Picture a dam built for a calm river. Now the river's running way faster than anyone planned for. And the water is pouring right over the top. That's what AI-generated code is doing to everything after the pull request. The reviews and approvals downstream were built for people working eight hour days.
My guest today says AI feels like alien technology that dropped from the sky and we don't know how to control it yet. But his answer isn't less automation, it's more. And he thinks that we should onboard an AI agent the same way we'd onboard a new engineer. I'm Brian Teller from Teller's Tech, and this is Ship It Weekly.
Welcome back to Ship It Weekly, where I filter the noise and focus on what matters when you are the one running infrastructure and owning reliability. Most weeks, it's a quick DevOps, SRE, platform, cloud, and security news recap. In between those, I do conversation episodes with people building and operating these systems. Ship It Weekly turns one this November.
To celebrate, through mid-December, you're getting two conversations a week, Mondays and Wednesdays, with the news every Friday. Then, we're back to our usual rhythm. Today, I'm joined by Darko Fabijan, CEO of SuperPlane and co-founder of Semaphore, the CI/CD platform.
We talk about what starts breaking after the pull request when AI is writing more of the code, what should sit between an AI agent and production, why he thinks you should onboard an agent the way you'd onboard a new engineer, and the first workflow he'd actually trust AI with on an ops team. All right. Let's jump in. Darko, thank you for joining me. Thank you for having me.
Developers are using AI to write more code faster. What starts breaking once that code hits the real software delivery and operations workflow? Well, yeah, I think that we are very fast with generating code. I mean, some are also very fast at putting it into production, but there are just some numbers that are not as they used to be.
So I think definitely it's a time of, you know, stability and, you know, various other metrics when it comes to the quality of, you know, what is more hitting the developers back. But, you know, for the operations people, it's just that code that we are shipping, you know, it just has more bugs and it's just by nature of things. Pull requests are bigger. They are more frequent.
The pressure is all over the place just within the industry of like making software. So there are like a couple of forces that are, you know, just very new and things are breaking in various shapes and forms. Yeah, I would not say only in the production and operation, but kind of also in the society a little bit. It's just the velocity and the amount of things that we're putting out there.
Is the post-PR world becoming the harder part of engineering? Or you had mentioned the social part as well, but what is becoming harder specifically from the changes with AI? I would say, I mean, operation side of things is becoming harder. So if I'm just thinking from these numbers perspective, so the complexity in the software is increasing just by the numbers of lines of code that are being put out there.
The frequency of those deliveries is increasing. I mean, in the good old, you know, CI, CD practice, continuous integration as a practice is you want to ship, you know, small chunks of changes because then you minimize, you know, the blast radius, you minimize the surface area. Now the pull requests are just, you know, bigger.
It's getting kind of out of hand, you know, what's the size of the thing that we are shipping. I mean, I'm seeing it in, you know, in our team and I'm hearing it all over the place. You know, those pull requests of a couple of thousands of lines of code, you know, who can manage that.
And I think that, I mean, it's not, I think it is getting out of, you know, cognitive capacity of people to just want to review all those pull requests. And it's just that, you know, things are getting into production that are not really that well reviewed and are causing the issues on the other side. I don't know the exact URL, obviously, by heart, but there is a Faros.ai report that recently came out.
And I think we can link it probably somewhere. There is an increase in the number of defects that are being shipped, in the number of incidents. So it's all between 50% and 200%. That's my kind of rough understanding and memory of how those charts look. And it's all in red. So yeah, that's one side of it. I've seen this in my day job as well.
Like we have AI agents creating code using Claude Code or Cursor or whatever. And then you have AI agents reviewing code. Like you have like a Codex plugin on maybe GitHub that reviews the PR. And then you have maybe even AI in your CI/CD pipeline, like in your actual delivery pipelines as well, where it's interfacing with the production. Where should AI agents not exist?
Like where should we not have AI or still have like very specific? Because again, it's non-deterministic, right? So at a certain point, how can we reduce that risk? Well, one way we can reduce it is by having more AI. But it takes time to figure out how exactly and where to put that AI. So in terms of unsolved problems, it's definitely the right access to production systems. So people are not really into that.
