Your AI Agent Doesn't Sleep. Are You Ready for That? NVIDIA Answers.

The most important shift in AI isn't better chatbots. It's the arrival of agents that can work independently.

Give them a task before bed and they may finish it before you wake up. They can write code, access tools, analyze information and make decisions without a human in the loop. The productivity upside is enormous but so is the potential risk.

In this episode, NVIDIA's Dave Barry breaks down what enterprises are learning about autonomous agents, from secure runtimes and policy controls to observability, governance, and limiting blast radius when things go wrong.

The technology is getting more capable by the week.

The real challenge is making sure it stays pointed in the right direction.

Support for this episode provided by: CrowdStrike

More about this week's guest:

Dave Barry is a Senior Solutions Architect and Data Scientist at NVIDIA. His team focuses on NVIDIA AI software for Agentic AI and accelerated computing. While his background is rooted in computer vision, he enjoys tackling problems across every corner of AI. 

The AI Proving Ground Podcast leverages the deep AI technical and business expertise from within World Wide Technology's one-of-a-kind AI Proving Ground, which provides unrivaled access to the world's leading AI technologies. This unique lab environment accelerates your ability to learn about, test, train and implement AI solutions. 

Learn more about WWT's AI Proving Ground.

The AI Proving Ground is a composable lab environment that features the latest high-performance infrastructure and reference architectures from the world's leading AI companies, such as NVIDIA, Cisco, Dell, F5, AMD, Intel and others.

Developed within our Advanced Technology Center (ATC), this one-of-a-kind lab environment empowers IT teams to evaluate and test AI infrastructure, software and solutions for efficacy, scalability and flexibility — all under one roof. The AI Proving Ground provides visibility into data flows across the entire development pipeline, enabling more informed decision-making while safeguarding production environments. 

Transcript

The Agent Problem Nobody's Ready For

SPEAKER_00 0:00

See, I would take your metaphor a little faster. I would say we've moved past uh uh you know hammer and chisel all the way to like uh like a jackhammer, right? You know, you can you can you can uh uh pour some diesel in this and and do a lot of uh uh building or damage very, very, very quickly. And that's why we always kind of advise with with anything this powerful that you know you just step carefully.

SPEAKER_01 0:24

If you're not already, every enterprise is about to have an agent problem. Not because AI agents are failing, but because they're working. They're writing code, reading documents, calling tools, moving through workflows, and running when no one is watching. You give an AI agent a task before you go to bed, and by morning, that work is done. That's powerful, but it also raises new questions for enterprise leaders. Do you know where your agents are, what they can access, or how to stop them? On today's episode of the AI Proving Ground Podcast, we're talking with Dave Barry, a senior solutions architect and data scientist with NVIDIA to help translate NVIDIA's most advanced AI capabilities into real-world enterprise architectures and what it takes to take agents from testing to validation to production without creating risk. So, you know, AI budgets are certainly rising sharply, especially as agents become more of a reality. PWC has 88% of senior execs plan to increase AI-related budgets because of agentic AI. And two-thirds of those companies are already actually seeing, you know, meaningful ROI in a lot of that realm. But at the same time, IBM breach research shows 97% of AI-related security breaches involve systems without proper AI access controls, operationalizing layers, things like that. So it's clear organizations want to delegate more work to agents, but they still need a lot of that scaffolding and scaffolding in place. I'm curious, is that where you see Nemo Claw kind of entering the fray? Or maybe let's just start with, you know, what is Nemo Claw and what do we need to understand about it in today's AI landscape?

A Blueprint for AI That Works Alone

SPEAKER_00 2:07

Okay, so what Nemo Claw is, uh Nemo Claw falls under the category of what we call a blueprint. And right, a blueprint is in the name. You wouldn't want to build a house by guessing at things. You'd want some plan, some you know, reference architecture, right? Before you went in and built something like this. So Nemo Claw basically is our blueprint for building an autonomous ageun. It utilizes a lot of the popular ageun harnesses that are out there. Obviously, OpenClaw being one of them. Hermace is another one that it supports. And what we kind of build on top of that is, and and what Nemo Claw, you know, what makes it different from just OpenClaw by itself is this concept of uh secure runtime. And so really, again, it brings it into that category of being kind of enterprise ready. OpenClaw by itself had some security concerns, and and we've tried to address that.

SPEAKER_01 3:01

So the so it's really more about agent control and less about necessarily capability, fair?

