Tyler Akidau of Redpanda, chief technology officer, participates in theCUBE Research discussion with hosts John Furrier and Dave Vellante for the NYSE Wired Mixture of Experts series. Akidau examines agent adoption, data pipelining, agentic governance and out-of-band enforcement, and explains how streaming data infrastructure and structured query language SQL engines integrate with artificial intelligence AI agents.
Akidau recommends placing governance outside agents to ensure deterministic controls and enterprise-wide consistency. They emphasize out-of-band enforcement to mask sensitive data, kill-switch quarantine mechanisms, full transcript recording for audit trails and leveraging deterministic engines for efficiency and accuracy. Analysts on the program highlight the need for a holistic cross-system governance plan and identify significant return on investment ROI potential from targeted agent deployments.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: Mixture of Experts Series. If you don’t think you received an email check your
spam folder.
Sign in to theCUBE + NYSE Wired: Mixture of Experts Series.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open this link to automatically sign into the site.
Register For theCUBE + NYSE Wired: Mixture of Experts Series
Please fill out the information below. You will recieve an email with a verification link confirming your registration. Click the link to automatically sign into the site.
You’re almost there!
We just sent you a verification email. Please click the verification button in the email. Once your email address is verified, you will have full access to all event content for theCUBE + NYSE Wired: Mixture of Experts Series.
I want my badge and interests to be visible to all attendees.
Checking this box will display your presense on the attendees list, view your profile and allow other attendees to contact you via 1-1 chat. Read the Privacy Policy. At any time, you can choose to disable this preference.
Select your Interests!
add
Upload your photo
Uploading..
OR
Connect via Twitter
Connect via Linkedin
EDIT PASSWORD
Share
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: Mixture of Experts Series. If you don’t think you received an email check your
spam folder.
Sign in to theCUBE + NYSE Wired: Mixture of Experts Series.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open this link to automatically sign into the site.
Sign in to gain access to theCUBE + NYSE Wired: Mixture of Experts Series
Please sign in with LinkedIn to continue to theCUBE + NYSE Wired: Mixture of Experts Series. Signing in with LinkedIn ensures a professional environment.
Are you sure you want to remove access rights for this user?
Details
Manage Access
email address
Community Invitation
Tyler Akidau, Redpanda
Tyler Akidau of Redpanda, chief technology officer, participates in theCUBE Research discussion with hosts John Furrier and Dave Vellante for the NYSE Wired Mixture of Experts series. Akidau examines agent adoption, data pipelining, agentic governance and out-of-band enforcement, and explains how streaming data infrastructure and structured query language SQL engines integrate with artificial intelligence AI agents.
Akidau recommends placing governance outside agents to ensure deterministic controls and enterprise-wide consistency. They emphasize out-of-band enforcement to mask sensitive data, kill-switch quarantine mechanisms, full transcript recording for audit trails and leveraging deterministic engines for efficiency and accuracy. Analysts on the program highlight the need for a holistic cross-system governance plan and identify significant return on investment ROI potential from targeted agent deployments.
>> theCUBE Palo Alto Studio, Silicon Valley and Wall Street.
Tyler Akidau
>> I'm John Furrier, the host of theCUBE, here with Dave Vellante, my co-host.
John Furrier
>> Hello, I'm John Furrier, host of theCUBE. We are here at theCUBE's NYSE studio. Of course, we have a Palo Alto studio connecting Silicon Valley to Wall Street. This is our mixture of experts series where we talk to the leaders, the experts, in AI and in the infrastructure that are making it happen. We are seeing a massive inflection all over the news. You're seeing more and more AI infrastructure coming online that is fueling the superintelligence, the agent capabilities, a lot more. Tyler Akidau here, Chief Technology Officer, Redpanda, featured here on theCUBE many times, certainly at the NYSE Wired program. Tyler, thanks for coming in. Good to see you again.
John Furrier
>> Thank you. Great to be here.
John Furrier
>> Gemma, I think interviewed you last time with Alex.
Tyler Akidau
>> Yes.
John Furrier
>> I've been here many times. You guys are really innovative. So listen, we've covered the momentum, but you got to love the agents right now. it's prime time, all this data pipelining, dynamic data feeds coming in, GPUs and AI is just math and you got to get the data into math and feed the math. So it's like, how do you feed the beast? The AI beast. This is where agents are coming in. Give us the update on what's going on with Redpanda.
Tyler Akidau
>> So the big thing for us is, in addition to sort of feeding the beast, as you say, I think the thing that we've found with our customers and just talking to enterprises in general is everybody's really excited about AI. They want to get agents in, but nobody really sees how to do it in a way where they can trust that they have control over the agents. So the governance is really the big thing that's blocking everything now. So that's where we're headed with Redpanda now is really expanding into the AI governance space.