There are some people that are. I think it's, you know, sub 1% of the industry at best, but people are trying to solve that. And we are also trying to solve that also in our own ways and industry as a whole. In the broadest possible sense, I would say I understand the situation.
Yes, AI exists for the last couple of years in this form that is, you know, usable, but in just like last, you know, six to nine or something like that, it has really become, you know, super powerful, you know. And I perceive it as that, you know, alien technology has dropped from the sky and we don't know how to control it.
Part of it is that in our process of producing software, there are still so many things that are manual in various ways. I'm not saying that we are creating EC2 instances by hand from console and all of that, but the processes are just not made for the AI. Processes are made for humans. So there is, you know...
The time that the human being, you know, can, you know, review something, understand something, you know, work those eight hours a day or, you know, a little bit less, more on the day and so on. So that is a very recent phenomenon of us being to create so much code at such a velocity. But other parts of the system are just not made for the AI yet. And that has to change, you know, iteratively over time.
In the same way that we are introducing change in the production systems, we want to make one change at a time and control the risk. And that whole pipeline has to change fundamentally, but it cannot change overnight. We have to figure out and understand how to change it so that velocity that we got to that AI creation side.
So yes, I can write like 10,000 lines of code in three hours of live coding, but how the rest of the system... Can really deliver it and work and how we can handle that amount of change. So that has not been figured out.
So kind of our main hypothesis is that there needs to be more automation and we have to figure out what that automation is and what's the nature of it to be able to really fully harness and control, I would say, mainly control the risk that this power that we just recently got has introduced. So I would say that's... Part of the, that's a part of the, of the answer.
I mean, definitely, other, the other piece that I mentioned is how do we give agents the right access to the production systems. So policy enforcement, that sort of thing. Yeah, or least privilege. Yeah. We need to figure out like what that kind of automation is and how that layer works. Is that like a fully deterministic layer that exists between, you know, AI agent and the production systems?
I currently believe that that's a part of the equation and then it's going to be part of the equation for, you know, significant period of time. There are some people that we are operating against, like, that hypothesis.
There has to be a deterministic layer in between and there are automations and right access that we give to the to production system to agents, but it's very scoped, very limited, goes through the approval process. But as I'm speaking with, you know, various people from industries and some, you know, very, you know, advanced companies recently spoke with a guy from Nvidia.
So they are trying to figure out how to give more, you know, right production to their systems, you know, in their own way. And that's also our hypothesis that there has to be that guardrail, you know, layer in between. To be able to do that. Yeah. I'm curious, what are your thoughts on that right access to the production system? Because it is a very uncomfortable thought to a lot of people. Yeah.
So, well, OPA policies, Sentinel policies, as far as like allowing changes and then on the actual production side change, having approval gates, like having least privilege and maybe gating those AI agents, I guess would be my answer. I don't know that it's... That's a full answer, though. A lot of the AI safety advice says keep a human in the loop.
But what I find is in a lot of real ops, humans end up just rubber stamping stuff all the time anyway. So that part, I don't know how to fix. How do you avoid those fake approvals where it's just rubber stamping the AI changes? That's a harder challenge that I'm not sure. And I think if you can solve that, it makes way to the least privileged conversation, the OPA policy conversation is an easier gating strategy.
The harder part, I think, is the human side. Like, how do you fix that human side? I mean, I recently heard that, yes, AI agents hallucinate and are unreliable, but also humans are hallucinating and are unreliable. So that side of the thing. I mean, whenever there is a system that has a lot of something in it, you have to... Quantify various kinds of things.
So there is like a code coverage, let's say, if you're developing something, is everything covered by tests? What's the quality of that code? There are like various metrics to assess that. And there's, you know, change, any change, you know, pull request, either your Terraform code or on your, you know, Java code, it doesn't matter.
There are various, you know, heuristics, numbers that can quantify how dangerous is that change. So as we are shipping, you know, this code, I mean, last week, I mean, we, we shipped kind of 50, 50, you know, thousand lines of code, you know, and it's, you know, five, six people. Okay, we are rushing and, you know, it is, you know, kind of pre-production software and, you know, all of that.