SPEAKER_00 3:07

Correct, correct, correct. It's the idea that we're not necessarily trying to redesign one of these agent harnesses. We're just trying to basically make it more secure, make it more enterprise ready. Yeah, I mean, that's that's the big thing. And we can in in a minute maybe get into some of the the different components that are inside Nemo Claw, things like open shell.

SPEAKER_01 3:30

Well, Dave, there's a lot of industry terms being thrown

Who's Actually in Control Here?

SPEAKER_01 3:33

around there. So can you just explain the difference between open claw, open shell, and NVIDIA Nemo Claw?

SPEAKER_00 3:39

Well, you know, there's lots of claws going on there. I really like that we've we've leaned into all the lobster metaphors, you know. At at our hackathon today, we've got everybody wearing little little hats that have claws on them. So I definitely was missing that for this interview. But um, yeah, let me let me take a stab at that. So basically you can kind of break this down into like three layers, right? So we've got the agent layer, okay? That's open claw, that's Hermace, one of these agentic harnesses. And this is basically, you know, the intelligence layer. This is the autonomous agent that can, you know, connect to tools, it can connect to various messaging channels, things like Discord, Slack, Telegram. It can connect to files, it can connect to APIs, all kinds of different workflows. So, you know, again, the the open claw piece is the the you know the the brain of it, really. Open shell is really kind of like the control plane. Basically, it's it's a secure runtime that acts as kind of a sandbox and a firewall around that intelligence layer, around that open claw, around that agentic agent or that agentic intelligence. And it basically handles all of the the kind of you know connections into the sandbox, right? So everything from external requests to file system limits to policy enforcement, really all of the kind of, yeah. I mean, think of it like a like a firewall. That's a good way of thinking of it. Sure. And then Nemo Claw is really the the wrapper around all of that, right? So it's a kind of package enterprise stack that has all of the pieces included in it for you to easily set up one of these autonomous agents to set up OpenShell, to set up the models and you know, do it in a safe kind of enterprise ready way. So, really, I guess the simplest way of thinking about it is you know, OpenClaw does the work, it does the thinking. Open shell governs what it can do. So, really, again, like the firewall. And then Nemo Claw is the package around all of it that makes it easier for an enterprise to deploy it and then run it safely.

SPEAKER_01 6:01

So, OpenClaw does the work, OpenShell governs what it can do, and Nemo Claw is the package around all of it, the enterprise ready wrapper. That's the architecture.

The Night Shift Never Logs Off

SPEAKER_01 6:09

Now the question is, why does any of this actually matter? What happens when you skip it? What is materially different about managing these always-on agents versus the traditional chatbot experience that we've had with AI over the last X number of years?

SPEAKER_00 6:25

You know, that's a great thing to ask. You know, when you start working with these autonomous agents, and of course, inside NVIDIA, we we we do a lot of building with these things. And it's like anything when you when you get in and you start building with them, you you see the the pros and the cons. And and what I would say is that when something is truly autonomous, okay, and and you look at something like OpenClaw, you look at something like Nemo Claw, it's this idea that you have this autonomous, self-evolving agent that is running in the background all of the time, right? So I have one running and I can give it something to work on, and I can go to bed and come back the next day and it's done some work for me. Or it's working in the background all the time, processing some data for me. You know, it's amazing the amount of data that's created every minute. And having an AI helper, you know, an AI personal assistant that can go through and process this data in real time all of the time is very, very powerful. But it's also kind of dangerous, right? Because this thing is autonomous and it kind of has kind of a mind of its own, it means that sometimes it will do things that you don't want it to do. It will decide to, you know, one great example is, you know, if you're using a frontier model or you're using a model in the cloud, sometimes it can decide to spend a lot more tokens than you intended. So it can be a very costly thing where maybe you told it to go in and let's say look at news articles and it decides overnight that, well, that's not enough. Now I want to, you know, create all of these extra things around it. And you wake up in the morning and it's spent a thousand dollars or it spent ten thousand dollars. So you want to have limits on these things. And that's before we even get into all the pieces around, you know, what it can do with internal enterprise data. You know, we just want kind of safety first.

SPEAKER_01 8:16

When when an agent is now working after I go to sleep or as I go eat lunch, knowing that it can kind of run away with those token costs or with doing whatever it is it deems it needs to do, what do we need to be able to see within how that model or that agent works to keep it safe? What do we need to observe within that process?