John Furrier
>> You got to love the growth because whenever you have a new market like this, you see, you know, chaos reigning, then you got to rein it in. And when you go across MCP servers and across boundaries, different data sets, different companies. Now you have fleets of agents. Agents are talking to other agents. The domain expertise, the governance, identity, all these security challenges bring up the issue, which is how do you make these secure and reliable? Because in the enterprise there are some deterministic workloads and probabilistic, but you can't be wrong if you're doing a financial calculation or physical AI with robotics. You can't have an error because of the physical safety. So you start to get into this, hey, your chatbot, you get a wrong answer, okay, do it again. But you really can't get them wrong in agents, and there's a security posture now that needs to be addressed.
Tyler Akidau
>> Yeah, and I think that's the key issue that's really sort of blocking agent adoption in the enterprise, is that there is, as you said, a material difference between a chatbot, where you're just talking, and if it says something crazy, you can tell it says something crazy. and an operational agent. And I think the big challenge we're seeing is, as you were alluding to, people are kind of trying to use either the security and the infrastructure built for humans or for this classic deterministic software. And agents are not like either of those, right? They are vastly more capable in a lot of ways. They're also far more clueless. They don't have the context that humans do, but they can also do things really fast, right? So, if they mix something up or hallucinate something, off they go at 100 miles an hour, just spreading damage everywhere. And so it's a very different class of kind of participant that you need to sort of secure and govern.
John Furrier
>> One of the things I want to ask you as an expert, because your work at Google, now Redpanda, you know the data business. we were talking before we came on camera, Flume, which is Apache Flume, and you guys have your own Flume at Google. You worked on that and a variety of other things. There's a whole infrastructure-first principle, almost a full pillar. So cybersecurity's moved down to the infrastructure. The data piece has kind of moved into this infrastructure layer where you got to get the plumbing right. Yep. What's the core issue? How are people doing this? What's the right strategy to get the data engineering right? That knowledge layer, that control plane, you got resource management because agents can touch things.
John Furrier
>> Yep.
John Furrier
>> it's not just data moving on some sort of semantic layer. They're reasoning, they're taking action.
Tyler Akidau
>> Mm-hmm.
John Furrier
>> What's the prescription for this? What should people be thinking about? It's not your grandfather's data pipelining strategy. It used to be.
Tyler Akidau
>> No.
John Furrier
>> It's different now. What's different?
Tyler Akidau
>> I think there's two key things you have to keep in mind to do this successfully. One is, in the limit, you want your agents touching everything in your business, right? They're basically like your new digital workforce alongside your human workforce. So they can't be running in some walled garden, right? Some platform that says, hey, come here, get all your data here, manage your agents. But what about your operational systems and other things? You want your agents able to touch everything.
John Furrier
>> You want them to move around.
Tyler Akidau
>> You want them to move around. So you need a governance layer that can work across everything, can interoperate, using open standards, things like that. And then the second thing is you really need a system that takes governance outside of the agent. A lot of the way people are doing governance with agents today is they put in the prompt. They say, hey, Don't— no, listen to what I say. Don't go share Social Security numbers. Don't go deleting the database. But as we all know, LLMs, they forget things, they hallucinate. There's all sorts of things that can happen if you leave the agents to govern themselves. You're asking for trouble. And so you've got to take that governance outside of the agents and into deterministic infrastructure that, as I said, spans your whole enterprise and allows you to connect your agents safely. To everything in it.
John Furrier
>> In the old days in IT, you'd stand up infrastructure, rack some servers, the workload runs on it, kind of vertically integrated, full stack. You mentioned walled gardens, now you have these silos. In the modern era, we're hearing the language change. We don't hear the word stand up, we hear turn on. Turn it on, not stand it up, because the infrastructure is now rack scale. So that's one feature. The other one is the data layer, you said data's got to move around. What's changing in there? Because I think this is something that people are scratching their heads saying, okay, what do I do? Because IT is everywhere now. You got to let the agents run around, as you said. But now you say IT's got to be more horizontal. What is the current state of the information technology architect's role? Not for deployed agents. I'm talking about, I want to set the generational foundation for agents. I got to nail the compliance. I got to nail the governance. People get that. What is that playbook of IT now?
Tyler Akidau
>> I think as of now, the playbook really needs to be to establish a strategy holistically across everything that you do, because you can't be having agents governed here one way and agents another way. You need to do it the same way everywhere. You've got to be recording everything they do. So everything an agent asks for and everything it gets back needs to be recorded, so you've got full transcripts of it. You can monitor it, you can debug it after the fact, you can build evaluations on top. You can have kill switches that look for agents that are trying to break out or that are drifting over time and getting worse at their jobs. You want to be able to auto-detect these things and shut them down or manually shut them down. So it's really, you need to take this holistic approach and make sure it's not an opt-in system. If you've got agents, they're going through your governance system.
John Furrier
>> Explain the kill switch because this comes up a lot. What does the kill switch mean? How do you think about that? Obviously, if you have a rogue agent out there, what does the kill switch mean? What is that? People can imagine, okay, just kill the agent.
