AI SRE: When AI-Written Code Outpaces Human Understanding
Five agents working on prod can't each keep their own memory. Mezmo CEO Tucker Callaway on curating shared context and runtime policy for production agents.
Tucker Callaway, CEO of Mezmo, joins Mike Vizard of Techstrong and Guy Currier of Futurum at KubeCon + CloudNativeCon to pin down what AI SRE actually means. Tucker frames it as production operations rather than a role: teams are shipping more AI-generated code than they understand, while AI still only suggests RCAs and every production change waits on a human who can follow it.
The conversation covers where governance stops (the deploy, after which an AI action like restarting a pod is ungoverned), why incidents get rarer but harder as AI clears the easy ones, runtime policy that lets an agent raise or lower its own autonomy based on system state, and why agent memory has to be curated once and shared across every agent working on production.
Mezmo is built for SRE and platform engineering teams who need AI agents running production operations without handing control to a black box. Tucker's closing argument: production is a multiplayer environment, so the teams running it need building blocks they assemble and own, not opaque AI SRE services.
0:00 What AI SRE actually means
1:18 Why production operations are multiplayer
3:03 Where governance stops: the deploy
4:40 Building blocks vs. opaque AI SRE services
5:11 Scaling production beyond human understanding
7:07 Quality, not quantity
7:47 Fewer incidents, but harder ones
9:11 Runtime policy for agent autonomy
9:27 Are engineers becoming agent managers?
10:33 Managing agent memory
11:18 Curating shared context across agents
#AISRE #SRE #KubeCon
Transcript
What AI SRE actually means
0:00 Mike Vizard: Welcome back to KubeCon + CloudNativeCon. We're playing a little bit of musical chairs, so let me welcome Tucker. How are you doing?
Tucker Callaway: Hey, great. Thanks for having me.
Mike Vizard: Welcome to the show. We're going to have a little chat about AI and DevOps and what's going on, because you may find it surprising, but there are probably, I don't know, 100 announcements related to something to do with DevOps and AI at this show. One of them I'll just call out: Cloudsmith does a thing where they're putting AI skills in a registry so that people can reuse them safely and not have all these things poisoned. And you guys have some stuff that you're working on in this space as well.
So there's this term now being kicked around, AI SRE, and nobody seems to know exactly what it means. Everybody kind of nods their head and says yes. But from your perspective, what's going on, and how does all this feel different from what we did before?
Tucker Callaway: Yeah, it's an interesting one. I think, actually, in the vendor landscape, we all don't like the term.
Guy Currier: Which part, AI or SRE?
Tucker Callaway: But I think we can't come up with a better one yet. The way we think about it is increasingly more about production operations, because an SRE is a role. We're really thinking about how we use AI to manage production operations autonomously.
Why production operations is multiplayer
1:18 Tucker Callaway: What's interesting about that is it relates to the skills registry, which we're a big fan of. The skills registry is really giving you more individual capabilities and helping improve the consistent state of individuals. But when you talk about the AI SRE concept, or production operations, it's really a multiplayer game. It's really more about how I take a group of individuals and allow them to work together in a multiplayer way against production.
That's super important, because we are shipping more and more code that we don't understand into production, but production is still gated by human understanding. So to me, the term is: how do we get prod caught up with dev, to a point where we can scale production beyond human understanding?
Guy Currier: I have felt, and I think I even used this word last week at AI Infrastructure Field Day, which the Futurum Group puts on, that AI, writ large, and the coding side in particular, is kind of a Mandelbrot set of generation, if you like. And this is an example of that to me, because there's, so to speak, the content of what's being generated by your AI system as an enterprise or organization. Then there is the management of it, with its elements of AI, to ensure that this code is inspected, that it's proper, and that it doesn't go into the library improperly. And then there's the next set, which is the happy home, the safe and secure home, which is where we start to touch on site reliability engineering and platform engineering. I feel like that just goes on and on and on.
Where governance stops: the deploy
3:03 Guy Currier: So where's the dividing line now? I mean, I think there's a big divide, if some of the code that you're generating is to manage the code that's being generated over here.
