AI Agents Need a Flight Recorder, Not Just a Sandbox
AI agent governance has become an urgent concern for IT leaders. Chatbots that go wrong produce a bad paragraph. Autonomous agents that go wrong can fail deployments or leak API keys.
SUSE’s Abhinav Puri examines that shift with Mike Vizard at KubeCon + CloudNativeCon in Salt Lake City. Puri is Vice President and General Manager of Portfolio and Community at SUSE.
Sandboxing Is Only the Start
Puri believes the era of chatbots is over. Organizations now want agents that act on their own to deliver productivity.
Sandboxing caps the potential damage. NVIDIA donated OpenShell to CNCF, and SUSE has integrated it with SUSE AI Factory to isolate agents.
Isolation is not enough, though. Teams also need to trace who performed an action and with what authority.
The RAMP Test
Puri offers CIOs a four-verb test called RAMP. It stands for replay, audit, move and patch.
Replay means reconstructing what an agent saw, which model versions ran and which tools it called. Like a flight recorder, the goal is to inspect the execution flow rather than rerun it.
Audit means knowing which agent acted and on whose behalf. Today’s logs often show only a generic service account.
Move asks whether models, weights and policies can shift to another cluster next week. Patch asks how quickly a team can update its fleet when a critical vulnerability hits.
Scheduled maintenance windows won’t hold up when long-running agents carry sensitive credentials. Puri says most organizations can’t pass all four tests today.
Platform Teams Become Governors
SUSE’s assistant, Liz, now works with an AI crew of specialized agents. Customers can also connect it to their own MCP servers and scale the crew as needed.
Puri expects MCP gateways to define what agents can do and when humans step in. He urges IT leaders to treat agents like a workforce, with identities and supervision.
With skilled talent scarce, existing platform teams will need to upskill. Puri sees them as the governors of agents and the key to AI agent governance at scale.
Transcript
Hey folks, we're back here at KubeCon + CloudNativeCon North America 2026. Boy, that rolls right off the tongue, doesn't it? Bang cloud.
We are here talking about, and again, AI agents. Our friend Abhinav's joining us from SUSE. Hi.
How you guys doing? We were talking in the last two segments about AI agents and how to manage them, and how the function of an SRE is gonna change, how the function of the development team has changed, and there's more unknown than known. But what's your assessment of what's going on here?
Because a lot of these workflows on the IT side are deterministic, and a lot of the AI agents are probabilistic, and we're- Yeah ... trying to marry these things together, and I need the IT thing to be done the same way every time. Yeah.
That's absolutely correct. Let's take a step back. Everyone has implemented chatbots one way or the other in their organizations.
Now, in my view, the era of chatbots is over. It's how you implement autonomous agents to deliver productivity in your environment. And these agents, unlike chatbots, can actually create and do damage.
Because when a chatbot goes wrong, you'll probably get what? A poorly written paragraph. When an agent goes wrong, we're talking about failed deployments.
We're talking about leaked API keys. We're talking about another thousand agents also doing something wrong. Something same.
Yeah. Exactly. And then comes the whole conversation around sandboxing, which I believe is absolutely essential.
And Media donated the OpenShell, to CNCF. SUSE, from our perspective, we've integrated OpenShell with SUSE AI Factory to ensure the isolation of these agents is controlled. So sandboxing is absolutely essential.
That's one way. Now you extend that. That's not, in my opinion, enough.
You extend that because sandboxing caps your potential damage. However, you also need to know exactly what happened, and why it happened in the first place. So you need to have an ability with these agents, one, a hundred, a thousand, to trace, to audit who performed an action.
What kind of authority did they have? So I think that's the era we live in, and that's what we need to work for in this industry right now. We were just talking about this issue about the memory of the AI agents.
And part of the challenge seems to be, A, once they learn something, they don't forget it, and sometimes we want them to forget it. And other times, when we tell them to forget something, the AI agent talks to another AI agent and relearns the thing we told them to forget anyway. Yeah.
So how do we manage this thing and wrap our arms around that? Yeah. The way I think it is, I abbreviate to a four-verb test, which abbreviates RAMP: replay, audit, move, and patch.
As a CIO, when you are thinking about implementing agents in your organization, you need to understand whether you have the ability to do all four of these. Let's start with the first one, which is replay. Can you reconstruct exactly what the agent saw?
Which model versions it ran, which tools it invoked? And to your point, Mike, you may not get the same identical output every single time. It's not deterministic.