Tyler Akidau
>> Yeah, but I think what it means in the context of, again, thinking about a large enterprise, you've probably got 6 different frameworks of agents running across thousands of different systems. And so one way you can approach it is to say, well, with each of those systems, I know how to shut off an agent. Or if you have every agent going through a governance gateway, both on the LLM access side and the model side, you've got this nice clean path to just say, even if the agent's still running, it now has no access to anything. You just shut everything off. So really just put them in a box.
John Furrier
>> Like quarantine them.
Tyler Akidau
>> Quarantine them, basically. And they stay there until you choose to let them out or shut it off.
John Furrier
>> And you have to have access to all that data. That's key to a kill switch. So you got to know it's out there.
Tyler Akidau
>> Yeah. And so what you want is you want the governance layer to be the thing providing the access.
John Furrier
>> All right. You wrote an article, O'Reilly, piece on out-of-band enforcement.
John Furrier
>> Yes.
John Furrier
>> Explain that. What does that mean, out-of-band? Does that mean out of band of scope, context? What does this mean?
Tyler Akidau
>> It's out of, from the context of an agent, out of band from the agent. The agent has no idea that the enforcement is there, has no way to touch it or alter it, and so you know that it's guaranteed to happen. Versus, as I was giving the example earlier, of saying, hey agent, you might see Social Security numbers, but don't share them, or don't record them, or maybe you're going to come across security credentials don't go try to access that system. Out-of-band means— so to give a concrete example, using the Social Security, for example, you might have an MCP tool that deals in user data. With our system, you can actually say, okay, when this— when an agent accesses this MCP tool, anytime you see a Social Security number, mask it out so the agent can never see it. So the infrastructure itself guarantees no data, no confidential data or security data ever gets—
John Furrier
>> So who's out of band, the agent or the data?
Tyler Akidau
>> The data and the metadata related to governing the agent are out of band relative to the agent. So basically saying, take governance away from the agent.
Tyler Akidau
>> So you scope the agent, basically.
Tyler Akidau
>> You scope the agent as much as you can, yeah. Keep it as tight as possible.
John Furrier
>> Explain the update on Redpanda. What's changed the most with this governance layer? Because you guys were very successful with pipelining streaming data. We saw that before. How has that played into the velocity of adoption and chaos and complexity? but you have intelligence now, you can manage that complexity. What's the biggest impact of Redpanda?
Tyler Akidau
>> I think for us, it's been a big mindset shift, but it's really an expansion for us. As you said, we started out in the data streaming business. When we were on last year, we had announced that we had acquired the company Oxla, which is a query engine, a SQL query engine. So still very heavily in the data space, but then we've also expanded into this AI governance space and built out an entire agentic governance plane that basically does all the things I've been talking about to you here. But it's a virtuous flywheel between them, really, because as you said, you want all this data access, you want data flowing through these systems. So we're sort of well-poised to be the infrastructure that sees the data first. Plus we have now the governance infrastructure that helps manage all of those systems. High-volume data as well as just the operational systems that you're
John Furrier
>> accessing.So you guys are leveraging that work you've done on the data pipeline and all the streaming data.
John Furrier
>> Yeah.That speed and intelligence, both transactional.
Tyler Akidau
>> Yeah.And if you think about what I was talking about earlier around wanting to just track everything that agents do, historically we've gathered audit logs, for example, but it's very metadata-focused. It's like, oh, Tyler accessed this database, right? But it doesn't say what was the query and what was all the data it pulled back. But with agents, you actually want to see, you want to capture all that data. That's a lot of
John Furrier
>> data.Because they're taking action
Tyler Akidau
>> too.Because they're taking action and you can't debug the agent after the fact. It's just this stochastic kind of model thing, right? And so the only thing you can really do is look at all of the data that they had and say, well, based off of what came in, what came out, maybe what happened. That's a lot of data. And so having really scalable, high-performance data infrastructure like we do is a key part of that.
John Furrier
>> All right, so the next question is, as a CTO, how do you think about ROI? Because you're starting to see visibility into performance of agents taking action, whether it's coding or a transaction. You have this business benefit popping out of these agents. You've got revenue potential impact. You got business model impact. You got certainly productivity. How are you thinking about the ROI? Because people are kind of like looking at the value side of it. It's not your traditional closed loop mechanism because there's actual impact.
Tyler Akidau
>> No, it's crazy. And it's also not the traditional, a lot of what we've dealt in historically has been, you know, how much data is there? The value comes from the volume, right? It's not a volume business anymore either. Like, it's crazy. Some of our customers, they deploy agents, data volume's just tiny, but the business value is just crazy. Millions, billions of dollars. It's insane.
John Furrier
>> Yeah, the whole token maxing wave is kind of turning into value maxing, as they say on, I've been following all the threads on X. But if you think about it, it's like almost, you could also go back to the old school days of when we were coding, in the '80s, you had very limited memory. You had to write memory managers and swap out memory to disk. But you're starting to see that efficiency mindset of, I did this with this amount of tokens, so small is better. So if you can get leverage from the tokens versus just blowing tokens, which is like writing tons of code. You don't get measured on the volume of code back in the old days. It's like, your code's tight, it's strong. We're kind of coming back around to this performance tuning of agents. What's your thoughts on that? How do you look at that? How do you frame that?