Tucker Callaway: The interesting thing to me is I think the dividing line is really around the deploy. And what's interesting is, if you think about...
Guy Currier: Deploying to any environment, just to be clear, through the life cycle.
Tucker Callaway: Through the life cycle, yeah.
Guy Currier: Starting with sandbox, in the old version of the word sandbox, and ending with prod.
Tucker Callaway: The old version of the word, yeah, and prod. Because what's really interesting, and this is kind of obvious, but sometimes stating the obvious is really helpful: there is a governance process that has existed for 40 years to get code into production safely. That's getting automated, but it hasn't fundamentally changed yet. So when AI creates a file of code, it is deterministic from that point forward, and the blast radius of that change being right or wrong as it goes through its governance process is virtually zero. You can roll it back, you test it, you use Greptile, you do all these things along the way to get it to a deploy.
Now, if you're trying to run something in production that you don't understand, and you're counting on AI to understand when you hit that button to restart that pod, that is like real life. It's ungoverned at that point.
So to me, we move from this world of individuals getting incredible productivity gains, which is where the skills play in. But when we get into production, we have multiple people deploying, we have multiple people making changes, and the systems that sit behind the skills have to accommodate a multiplayer environment.
Building blocks vs. opaque AI SRE services
4:40 Tucker Callaway: One of the challenges I have with the SRE space right now is I feel like we are trying to serve this group of builders, the SREs and the platform engineers, by giving them opaque, contained outcomes, like SaaS solutions, for them to get AI SRE capabilities, instead of giving them the building blocks they need to build critical infrastructure. People need to assemble and own their critical infrastructure in a multiplayer way, not just consume it from someone where they don't know what's really happening.
Scaling production beyond human understanding
5:11 Mike Vizard: So the part that I get confused with: if I look at the old way, the DevOps loop, to your point, it was a fairly structured process of moving from left to right. In the world of AI agents, it seems to me like more of these things are now happening in parallel with each other. It's not nearly as sequential, and these AI agents are doing various tasks continuously at the same time. And the cognitive load of managing that is going to be hard.
Tucker Callaway: It's impossible. If you go to our website now, it says we have to scale beyond human understanding. Because we're basically in the cycle, and the cycle is kind of working the way it's always worked, just really fast. The problem is the things it's deploying, we don't understand anymore, or we don't understand in the depth that we did when we used to handcraft these things. I think there is still understanding, but the depth of that understanding, and figuring out all the interdependencies you have in a production environment, is really hard to do now.
So we have to actually leverage AI to do that. You can't develop with AI and not run production with AI. It has to go beyond what a human can understand, because right now we're using AI to suggest RCAs and to make recommendations, but it's still gated on a human. A human still has to look at it and say, "I understand what's happening. I get this, and I'm going to make this change." And that's just not going to work when you have this avalanche of code coming your way for production.
Mike Vizard: So I have a bad feeling in the pit of my stomach about this first generation of all this, because I talk to people, and a lot of them are saying, "I'm overwhelmed. I've got too much code. I can't review it all. When I do review it, it's very verbose, and it probably has a lot of vulnerabilities in it, but we're under pressure to ship." So we're shipping more ungoverned code than ever, and that seems like a recipe for a bad thing to happen. It's going to happen.
Quality, not quantity
7:07 Guy Currier: Can I pile on that just a little bit? My contention since the GPT-3.5 moment in late 2022 has been that the advantage here with AI is actually not quantity. It's quality. It's not productivity, it's quality. You can measure productivity as, for example, the number of cycles through a given bit of code before the push or pull. In research, in testing, in all kinds of ways, you can do a whole lot more with that piece than before. And I'm just wondering if this pressure for massive quantity is just the wrong focus.
Fewer incidents, but harder ones
7:47 Tucker Callaway: Yeah, it very much could be. One of the things we're going to see, either via the improvement in quality or the increased velocity, is the nature of an incident becoming more and more complex. So if it does what you said, we'll probably have fewer incidents, but the incidents we have will be infinitely more complex, which is a good thing.