But it's, it's such a, such a big, you know, such a big change that we are introducing the system that I was thinking in terms of just pull requests, so can we have some kind of like metric in place? So if this change is just changing maybe the, you know, but name of the bucket, you know, if it's like, you know, Terraform or, you know, changing some, some smaller thing, you know, it should get the score.
So it's not very dangerous. If I'm deleting the resource which is a data storage thing, either S3 bucket, RDS instance, I mean, that's 10 out of 10. That, that's a red. I have not, maybe I don't know, but I haven't seen that kind of like metric that exists yet in the industry either on the development side or the ops side. I think that we need that kind of thing because of the rubber stamping that you're mentioning.
Sooner or later, someone is going to be tired and there is going to be a line which is going to be like drop all tables because, I mean, we have seen that recently, last couple of months. There have been some drop all tables scenarios around. So, yeah. Yeah, I was trying to think of that because, yeah, dropping an instance or deleting something from a bucket or removing a database, that's an easy... Gate to have, and that's an easy score to score on a metric.
The problem I would think is with ancillary processes, like a Lambda that maybe does some data manipulation that could potentially also, maybe there's an unintended change and it's down the road and you're deleting objects, you know, and that's not, it's harder to understand, I think, to actually quality gate that because there could be related resources with unintended consequences further down the chain.
Absolutely. Absolutely. But I mean, that's an area where theoretically, I mean, we spoke about AI writing code and reviewing it, then, you know, approving it. Yeah, no, I know. Yeah, that's a hard problem. Yeah, it's not solved either. I would say just the area of like qualifying the risk. That's a billion dollar startup idea over there. And I mean, it's just, there's a lot of work to be done in the area.
I guess there are some, you know, research going around some universities around qualifying that. So you have all of these different services that are part of your control plane, and you have scripts that are all glued together.
Most ops teams already have some version of this held together with GitHub Actions and a Slack app and maybe Argo CD and Argo Workflows and scripts and runbooks, maybe Ansible, whatever, dashboards, and probably in someone's memory somewhere. What's wrong with that model and how do you solve it? I don't think there is like anything particularly wrong with that model.
It's just the question of like how AI can interface with that in the best way possible. So part of our hypothesis that, you know, there needs to be like a bigger context for AI to be able to understand the system. And a lot of what you mentioned ends up being, you know, what I like to call those point solutions. So there are things which are in PagerDuty. There are things within your GitHub Actions.
There are some things in your repo. Some things are in CloudWatch or Datadog. So things are all over the place. And I think it's a situation for the vast majority of the companies. Yes, there are companies that are fully AWS. Nothing goes out. I think in that situation, it's a little bit easier. It's about giving that context and figuring out, you know, what should go in that context so decision can be made.
So I would say that's one piece of the puzzle. There is also a question of coordinating and how AI can coordinate those things and, you know, stitch them together. And when we get to that question, we have to get into that question mark of which API actions can AI call and which API actions AI should not call. That can be solved through, you know, policies, you know, various guardrails, approvals and all of that.
But as our industry, it's not science. It's not art. It's between art and science. And we have been running in loops in our industry for, you know, many times and, you know, creating things over and over, like every decade where we create things. So I would say... We are now at a stage where everything is getting recreated.
Each service is going to be recreated from scratch to maybe suit the agents a little bit better. And it's about, I would say, scale and velocity. So historically, everything was made for humans. I mean, each service that you can imagine, it's made for humans, including programming languages and so on. So now the more the conversation is, okay, all of those services need to speak with AI.
Some need to be rebuilt from the ground up. It's great for the new generation of builders. Everything that my father and grandfather built, let me rebuild it from scratch. Some for very good reasons, some for not. And then in that rebuilding, the question is, you know, how do you treat actually human? How do you treat agent? Are they on the same level? Are they somehow different?
So it's about, I would say, becoming the reality in every part of the system. How is SuperPlane different from Argo Workflows or n8n? I mean, our main hypothesis, let's say, the space that we operate in is essentially software engineering operations, both of those domains. And we specifically want to treat, we want to start with that, that both agents and humans should be kind of on the same on the same plane.