However, you need to focus on replaying the execution flow, much like a flight recorder. You don't re-fly the plane every time. Instead, you inspect the black box.
Let me see if I can clarify this for Nay. Yeah. You actually don't know what runs- Yeah ...
by the nature of this system. Unlike three years ago, let's say, you don't know the sequence of steps necessarily because that's being determined by an agent to begin with. So you need to be able to know what those steps are, and be able to replay those exact steps as opposed to- Exactly ...
the next run that might follow different steps, even if they're the same goal. Exactly. Okay.
Precisely. Because inevitably, when something goes wrong, your organization's auditor is going to ask what exactly ran. And if you have the ability to replay, you can answer that question.
So that's part one of the four-verb test. Part two is audit. Do you know which agent acted, and on whose behalf?
Because at the end of the day- Or what's behalf. Yeah. Or what's behalf.
Exactly. If you look at the standard audit logs today, you'll see effectively a general service account indicating that a port performed an action, and not necessarily telling you on which agent instance's behalf it did that. Yeah.
So your post-mortem logs, if you think about it, they end up concluding that something with elevated rights performed an action. Yeah. So that's the second thing from a CIO's lens, the RAMP test.
Replay done, audit done. The third angle is about move. So today, do you have the ability to move your entire stack on a completely different cluster in a completely different environment, even on premise if you need to, next week?
And to do that in the world of agents, you need to ensure actually that your models, your weights- Yeah ... your container policies- Sure ... your governance policies, they're all portable.
Text, data. Yeah. Exactly.
Yeah. So are we talking about I'm gonna have a small army of AI agents performing tasks- Yeah ... and I don't think a human's gonna be able to keep track of that, so I'm gonna have another set of AI agents that are essentially performing the governance function and monitoring those AI agents, and it's kind of a checks and balance kind of thing?
Somewhat. What you'll see is, one, MCP gateways coming into play, which will start governing at the end of the day what your agents can do. So what can they do?
Who can they reach? On whose behalf can they act? Who stops them?
When does a human come into the loop? So you'll see a whole thread of agent governance come up in the next-- It's already being talked about right now. It'll come into mainstream and production environments in the coming months.
So you guys at the show were talking about, you already have Liz, which is kind of an AI interface or an assistant. Yeah. And now you've added an AI crew, which is a bunch of agents that sit underneath that, that perform a specific task.
How big is that crew gonna get? Yeah. So from SUSE's perspective, the technologies that SUSE brings to the table, whether it is around observability, whether it is around security, we've enabled Liz, our assistant, to interface with them.
So that's one part of the agentic group. At the same time, we've also enabled Liz to talk to any MCP endpoint. So from a customer's lens, if they have their own MCP servers that they would like Liz to drive automated actions with, they can connect it to the Liz, and then the army of agents is actually our customer's choice.
If they want an army of 10, if they want an army of 100, we allow them to scale. That's how we've built the agentic crew with Liz. All right.
Will the agents kinda-- Are they gonna live forever and we're gonna give them names, or are they gonna be more intermittent and come and go because over time they get new skills and new capabilities? How does that kinda flow? Which agent are we talking about?
If I have an agent that's assigned a task for observability, say. At some point, that AI agent that I trained, its skills start to fall off because- Right. Yeah ...
there's drift. There's also new capabilities or the environment changes. So how dynamically am I kinda retraining the AI agents, and should I get attached to them, or should I consider them to be more like cattle than pets?
Yeah. So I would consider them more like cattle than pets. So that would be my- Well, we all want them to be more like cattle than pets.
How do we ensure-- Whichever bucket we put them in, we need that cattle-like control. Yeah. You know?
The way I see this evolving, actually, is that we need to start treating, CIOs and IT teams will need to start treating an agent force like their human workforce, with identities, with supervision, with controlling and tracking what they can access, how they can act. That is the way I see this going forward. Mm.
And those organizations who are able to scale this in a safe manner by providing the agents the right identity will be the ones who can actually win and who can actually see return on their investments that are going to put in. That's my take. So what does the future of the IT department look like in that construct?
Because today, we've got IT administrators using GUI interfaces to manage VMs. We have got DevOps teams with programming skills to manage infrastructure and the application development, and I even have a hierarchy above that, platform engineering teams that are doing stuff with DevOps at higher levels of scale. Does that collapse?
Does that stack become flatter, or what does it look like? So my view on this, Mike, is that your current platform teams will have to be upskilled, and they'll start becoming a governor of agents. And this is the way that it's going to move towards.