Tyler Akidau
>> Oh yeah, there's so much work to be done there. There's, right now, very similar to what I was saying about governance earlier, everyone's just kind of asking the agent, go do this, right? But you ask an agent to join two datasets and compute a weighted average over the data, good luck with the results. A, it's going to burn a lot of tokens. And so we've actually done benchmarks on this. Half the time they're going to get the answers wrong. Whereas if you can have them write a SQL query, for example, and say, "Hey, here, you're going to join some Salesforce data and some Jira data and do a weighted average." They're really good at writing SQL queries, and then you've offloaded all that to this deterministic engine that then is vastly more efficient, does it faster, and gives you more correct answers.
John Furrier
>> I was talking to a Python developer in their 20s, and he's a SQL guru, but he loves writing SQL code. I was talking to a business person who does SQL, and the comment was, I used to have to figure out the business logic, then do the SQL query, versus now I just say, give me the third quarter sales forecast, and that writes the queries. So the logic is being embedded into the AI or the intelligence layer. You guys have done some research in this area, the role of the human interaction piece, because I think you mentioned the humans have to manage the agents. They work on behalf of the humans and the business. As the users are going to abstract away to just voice prompts, if you will. Hey, get me that data. Give me the weighted average. Not even say merge the two datasets. It might know the datasets.
John Furrier
>> Yep.
John Furrier
>> This becomes fundamental in this next wave. What's your thoughts on that? What's the research telling you? The role of the human in the interaction? Because it's not GUI-based, it's not query-based. the SQL is being abstracted away from the user, certainly being used in the logic. So as the business logic gets into the agents, The humans play an important role.
Tyler Akidau
>> Yeah, no, the humans definitely play an important role in sort of shaping what needs to be done. I think the other part that we're finding that's interesting with agents in business, agents in enterprise, that's very different from kind of the copilots that we all kind of like to use or the chatbots, is that the business process agents are much more sort of structured in what you want them to do, right? Like, it's not a greenfield, like, oh, go research some random stuff for me. It's like, look, I've got a job for you. To do. It's X, then Y, then Z. And there's probably some choose-your-own-adventure between those, but roughly the flow needs to be X, Y, Z. It can never be Z, Z, Z, X, Y, or something like that. And so what we're finding and what we're working on with a bunch of the stuff that we're making is finding ways to kind of bridge— you want a little bit of determinism there, you want a structure for the agents, you want to define what the flow is at a high level while still leaving them the opportunity to do what they're good at, which is non-deterministic kind of thinking style actions where it's helpful, you know?
John Furrier
>> Yeah. It's interesting, back in other waves and other inflection points, smaller inflection points, certainly, anytime a new technology came out of the enterprise, it was okay, cool, great, an IT innovation mechanism. Virtual desktops. No one really cried a river when they went home and said, I want my virtual desktop. With AI and agents, there is a huge end-user appetite. People are adopting them. They're deploying them like crazy. So you have that end user or user perspective, and then you got the infrastructure piece of it. You guys are the center of that wave.
John Furrier
>> Yes.
John Furrier
>> The demand's there because people are pounding agents, throwing them into the system. This is kind of putting a tsunami effect on the data players.
Tyler Akidau
>> No, absolutely.
John Furrier
>> What's your reaction to that? Because it forces you to go faster, but you're also dealing with all these agents coming into the fold. You got to rein them in.
Tyler Akidau
>> No, it's— yeah, and I think that's really the sentiment we see everywhere, is everyone's really excited about this and they're horrified. They're just absolutely horrified, and they're trying to figure out how can I do this in a way that I feel like I have control, I know, I have confidence that the agents are going to be able to do what they need to, and if they aren't, that we know and that we can shut it down and not have damage. That's really what they're aiming for.
John Furrier
>> All right, so I'm interested. I'm an IT person. Or a business line manager, people running agents are all over the place. What do I do? What's the Redpanda playbook?
Tyler Akidau
>> So our playbook is the agentic data plane that we've built. The idea is really you've got this kind of foundational layer of agent access to LLMs, models, and tools all goes through the governance layer. That's sort of the ground level. Once you have that, then via those gateways and via other mechanisms, you can collect all the data around the things that they're doing. So you have these transcripts, you know what the agent is doing holistically. You can also insert into that these out-of-band governance layers that allow you to narrow the scope of what agents can do. Like an agent acting on behalf of me shouldn't be able to go and delete the entire database, even though I have that permission, right? Like, I should know, like, I really just need this agent to read.
Tyler Akidau
>> It's like credentials.
Tyler Akidau
>> Yeah. So it can adopt my credentials, but then those credentials get narrowed so that it can only read and it can't go drop a thing.