They're not today. It will take time until that becomes the reality. But where we are building service, we are building it from the ground up from that perspective. The other element is how do we give AI right access to the production, which is a big unsolved question within the industry. And our solution here is that we need more automation. Automation in that very, I would say, traditional sense.
But we need to expose that in a very controlled way. So I like to think about that story as we are onboarding, let's say, the DevOps engineer within the company, or we are onboarding the agent. So what are the things that are available to me? Where can I apply my Terraform change? What can I do? What can I introspect? So I like to think about the agent as something that has the identity. It's entering the system.
Okay, let me do the introspection. Let me see what are all the things that are available to me so I can be productive. And that onboarding piece is an angle how we are telling our story. In the same way that we want to make super easy for DevOps human to onboard the production system as we understand everything, and that should be there is no half-rotten documentation, tribal knowledge, all of that.
It's all codified. It works, flows. I can click on everything. I can understand how things are flowing, how things are being done. And that hit by that bus factor is removed. So it's the source of truth is, you know, there. Once we reach that level, and that's going to take time for the industry to, you know, just introduce more automation.
I would say at that point in time, it would be easier to onboard also like a dev, you know, DevOps AI engineer. Like in a way to be able to help with things. And I generally don't go for that, you know, doomer scenario where, you know, us engineers and, you know, crafters are going to be removed from the equation.
I mean, it's obvious that things are going to change, but, you know, taste, direction, you know, all of that is going to be present. So I'm more in that scenario that things that we have not historically, we didn't have time to solve. Because it was too expensive. It was too expensive to write software.
So I'm more in that category that, you know, we will need more engineers on all sides because there will be just more software. Yes, the roles will change and some will change drastically. And we'll have to figure out how that works. So when you say more AI and onboarding, are you talking about like a multi-agentic workflow? Like where maybe I have like an agent for just Terraform or just IaC in general?
And like that agent only handles that? Or is it even more fractured than that? Like an agent that only handles like creation of objects or resources? And how do you think about that model? Like, how do you frame that? My hypothesis is that it's not going to be like that. It's going to be within this team that is handling maybe... Encoding and stream in Riverside.
I don't know how they're organized, but let's say they have a department which is more for this, you know, video transformation, production, whatever team, there they will have an agent who will have a specialized set of skills and understanding. So it will know that there is like a throughput and latency are very important. So maybe I'm not, I don't even write, you know, Terraform. I don't do this.
I have some specialized functions for my team and the context is very based on that. And I'll have that maybe one monitoring agent, which is really in charge of, you know, maybe scaling and latency, understanding what's happening in the world, you know, within the AWS network. And there is a glitch and then what I can do. My sense is we're going to be very specialized and small and scoped.
And that agent is going to take care of that. Of course, it does make sense that there are agents and there will be agents that will be very, very specialized in Terraform, OpenTofu, whatever, resource creation. But I'm more on this side. How do we fit? How do we find help? How do we find that small agent?
Not replacing anyone, helping someone with a very specific task, which is maybe more suited for AI because there are a lot of numbers and metrics and less for human. Yeah. Do you think that compliance and audit requirements change when part of the workflow came from an AI agent? I think that they are going to change. I think the change is going to come by this, you know, velocity, which is coming in.
So whenever I think about this, you know, code creation in AI, I imagine like a dam and a river. And like river is so strong and it's, you know, overflowing the dam and, you know, you have to do something different. So I think that, you know, compliance piece will have to change and we have to figure out how that will work.
At some point, agent is probably going to have to carry, you know, some element of responsibility in some shape and form. No clue how is that going to work. I would say... Maybe this, you know, self-driving cars industry will maybe, you know, teach us something. So if I get killed on the streets of San Francisco by Waymo, who is responsible? I know that they have some framework.
They figure it out, you know, one way or another. So if like a database, you know, a database table or, you know, whatever gets dropped or, you know, some data corrupted or leak or whatever, whatever other way. There is going to be some legal framework around that. I wish I have an answer, so I don't. I'm just, you know, thinking out loud. I think it touches the law and those areas also in various ways.
