more, you're like, okay, here are other areas of toil, and then you naturally end up building out kind of this substrate that's like, hey, I've got a platform engineer agent or a DevOps agent or whatever it is within the organization that's there. And I will say that the operating model portion of that is something that I think most enterprises are struggling with.
We're in kind of this in-between transition phase where enterprises are still wanting to have one foot in the old model and another foot in this new model, and they don't know how to make that transition very cleanly. So we see a lot of organizational debt within the organization where it's like, well, we've got to use all these AI agents, and we'll just name them as a role within the organization, and they'll just be out there. And then you go and do kind of an assessment and you're like, well, you've got 500 agents that are just out here doing random things.
Have you figured out how you're going to kind of figure out what your operating model is for these things? And that naturally should get pushed to the platform team, right? They should have most of the governance and observability for that for an entire organization.
But you have to have that kind of organizational maturity or at least some direction organizationally to be able to figure that out. Well, and I think to your point is where do you start from, right? " Where do we begin with?
And I think the same old kind of best practices again that we've been talking about in platform engineering, which is to your point, look for the pain, right? We used to build this matrix, for example, for teams, right? It was like how much time does something take application developers and DevOps infra teams, and how often is that thing required?
Then you multiply that, and then you basically get a pain score at the end of that, right? And so you can apply the same exact framework to citizen developers, the new type of users, right? And I think that will give you a pretty good indication of okay, let's start there.
One of the things I love about vibe coded citizen developer, it's a horrible thing, but it has a wonderful part to it. It is that the actual business unit teams are starting to go through the process of how do I describe what I actually want and what the outcome of success is, what my requirement-- The things we always as dev teams, as infrastructure teams, is business tell me what the heck you want. And this is randomly, we have this mechanism where business units are choosing to use a tool that gets them to describe what they want, and that's a great on-ramp to then finding those points of pain in those places where we can make a difference.
Yeah, I think that's a really good point, sorry. Because I think good platforms at the end of the day become a discovery mechanisms, right? And so there's this, I think, powerful image that we always bring up, which is the typical college campus thing, which is funny because I always say that, but I've never been to a college campus actually because I'm not from here.
But where you have the paved paths, and then there's kind of the path that then people end up walking around, that becomes the main thing. And obviously, a good platform team and a good platform can become an automated discovery mechanism to actually detect that that's the path that's of least resistance, or at least where the users are telling you what they want and then go and pave that path, right? And so I think that's a really good way of thinking about it.
And to your point, Alistair, something that I think before was harder to do, and now with this, it becomes very obvious what people are trying to do, right? Yeah. I'm going to take an extremist position on this.
Sure. This is going to be, yeah, a shock. I actually have recently preferred working with the quote-unquote "citizen developer" more so than my normal dev teams, both internally within Thoughtworks and with our clients.
They have been the most excited to get their hands on these tools. They have been the most creative when using these tools, and it's an easier through line to actually solving real business problems. So we've been leaning really heavily into how do we get those vibe coded applications into production as quickly as possible.
That is the golden pathway. " And then we're going to codify that into a new platform golden path, right? Because that's an area where I've been talking about this for the past year.
Most of these edicts are coming down from using AI. It's a bunch of, and I don't want to disparage any CIOs or C-suite people, but I'm going to do so anyway. " No measurement, just an anecdote.
And they're like, "I got to bring that back to my company because I need 50% productivity. " And those kind of large-scale edicts don't really work out. Okay.
But I think under the umbrella of that, and I can't remember who said on a panel previously, but everything is AI. That's how you get the budget. You take a project and you say, "Hey, we're going to AI-ify the platform or Genfi the platform.
" And then it's your job to be able to go and identify the use cases there. So I think that that means the platform teams now need to be more ROI focused than they've ever been before, needs to be a bit more business focused than they ever been before, and they need to be leaning into, "Hey, I want to go talk to the business. " Excellent.
Well, we have a few more minutes left, so any questions from the floor? So on behalf of the underrepresented CIOs, what advice would you give a CIO? Which action should they take?
How would you advise them? Thankfully, I get to talk to a bunch of CIOs all the time, so I have some very specific advice for them. I have been telling them to really focus on two things.
No tools. What are the actual use cases that we want to actually affect? And it should be tied back to business value.
" Great. That's a great starting. That's a use case.
" It's like, well, I don't know how helpful that is to the business. A measurable outcome. A measurable outcome, use case with a measurable outcome.
And then the other part of that is, if they're not talking directly to the business, they should be. They should be talking to them about what is pain point, what do they think their future outlook looks like, and they should be able to tie their strategy directly to that. A lot of times, the technical strategy feels completely divorced from what the business is doing.
" Those feel like the same things, but they may be two sides of the same coin of it. Yeah, I think second that token spend does not equal performance. And I think what we're seeing is also this sort of decoupling of everybody has access to the same models, right?
Everybody has access to the same APIs, but you already start seeing this normal distribution of very different outcomes, right? In one of the last research pieces that we published, it was something like 60% don't really see much ROI, and then you have already a few, something like 20% see 2 to 3X ROI, and then 5X plus for the last 10, 15%. They all have the same models.
What changes the maturity of the platform, and really essentially the platform level harness that they're building around the models. " the number one starting point that I would look at. And I'm amazed that I didn't say harness at one time during that entire conversation.
Where's the bonus? Yeah. The bonus money that I would get is missing.
I will say, to just add one more thing, the token cost actually is something that CIOs should be getting ahead of now. They should have some sort of-- I saw Amazon just move away from their token leaderboard to something that's a bit more sustainable, versus just, "Hey, we're going to pay. " I think that this idea of building in some sort of token spend telemetry, tokenomics directly from the start and tying that to ROI also is actually pretty important.
So I would tell them to balance those three things, and then the parts that Luca was talking about. Well, I think I'm being given the hook here. I think Alan would like the stage back again.
He wants to prevent me from talking for another 30 minutes. He already knows that I'm already at 10. Yeah.
" Ricky Zachary, Luca Galante, thank you so much for joining me.