John Furrier
>> It's the rules of the road.
Tyler Akidau
>> Exactly. Or like visitor badges, you can have a visitor badge, but that doesn't let you go into the secret server room, right? And so then you need to layer that stuff on top, and then you also need to start building on top of it all the things around evaluations and guardrails and kill switches and that sort of stuff. And again, that's kind of a bunch of data—
John Furrier
>> Is there a prerequisite for a customerto be in good data shape, if you will? What's the prerequisite?
Tyler Akidau
>> I think the big prerequisite is to really understand that for this to be successful, it's going to be across your entire enterprise, and you've got to be prepared to govern across everything. You can't just focus on one part.
John Furrier
>> Get your data in line, basically.
Tyler Akidau
>> Get your data in line everywhere and have a holistic plan.
John Furrier
>> Awesome. All right, well, thanks for coming on theCUBE. Appreciate it. Good to see you again. Say hello to Alex and the team at Redpanda again. You guys again did a lot of work pre-Agent. Now you're in a good position.
Tyler Akidau
>> Yep.Thanks for coming on. We're in it deep.
John Furrier
>> Appreciate it. I'm John Furrier. This is a mixture of experts. You got to get the data in line. Agents need to be everywhere. They need to see and touch things. They're touching resources, writing code, taking actions. And if you don't rein them in with the right governance layer, they could go off the rails. And of course, you have now out-of-band capabilities. Scope them. That's the discussion we're hearing a lot more on theCUBE. Thanks for watching.
>> theCUBE Palo Alto Studio, Silicon Valley and Wall Street.
Tyler Akidau
>> I'm John Furrier, the host of theCUBE, here with Dave Vellante, my co-host.
John Furrier
>> Hello, I'm John Furrier, host of theCUBE. We are here at theCUBE's NYSE studio. Of course, we have a Palo Alto studio connecting Silicon Valley to Wall Street. This is our mixture of experts series where we talk to the leaders, the experts, in AI and in the infrastructure that are making it happen. We are seeing a massive inflection all over the news. You're seeing more and more AI infrastructure coming online that is fueling the superintelligence, the agent capabilities, a lot more. Tyler Akidau here, Chief Technology Officer, Redpanda, featured here on theCUBE many times, certainly at the NYSE Wired program. Tyler, thanks for coming in. Good to see you again.
John Furrier
>> Thank you. Great to be here.
John Furrier
>> Gemma, I think interviewed you last time with Alex.
Tyler Akidau
>> Yes.
John Furrier
>> I've been here many times. You guys are really innovative. So listen, we've covered the momentum, but you got to love the agents right now. it's prime time, all this data pipelining, dynamic data feeds coming in, GPUs and AI is just math and you got to get the data into math and feed the math. So it's like, how do you feed the beast? The AI beast. This is where agents are coming in. Give us the update on what's going on with Redpanda.
Tyler Akidau
>> So the big thing for us is, in addition to sort of feeding the beast, as you say, I think the thing that we've found with our customers and just talking to enterprises in general is everybody's really excited about AI. They want to get agents in, but nobody really sees how to do it in a way where they can trust that they have control over the agents. So the governance is really the big thing that's blocking everything now. So that's where we're headed with Redpanda now is really expanding into the AI governance space.
John Furrier
>> You got to love the growth because whenever you have a new market like this, you see, you know, chaos reigning, then you got to rein it in. And when you go across MCP servers and across boundaries, different data sets, different companies. Now you have fleets of agents. Agents are talking to other agents. The domain expertise, the governance, identity, all these security challenges bring up the issue, which is how do you make these secure and reliable? Because in the enterprise there are some deterministic workloads and probabilistic, but you can't be wrong if you're doing a financial calculation or physical AI with robotics. You can't have an error because of the physical safety. So you start to get into this, hey, your chatbot, you get a wrong answer, okay, do it again. But you really can't get them wrong in agents, and there's a security posture now that needs to be addressed.
Tyler Akidau
>> Yeah, and I think that's the key issue that's really sort of blocking agent adoption in the enterprise, is that there is, as you said, a material difference between a chatbot, where you're just talking, and if it says something crazy, you can tell it says something crazy. and an operational agent. And I think the big challenge we're seeing is, as you were alluding to, people are kind of trying to use either the security and the infrastructure built for humans or for this classic deterministic software. And agents are not like either of those, right? They are vastly more capable in a lot of ways. They're also far more clueless. They don't have the context that humans do, but they can also do things really fast, right? So, if they mix something up or hallucinate something, off they go at 100 miles an hour, just spreading damage everywhere. And so it's a very different class of kind of participant that you need to sort of secure and govern.
John Furrier
>> One of the things I want to ask you as an expert, because your work at Google, now Redpanda, you know the data business. we were talking before we came on camera, Flume, which is Apache Flume, and you guys have your own Flume at Google. You worked on that and a variety of other things. There's a whole infrastructure-first principle, almost a full pillar. So cybersecurity's moved down to the infrastructure. The data piece has kind of moved into this infrastructure layer where you got to get the plumbing right. Yep. What's the core issue? How are people doing this? What's the right strategy to get the data engineering right? That knowledge layer, that control plane, you got resource management because agents can touch things.