You Can't Govern What You Can't See

SPEAKER_00 8:37

So I'm glad you brought that up. That's another thing that's included in Nemo Claw is we have a lot of observability tools as well. So we can see the kind of things that it's working on. You know, really, really, really important that you kind of see every step of its process, even though the whole idea with these autonomous agents is that there isn't a human in the loop. You really want some observability and some guardrails and some protection around that to see that it is kind of staying on track. So if you look at how OpenClaw is set up, OpenClaw by default basically the agent had really no rules, no limits around the tool calling it could do. And so if you had made the error of doing something like connecting banking data, connecting cloud credits, connecting anything that has real money tied to it, that's where you heard some of these horror stories of people again spending a lot of money overnight or it cleaning out a bank account. With Nemo Claw, because by default every policy needs to be approved, and you need to go in there and you need to create a policy around all of the different pieces that are allowed. The idea is that you you are very mindful about what you give it access to. By default, it is access to nothing. And you go and turn those things on one by one by one. But the idea being that, you know, you again, you you are mindful about what you give it access to. You don't kind of go in there willy-nilly. You you select the things one at a time, and the hope would be, and uh, a lot of our documentation speaks to this as well. You know, we tell people to be very, very careful about what they give it access to with any kind of real money tied to it.

SPEAKER_01 10:29

How do we make sure that that policy, those policies, that documentation actually lives within the runtime and not just stored away somewhere? Is that is that a hard thing to conquer?

SPEAKER_00 10:39

I'll I'll try not to get too technical with some of the pieces on that, but inside the inside NemoClaw, when you're going to set it up, you set up this, again, this you know, policy router. There's another privacy router that's in there, and that's where you basically spell out the different things that it's allowed to do.

SPEAKER_01 10:56

And is that an organization by organization basis, or is that something that spreads kind of across all enterprises?

SPEAKER_00 11:02

No, no. When you get in by default, basically you would build that policy from scratch. Okay. Now, we would assume that on the enterprise level that there would be some guidance around, you know, enterprises would definitely need to probably have a template for how they set those policies up. I know internally inside NVIDIA, we have been super, super careful, even about running claws ourselves, you know, until the policy until our work policies were in place to create that policy template for for the claw.

SPEAKER_01 11:34

Is that a gap that you see if you're out talking to organizations or partners? Is that a system that organizations are following where they're waiting? Because there is this balance between innovation and speed and people want to do that. There is quick.

SPEAKER_00 11:48

And there is this tendency sometimes to get something set up and go, okay, like we'll, you know, what what was Meta's quote, you know, break things fast. I wouldn't necessarily advise that as being the route I I would build a claw with. I think you do have to be so careful with it just because it is fully autonomous once you have it going. We on our website, if you look at some of the doc uh documentation around NemoClaw, we have a documentation website, we have a GitHub, we have sample policy documents on there that give some guidance around, you know, what we would believe to be best practices. Also, a lot of, you know, workbooks and examples of us building some of these claws out.

SPEAKER_01 12:32

It's interesting that you brought up the phrase break things

Before You Hand Over the Keys

SPEAKER_01 12:34

fast. That's such a popular term. I wonder if in this age that's almost irrelevant now because it break things fast as soon as you had kind of a chisel, you weren't able to actually break everything. But this feels like an era where you have a a sledgehammer instead of a a chisel. A little bit more of like a personal thought, but like do you think the break things fast mindset is still appropriate or see?

SPEAKER_00 12:55

I would take your metaphor a little faster. I would say we've moved past, you know, hammer and chisel all the way to like like a jackhammer, right? You know, you can you can you can uh pour some diesel in this and and do a lot of building or damage very, very, very quickly. And that's why we always kind of advise with with anything this powerful that you know you just step carefully. A jackhammer.

SPEAKER_01 13:18

You can pour some diesel on this thing and do a lot of building or a lot of damage very quickly. That is exactly the problem. And the answer for NVIDIA is not to slow down, it's to sandbox first, to build in a controlled environment before you let the agent anywhere near real data. This episode is supported by CrowdStrike. CrowdStrike provides cloud native endpoint protection to detect and prevent breaches in real time. Secure your organization with CrowdStrike's advanced threat intelligence and response capabilities.