What's the framework there? If something really bad happens in a publicly traded company, they have various policies in place. I would say that's the domain where probably the biggest guys should solve it and let us know what to do. So I would say, you know, Google, Apple, whoever has the most AI, the biggest production, all of that, and the rest of the industry would follow us.
So I would say it's not my job and of many other folks. But it is an interesting topic for sure. And one thing that feels very real is how much production knowledge lives in people's heads. How should teams capture operational memory without pretending that documentation will stay updated? How do we solve that problem? Again, getting back to our hypothesis at SuperPlane that we live by, I think it's about automation.
So as things get more and more done through the automation, the more records do we have about what happened and when and why and whose approval and whose idea and which prompt and so on. So all that's tied into a chain of a decision chain, how we came to that. So that's one piece of the puzzle for sure. How in practice that gets pulled off, we'll see.
I mean, in terms of documentation, I think that all documentation should be, kind of AI-generated by those by those automations that we have, you know, in place, again, iteratively getting that piece by piece. But I think that's the main piece. That's the main piece of the puzzle of that.
Yes, AI definitely is better at maintaining documentation over practitioners, just given that, you know, practitioners can get lazy, right with comments, or just markdown files and AI will be much more verbose, but not overly verbose, like concise in its verboseness. I agree with that. It's also about generating the documentation.
So I think it will be each paragraph sentence, I think should be clickable in that documentation with like some decision tree. How do you came up with that conclusion? Because if I perform some action based on like incorrect documentation, I do something bad. I need to have that chain. Okay, you inform. You gave me something, you know, some misinformation based on some, you know, wrong reasoning.
And then we should be able to have like root cause analysis of what caused that. So in my head, I mean, we are not there yet, but we have some elements in our documentation generation where, you know, you have a sentence or you have a certain action and you click on it and it links back to some part of the workflow and you see, okay, how that, you know, documentation came came about.
So, and I keep going and kind of maybe defending not knowing the answer and saying about that has to be done, you know, iteratively. So the whole documentation generation problem.
I think we'll have to, we'll have to get into that check that we should know how each line of that documentation was generated and, you know, it traced back to some to some code or some fact some configuration somewhere And then how do we keep that in sync? Is that eventual consistency or whenever I change some configuration, my load balancer, is that immediately reflected on the documentation?
What's the gap, you know, between that? Again, one of the areas that deserves a special kind of attention, especially because of that, as you said, no one likes doing that. So our industry is awful, you know, when it comes to having like up-to-date documentation. Yep, for sure.
Every company I've worked at, you go on Confluence and there is, you know, a document that hasn't been updated in three years and the process is not even the same anymore. You know, that's the only document that exists. Or you're trying to hunt through Slack to find, you know, a conversation that engineers have had. And that's a real problem, I think, yeah, that is just not solved at a lot of companies.
And AI definitely can help with that 100%. Okay, so wrapping up, if a platform or SRE team wants to start using AI in infrastructure operations, What's the first workflow that you would trust it with? Like, where would you start? I mean, it's definitely this, you know, read-only mode where I would want it to explain me how something is done and point me in the right direction.
So if I'm onboarding like a brownfield project, let's say on AWS, it has some complexity in terms of the, you know, software stack, infrastructure stack, all of that. I'm generally, what I really like is that, you know, I pretend to be super, super stupid, you know, and just a lot of dumb questions and like onboard me, help me understand how system works.
And then I ask, okay, what's the smallest possible change that I can do on the system to just help me get, you know, some deeper understanding. I think that I could be, I could misinterpreted your question. So feel free to, you know, dig deeper. But I mean, for me personally, that's, you know, discovery phase is like what I would do as a first step. Yeah. Okay, so let's say we've done discovery phase.
We've done read only. It has the context of our entire IaC. It knows our deployment stack. It has access to GitHub, so it knows all of our GitHub Actions that we use for our CI/CD platform. It maybe even has access to Grafana and Prometheus endpoints, so it's getting monitoring. All read only. What would be the first write operation?