John Furrier
>> Yep.
John Furrier
>> it's not just data moving on some sort of semantic layer. They're reasoning, they're taking action.
Tyler Akidau
>> Mm-hmm.
John Furrier
>> What's the prescription for this? What should people be thinking about? It's not your grandfather's data pipelining strategy. It used to be.
Tyler Akidau
>> No.
John Furrier
>> It's different now. What's different?
Tyler Akidau
>> I think there's two key things you have to keep in mind to do this successfully. One is, in the limit, you want your agents touching everything in your business, right? They're basically like your new digital workforce alongside your human workforce. So they can't be running in some walled garden, right? Some platform that says, hey, come here, get all your data here, manage your agents. But what about your operational systems and other things? You want your agents able to touch everything.
John Furrier
>> You want them to move around.
Tyler Akidau
>> You want them to move around. So you need a governance layer that can work across everything, can interoperate, using open standards, things like that. And then the second thing is you really need a system that takes governance outside of the agent. A lot of the way people are doing governance with agents today is they put in the prompt. They say, hey, Don't— no, listen to what I say. Don't go share Social Security numbers. Don't go deleting the database. But as we all know, LLMs, they forget things, they hallucinate. There's all sorts of things that can happen if you leave the agents to govern themselves. You're asking for trouble. And so you've got to take that governance outside of the agents and into deterministic infrastructure that, as I said, spans your whole enterprise and allows you to connect your agents safely. To everything in it.
John Furrier
>> In the old days in IT, you'd stand up infrastructure, rack some servers, the workload runs on it, kind of vertically integrated, full stack. You mentioned walled gardens, now you have these silos. In the modern era, we're hearing the language change. We don't hear the word stand up, we hear turn on. Turn it on, not stand it up, because the infrastructure is now rack scale. So that's one feature. The other one is the data layer, you said data's got to move around. What's changing in there? Because I think this is something that people are scratching their heads saying, okay, what do I do? Because IT is everywhere now. You got to let the agents run around, as you said. But now you say IT's got to be more horizontal. What is the current state of the information technology architect's role? Not for deployed agents. I'm talking about, I want to set the generational foundation for agents. I got to nail the compliance. I got to nail the governance. People get that. What is that playbook of IT now?
Tyler Akidau
>> I think as of now, the playbook really needs to be to establish a strategy holistically across everything that you do, because you can't be having agents governed here one way and agents another way. You need to do it the same way everywhere. You've got to be recording everything they do. So everything an agent asks for and everything it gets back needs to be recorded, so you've got full transcripts of it. You can monitor it, you can debug it after the fact, you can build evaluations on top. You can have kill switches that look for agents that are trying to break out or that are drifting over time and getting worse at their jobs. You want to be able to auto-detect these things and shut them down or manually shut them down. So it's really, you need to take this holistic approach and make sure it's not an opt-in system. If you've got agents, they're going through your governance system.
John Furrier
>> Explain the kill switch because this comes up a lot. What does the kill switch mean? How do you think about that? Obviously, if you have a rogue agent out there, what does the kill switch mean? What is that? People can imagine, okay, just kill the agent.
Tyler Akidau
>> Yeah, but I think what it means in the context of, again, thinking about a large enterprise, you've probably got 6 different frameworks of agents running across thousands of different systems. And so one way you can approach it is to say, well, with each of those systems, I know how to shut off an agent. Or if you have every agent going through a governance gateway, both on the LLM access side and the model side, you've got this nice clean path to just say, even if the agent's still running, it now has no access to anything. You just shut everything off. So really just put them in a box.
John Furrier
>> Like quarantine them.
Tyler Akidau
>> Quarantine them, basically. And they stay there until you choose to let them out or shut it off.
John Furrier
>> And you have to have access to all that data. That's key to a kill switch. So you got to know it's out there.
Tyler Akidau
>> Yeah. And so what you want is you want the governance layer to be the thing providing the access.
John Furrier
>> All right. You wrote an article, O'Reilly, piece on out-of-band enforcement.
John Furrier
>> Yes.
John Furrier
>> Explain that. What does that mean, out-of-band? Does that mean out of band of scope, context? What does this mean?
Tyler Akidau
>> It's out of, from the context of an agent, out of band from the agent. The agent has no idea that the enforcement is there, has no way to touch it or alter it, and so you know that it's guaranteed to happen. Versus, as I was giving the example earlier, of saying, hey agent, you might see Social Security numbers, but don't share them, or don't record them, or maybe you're going to come across security credentials don't go try to access that system. Out-of-band means— so to give a concrete example, using the Social Security, for example, you might have an MCP tool that deals in user data. With our system, you can actually say, okay, when this— when an agent accesses this MCP tool, anytime you see a Social Security number, mask it out so the agent can never see it. So the infrastructure itself guarantees no data, no confidential data or security data ever gets—
John Furrier
>> So who's out of band, the agent or the data?