Because right now, the reality is that there is a shortage of talent, there is a shortage of skilled workforce who can actually implement this at scale, at enterprise level, in a safe and secure manner, in a cost-controlled manner. So you'll not get, in my opinion, a skilled workforce that you can hire on the spot for hundreds of enterprises tomorrow. So you need to reskill your existing platform teams, who will start to act as a governor of agents.
That's where we are headed. And on CIOs and IT leads should be thinking on how can they upskill their existing workforce in order to tackle this situation of tomorrow. Well, let me ask the same question as a little bit of pushback on that, just to maybe up-level the strategy here, because one of the trends right now in coding specifically, or software engineering specifically, is a bit of a decline in demand for the software engineering side and an increase in demand for the product manager side of things.
Mm-hmm. And when you think about the core functionality of a product manager, you can call it strategic. That's vague.
The idea is setting requirements, figuring out how to approach them, prioritizing them, ensuring an appropriate architecture, all of those sorts of things that can create a good, in this case, application or- Mm-hmm. Yeah, application, I guess, is the right word. Mm-hmm.
I don't see why you can't generally apply that. It's not upscale. Upscale is one way of putting it, but- Yeah ...
wouldn't you see that whether it's in SRE platform engineering, IT operations, or what have you, this sort of, let's call it a product approach- Yeah ... product development approach- Yeah ... might be more like what we're talking about.
Yeah. An orchestration of human and non-human and technological resources. Yeah.
So what are the odds that a product manager is going to have the foundation in good software engineering practices to execute that without a software engineer? So what I've seen is that I'm seeing the number of developers in an organization may go up and down, but more organizations seem to be starting to realize that that whole back end isn't going to magically happen on its own, and the product managers are not skilled enough to understand how the IT workflow works. So I think that the current structure stays, but the roles change.
I don't know. Yeah. I think there'll be a balance because what you said, directionally, I tend to agree.
What you said, I also tend to agree with. So there's a balance. And how I see- I'm saying this is a quantum question, and a few things can be true at the same time.
It will be, and here is why. Because code creation today, it's not difficult. It's easy, and that's no longer the bottleneck.
To your point, anyone could create an application. With the right logic, you can write code, good things today. That is not difficult.
What is going to be challenging is how you scale that safely and how you maintain what you have created. How do you ensure the security of the supply chain of what actually went into that code and do that at an enterprise level? And that is where things start getting trickier.
Yes. That is where the roles of the platform teams become important and relevant, however, with a different skill set. That's how I see it.
Mm-hmm. Do we need to maybe take a breath? And we were talking about this in the last segment, but we, and everybody here, worship at the altar of speed, and we are like, everything needs to be, from an application development process, done as quickly as possible.
And then I feel like there's two gates starting to show up. One is we have security issues, and we need to clearly resolve that. But third, and I think that people aren't really talking about, the rate of change is accelerating to the point now where I think many organizations are unable to absorb new features and new capabilities into software systems every other day.
Yeah. So, at what point does this thing get where the velocity is too much, and we need to throttle? Yeah.
So I was talking about the RAMP test at the time, right? Yeah. The four-letter P is, yeah.
The full test. Yeah. And this actually is an excellent segue, and Mike and I didn't talk about it, but this is exactly where the P comes in.
The P is the patch. Patch. Exactly.
So when a critical vulnerability hits our kernel or runtime, how long does it take for a team to completely update their fleet, and who controls that timeline? And now comes the part of speed. Most organizations, even today, have scheduled maintenance windows where they update their environment.
That is the reality of enterprise organizations today. However, in this era of long-running AI agents that can hold sensitive credentials, that strategy will not last long. So at an organizational level, companies will need to start thinking on how can they evolve and get more nimble in protecting and patching their environment.
And as I see it, most organizations today do not pass all the four tests, replay, audit, move, and patch simultaneously. So that's the take on what we just discussed. I think, to some degree, we were more talking about, again, the human factor, scope of expectations.
There is a market pressure, essentially, to produce more faster. If you are a first mover, there's a general feeling that that's innovative. That is differentiating.
That makes you stand out, whether as an individual or as an organization. Yeah. So what you're talking about is a way to manage the speed.
I think we also need to acknowledge that speed is something to measure itself as well. So the faster you can go safely and well, great. Yeah.
But I think that was more- We've got to end this. I want to thank you for coming by. Thank you.
Yeah, thank you. It was a pleasure. I would also just point out, the faster you go and the more you hit something, the more it hurts.
So just keep that in mind.