Where would you feel the safest starting giving that autonomy to an AI agent to start writing? I mean, I would say we have two big categories here. One is the creation of, you know, just the change, which is, again, writing some code. So it's either Terraform or, you know, Java code within our application. That's one domain that goes through a review process. And then that's, you know, being applied in production.
I would say that's the category where you have human in the loop approval process, which is lowering the risk there to some extent. And that's what happened in the industry at scale already. So, I mean, I haven't said anything new. I just kind of described the reality that we have today. I would say maybe the first next step after that are... I would say in that category of these loops that we keep mentioning.
So if we want to move some metric, you know, want to lower the latency of a certain system, then I would say I would work on those numerical things that are like easy to measure and easy to test. So yes, we have some response time somewhere. It can be due to the way the code is written or how infrastructure is configured. And maybe we want to have like a faster disk, faster database, whatever.
And I would say it's in that performance optimization area. What are some of the incremental steps that we can make? And then we apply that to copy of production or staging environment, whatever. Do some testing and apply that change in production. In the same way, I would map it to how can I improve the speed of my GitHub action, which is, you know.
Running my code and I would, it is also how can I lower maybe the number of CVEs like in my system, there are auto security tools that are minimizing that. So I would say I would keep it in that realm where we can easily measure the impact and it's a numerical value. And we, we, we, we go, we go from there. Yeah. Yeah. It makes sense. Any closing thoughts you'd like to leave our audience with?
I would say I would get back to the automation and I would say we as an industry and discipline, hardware and software together, we have been going out and changing factories and many, many other industries for many decades and automating things, automating things, you know, factories, removing workers, like all those kinds of bad things.
Now the lens has kind of, you know, shifted in the other direction and we are getting automated and, you know, also feel threatened a little bit and all those emotions that we're feeling. I think that's okay and everything is going to be okay. We just have to embrace that automation and go after it as soon as possible and just embrace the change that are happening to our roles as that is happening.
So with any technology, historically, you can fight it or you can embrace it. So I would say let's just embrace it. You know, bring that, bring that automation and, you know, figure out there is going to be a lot of, you know, a lot of infrastructure, a lot of everything. And yeah, those are my thoughts. So more automation. More automation. No, I got you. Where can people follow you? SuperPlane, Semaphore?
Yeah, superplane.com. That's, that's our thing that we, that we are doing and trying to solve that, you know, write access to production. Easy problem. Easy problem to solve. But yeah, we are doing it iteratively in that, you know, a lot of determinism in place. So we want to be cautious and, you know, do good things there. And semaphore.io, previous company in the CI/CD domain and DarkoFabijan at X. Yeah.
So feel free to engage. A lot of, you know, hard topics to discuss. We love them. And I'll put links for all those in the show notes as well. Darko, thank you so much for coming on. Really, really appreciate it. That was my conversation with Darko Fabijan from SuperPlane. The biggest takeaway for me is that AI didn't break software delivery. It exposed how much of it still runs at human speed.
Darko's dam and river image nailed it. AI-written code is the river and the dam is every review and approval we built for people. I also liked his take on rubber stamping. He pointed out humans hallucinate too. And sooner or later, a tired reviewer approves the line that drops all the tables. So he wants a risk score on every change. Renaming a bucket scores low.
Deleting an S3 bucket or an RDS instance is a 10 out of 10. He called that a billion dollar startup idea. And I'd buy it. His agents as new hires idea stuck with me too. Give the agent an identity, let it look around, and keep a deterministic layer between it and production. The work that makes an agent easy to onboard helps your humans too. No half-rotten docs, no tribal knowledge, no bus factor. Where to start?
Darko says read only. He likes to pretend that he's super stupid, ask a lot of dumb questions, and say, onboard me. When you do let it write, pick something you can measure with a number, like latency. You can find Darko on X, plus SuperPlane and his previous company Semaphore through the links in the show notes. Follow or subscribe to Ship It Weekly wherever you listen and find previous episodes at shipitweekly.fm. Thanks for listening, and I'll see you Friday for the news.
Scroll inside the box to read the full transcript, or expand for a larger view.