Tyler Akidau
>> The data and the metadata related to governing the agent are out of band relative to the agent. So basically saying, take governance away from the agent.
Tyler Akidau
>> So you scope the agent, basically.
Tyler Akidau
>> You scope the agent as much as you can, yeah. Keep it as tight as possible.
John Furrier
>> Explain the update on Redpanda. What's changed the most with this governance layer? Because you guys were very successful with pipelining streaming data. We saw that before. How has that played into the velocity of adoption and chaos and complexity? but you have intelligence now, you can manage that complexity. What's the biggest impact of Redpanda?
Tyler Akidau
>> I think for us, it's been a big mindset shift, but it's really an expansion for us. As you said, we started out in the data streaming business. When we were on last year, we had announced that we had acquired the company Oxla, which is a query engine, a SQL query engine. So still very heavily in the data space, but then we've also expanded into this AI governance space and built out an entire agentic governance plane that basically does all the things I've been talking about to you here. But it's a virtuous flywheel between them, really, because as you said, you want all this data access, you want data flowing through these systems. So we're sort of well-poised to be the infrastructure that sees the data first. Plus we have now the governance infrastructure that helps manage all of those systems. High-volume data as well as just the operational systems that you're
John Furrier
>> accessing.So you guys are leveraging that work you've done on the data pipeline and all the streaming data.
John Furrier
>> Yeah.That speed and intelligence, both transactional.
Tyler Akidau
>> Yeah.And if you think about what I was talking about earlier around wanting to just track everything that agents do, historically we've gathered audit logs, for example, but it's very metadata-focused. It's like, oh, Tyler accessed this database, right? But it doesn't say what was the query and what was all the data it pulled back. But with agents, you actually want to see, you want to capture all that data. That's a lot of
John Furrier
>> data.Because they're taking action
Tyler Akidau
>> too.Because they're taking action and you can't debug the agent after the fact. It's just this stochastic kind of model thing, right? And so the only thing you can really do is look at all of the data that they had and say, well, based off of what came in, what came out, maybe what happened. That's a lot of data. And so having really scalable, high-performance data infrastructure like we do is a key part of that.
John Furrier
>> All right, so the next question is, as a CTO, how do you think about ROI? Because you're starting to see visibility into performance of agents taking action, whether it's coding or a transaction. You have this business benefit popping out of these agents. You've got revenue potential impact. You got business model impact. You got certainly productivity. How are you thinking about the ROI? Because people are kind of like looking at the value side of it. It's not your traditional closed loop mechanism because there's actual impact.
Tyler Akidau
>> No, it's crazy. And it's also not the traditional, a lot of what we've dealt in historically has been, you know, how much data is there? The value comes from the volume, right? It's not a volume business anymore either. Like, it's crazy. Some of our customers, they deploy agents, data volume's just tiny, but the business value is just crazy. Millions, billions of dollars. It's insane.
John Furrier
>> Yeah, the whole token maxing wave is kind of turning into value maxing, as they say on, I've been following all the threads on X. But if you think about it, it's like almost, you could also go back to the old school days of when we were coding, in the '80s, you had very limited memory. You had to write memory managers and swap out memory to disk. But you're starting to see that efficiency mindset of, I did this with this amount of tokens, so small is better. So if you can get leverage from the tokens versus just blowing tokens, which is like writing tons of code. You don't get measured on the volume of code back in the old days. It's like, your code's tight, it's strong. We're kind of coming back around to this performance tuning of agents. What's your thoughts on that? How do you look at that? How do you frame that?
Tyler Akidau
>> Oh yeah, there's so much work to be done there. There's, right now, very similar to what I was saying about governance earlier, everyone's just kind of asking the agent, go do this, right? But you ask an agent to join two datasets and compute a weighted average over the data, good luck with the results. A, it's going to burn a lot of tokens. And so we've actually done benchmarks on this. Half the time they're going to get the answers wrong. Whereas if you can have them write a SQL query, for example, and say, "Hey, here, you're going to join some Salesforce data and some Jira data and do a weighted average." They're really good at writing SQL queries, and then you've offloaded all that to this deterministic engine that then is vastly more efficient, does it faster, and gives you more correct answers.
John Furrier
>> I was talking to a Python developer in their 20s, and he's a SQL guru, but he loves writing SQL code. I was talking to a business person who does SQL, and the comment was, I used to have to figure out the business logic, then do the SQL query, versus now I just say, give me the third quarter sales forecast, and that writes the queries. So the logic is being embedded into the AI or the intelligence layer. You guys have done some research in this area, the role of the human interaction piece, because I think you mentioned the humans have to manage the agents. They work on behalf of the humans and the business. As the users are going to abstract away to just voice prompts, if you will. Hey, get me that data. Give me the weighted average. Not even say merge the two datasets. It might know the datasets.