We had this interesting conversation with one of the SREs from Google, and the increase in incident time, in a way, is actually a positive, assuming it's because you've solved all the easy ones already. You're automatically solving the easy ones, so necessarily you're getting more novel failures, and that is much harder to solve for. So that could be true.
Guy Currier: Can I just say that sounds a lot more fun?
Tucker Callaway: Totally. Yeah, sure.
Guy Currier: For the humans involved. I just want to say that, granting the seriousness of breaches and security issues, that's kind of what we want: an up-level of our human activity.
Tucker Callaway: I very much agree with that, and we very much think about how we empower people. This notion that some of these roles are going to go away is ridiculous. We're actually empowering these roles to be more productive, to think about more novel failures, to think further ahead, and to take some of the routine work away from them and get that to agents.
Runtime policy for agent autonomy
9:11 Tucker Callaway: But the very hard thing is how you set policy for that. So one of the things we've created is a runtime policy enforcement, so that as the agent goes, it can go check policy in real time and make a decision: can I elevate my authority to be autonomous, or do I need to downgrade it based on the current state of the system?
Are engineers becoming agent managers?
9:27 Mike Vizard: So what does the role of the software engineer look like going forward? Because I've talked to some folks, and they're fairly advanced, but they're describing a world where they're building a software factory using AI agents. But the AI agents themselves are kind of like children. They squabble. They get jealous. If you don't pay enough attention to them, they go off and start doing things on their own. That requires a significant amount of effort to manage. One of them tells me he has company meetings with his AI agents now to define roles. Is that where we're headed with this? Is every software engineer about to become a, quote unquote, manager?
Tucker Callaway: I don't have as strong an opinion on the dev side of it. But what I can say is, I feel like this is one of those moments where the complexity that exists in the world is the complexity that exists in the world. If you shove it down in one area, it's going to blow up in the other. So there are wild productivity gains we're going to get, but there's also so much we need to manage about that.
Managing agent memory
10:33 Tucker Callaway: If you have all these agents doing all these things, you've got to manage their memory, right? And you can't just leave their memory, and memory poisoning, if you've done anything over any period of time: the things that my agent thinks that I think. Or I just ask it a random question, and now all of a sudden it has this whole opinion about me, and I don't know how to get that memory out of there.
So that's one of the things we've built too. In order to do this, you shift from doing the work into "I need to manage the memories. I need to manage how they think about semantic and episodic memory," and things like that. So while you might be guiding those agents, that becomes more complex, in the sense that I need to very carefully curate the brains of those agents all the time.
Curating shared context across agents
11:18 Guy Currier: But isn't this a human problem again? It's partly what your company is doing. A lot of times, what we wind up advising customers and vendors is that your people need to do X, Y, and Z better. Because when you describe that memory problem, I'm like, well, you've got to tell the thing not to remember that anymore. But to do that effectively requires a certain skill level in prompt engineering, if you like. And I think what vendors like you are trying to do is to take some of that requirement out of the workforce so they can just do their work, and institutionalize it: give them the systems they need to have systems work on their behalf. But it's still a human problem.
Tucker Callaway: Right. But even if I tell the AI agent not to remember something, this other AI agent that it talks to does remember that thing. In fact, they left messages for each other, and it just relearns that. And that's exactly why I talk about it as a multiplayer game versus an individual game.
So with our memory service, let's say you have five agents working on prod: we don't let them generate their own context. We curate the context, so that context is created consistently and distributed to the agents. If you're doing individual work against a shared environment, that's just not going to work. You need to curate the context to ensure it's engineered for all agents equally, that the memories are shared, and things like that. So when you improve a memory, you improve it for all the people working on it.
Guy Currier: This is what I meant by an institutional solution.
Tucker Callaway: Yes.
Guy Currier: Instead of an "everybody needs to learn better how to do these things" solution, which you can't rely on as an organization.
Mike Vizard: Guys, we've got to move on to our next segment, but I would just say a theme for the day is starting to emerge, which is that there's more unknown than known.
Tucker Callaway: Thanks for having me on.
Mike Vizard: Yeah, that was great.