John Furrier
>> Yep.
John Furrier
>> This becomes fundamental in this next wave. What's your thoughts on that? What's the research telling you? The role of the human in the interaction? Because it's not GUI-based, it's not query-based. the SQL is being abstracted away from the user, certainly being used in the logic. So as the business logic gets into the agents, The humans play an important role.
Tyler Akidau
>> Yeah, no, the humans definitely play an important role in sort of shaping what needs to be done. I think the other part that we're finding that's interesting with agents in business, agents in enterprise, that's very different from kind of the copilots that we all kind of like to use or the chatbots, is that the business process agents are much more sort of structured in what you want them to do, right? Like, it's not a greenfield, like, oh, go research some random stuff for me. It's like, look, I've got a job for you. To do. It's X, then Y, then Z. And there's probably some choose-your-own-adventure between those, but roughly the flow needs to be X, Y, Z. It can never be Z, Z, Z, X, Y, or something like that. And so what we're finding and what we're working on with a bunch of the stuff that we're making is finding ways to kind of bridge— you want a little bit of determinism there, you want a structure for the agents, you want to define what the flow is at a high level while still leaving them the opportunity to do what they're good at, which is non-deterministic kind of thinking style actions where it's helpful, you know?
John Furrier
>> Yeah. It's interesting, back in other waves and other inflection points, smaller inflection points, certainly, anytime a new technology came out of the enterprise, it was okay, cool, great, an IT innovation mechanism. Virtual desktops. No one really cried a river when they went home and said, I want my virtual desktop. With AI and agents, there is a huge end-user appetite. People are adopting them. They're deploying them like crazy. So you have that end user or user perspective, and then you got the infrastructure piece of it. You guys are the center of that wave.
John Furrier
>> Yes.
John Furrier
>> The demand's there because people are pounding agents, throwing them into the system. This is kind of putting a tsunami effect on the data players.
Tyler Akidau
>> No, absolutely.
John Furrier
>> What's your reaction to that? Because it forces you to go faster, but you're also dealing with all these agents coming into the fold. You got to rein them in.
Tyler Akidau
>> No, it's— yeah, and I think that's really the sentiment we see everywhere, is everyone's really excited about this and they're horrified. They're just absolutely horrified, and they're trying to figure out how can I do this in a way that I feel like I have control, I know, I have confidence that the agents are going to be able to do what they need to, and if they aren't, that we know and that we can shut it down and not have damage. That's really what they're aiming for.
John Furrier
>> All right, so I'm interested. I'm an IT person. Or a business line manager, people running agents are all over the place. What do I do? What's the Redpanda playbook?
Tyler Akidau
>> So our playbook is the agentic data plane that we've built. The idea is really you've got this kind of foundational layer of agent access to LLMs, models, and tools all goes through the governance layer. That's sort of the ground level. Once you have that, then via those gateways and via other mechanisms, you can collect all the data around the things that they're doing. So you have these transcripts, you know what the agent is doing holistically. You can also insert into that these out-of-band governance layers that allow you to narrow the scope of what agents can do. Like an agent acting on behalf of me shouldn't be able to go and delete the entire database, even though I have that permission, right? Like, I should know, like, I really just need this agent to read.
Tyler Akidau
>> It's like credentials.
Tyler Akidau
>> Yeah. So it can adopt my credentials, but then those credentials get narrowed so that it can only read and it can't go drop a thing.
John Furrier
>> It's the rules of the road.
Tyler Akidau
>> Exactly. Or like visitor badges, you can have a visitor badge, but that doesn't let you go into the secret server room, right? And so then you need to layer that stuff on top, and then you also need to start building on top of it all the things around evaluations and guardrails and kill switches and that sort of stuff. And again, that's kind of a bunch of data—
John Furrier
>> Is there a prerequisite for a customerto be in good data shape, if you will? What's the prerequisite?
Tyler Akidau
>> I think the big prerequisite is to really understand that for this to be successful, it's going to be across your entire enterprise, and you've got to be prepared to govern across everything. You can't just focus on one part.
John Furrier
>> Get your data in line, basically.
Tyler Akidau
>> Get your data in line everywhere and have a holistic plan.
John Furrier
>> Awesome. All right, well, thanks for coming on theCUBE. Appreciate it. Good to see you again. Say hello to Alex and the team at Redpanda again. You guys again did a lot of work pre-Agent. Now you're in a good position.
Tyler Akidau
>> Yep.Thanks for coming on. We're in it deep.
John Furrier
>> Appreciate it. I'm John Furrier. This is a mixture of experts. You got to get the data in line. Agents need to be everywhere. They need to see and touch things. They're touching resources, writing code, taking actions. And if you don't rein them in with the right governance layer, they could go off the rails. And of course, you have now out-of-band capabilities. Scope them. That's the discussion we're hearing a lot more on theCUBE. Thanks for watching.
John Furrier
Tyler Akidau