SPEAKER_00 13:49

So the way we look at a sandbox is you know, it is a it's basically a controlled environment, a controlled workspace where you can go in and and you can make some mistakes, right? You can basically get in there, play around, you can learn these tools and be safe in that, you know, this isn't going to hit anything external. It's not going to hit any external data, it's not going to hit anything any internal corp data. It's it's a safe place to experiment with these. Whereas, you know, if if you were doing some of these age unharnesses before and you didn't have a sandbox, it was really risky even deploying them in the first place, particularly on any kind of you know corporate computer. So what we have have kind of treated our sandboxes as is a place to again get in there, learn the tools, make the mistakes in a safe way. And, you know, coming back to what we were talking about with policy, being very mindful about what we give it access to. What we have done to experiment with our own claws because we don't want to use any kind of internal data is we do things like create a lot of synthetic data. And so we will go in and practice in that sandbox environment with, okay, I'll give an access to this synthetic data that might resemble my email or might resemble some kind of internal data, but you know, that's safe synthesized data. Yeah. And that way I can kind of prove out the concept, I can prove out the pieces before I was to bring this to prod.

SPEAKER_01 15:20

And and who owns or who should own the readiness portion when it's mature enough to advance out of the sandbox and start to go into production? Who who should own that within an organization?

How Big Is Your Blast Radius?

SPEAKER_00 15:32

Hmm. So I would say that's kind of a hard one to answer, only because I think that like any kind of new technology that is very powerful like this, I think enterprises really need to spend a lot of time in that sandbox before they move on to the next step. I think they should also evaluate basically like how successful they've been in some of those sandbox tests. So when they go in, they do these connectors, they handle this synthetic data, you know, what mistakes are they seeing in that sandbox environment? What kind of things are they learning to kind of be careful with, right? You you find with a lot of these tools that it isn't until you're using the tools that you learn really some of the the pitfalls and some of the areas where you need to be careful. And I think once you have a good workflow in place, you want to look at more than just how accurate the agent is. You want to look at, okay, like is this workflow useful? Is it locked down enough that if I was to bring this into a production environment that I'd be comfortable with it running? And that I know how to monitor it, and most importantly, I know how to stop it if something goes wrong.

SPEAKER_01 16:51

One of the terms I've heard a lot recently on discussions that we've had here on this show is blast radius. It's been out there before, but I feel like it's just it's every episode now. Somebody mentions the word blast radius. Is this is the sandbox environment the time to think about the blast radius? And how do you think about like how do you replicate exactly, you know, where the damage can be done?

SPEAKER_00 17:11

You know, I'm not surprised that that comes up a lot. I mean, again, coming back to the example with the, you know, the hammer, the chisel, the jackhammer, these are very, very powerful tools. And they could be used for good, but they also need to be used in the right kind of responsible way. Why, you know, a sandbox is so important is it is that idea that it limits the blast radius at the beginning. Okay. Anything that you've, any mistake you make in there is contained. Now, when these things go out to to prod, when they go out to the real world, the idea is is that because you've had to be so mindful about this, you've had to practice in the sandbox, you've had to build up the policies, you've had to one by one, you know, give it access to tools and give it access, very specific access to tools. The idea is that you you limit that blast radius if something was to go wrong. And then, you know, that ties into some of the pieces we were talking about earlier with observability as well. So the ability to observe my age and and even though there isn't a you know human in the loop, I can I can stop things if something was to was to go wrong.

SPEAKER_01 18:21

So the sandbox limits the blast radius at the start, and observability is how you limit it after the agent goes live. That brings up the obvious question. When is the agent actually ready to go live? Dave has a framework for that. Three things, and they have to all be true at the same time. When do we know that the agent is good enough to advance out of the sandbox? What what is good like for what does good look like for an agent in a sandbox environment to where you're gonna feel comfortable that it can be put into deployment?

SPEAKER_00 18:51

Okay, so our

Three Rules Before You Go Live

SPEAKER_00 18:53

kind of theory around that is that basically there's three concepts that need to be in place before you move to the next phase. The first one is governance, meaning that the agent needs to have a very defined role. There needs to be a very defined use case, very explicit boundaries. Basically, everybody should understand what this agent was designed to do, the full capability of what it's allowed to do, and most importantly, what it's not allowed to do. All of these policies need to not only be in the, you know, the the the policy of the Nemo Claw itself, but those policies should also be documented on an enterprise level. So everybody understands what this is doing and what it has what it has access to. So that's the first thing is governance. Second thing is security. Best practice with setting one of these up is everything should be least privileged. And what that means is that by default, this thing has access to nothing. And everything that you give it access to, you really have to think about okay, does it really need that tool? Does it really need to do this external call? Does it really need that? And you try to limit the amount of, you know, I guess you would say vulnerabilities, right? The amount of kind of things that it can access. You really, really have to think about what sensitive data you give it. What what you you really have to look at like every single action that this ageum could do and go, okay, should it be allowed to do that? You really just have to kind of spell out all the rules for it. Just by default, really, the less you give it the better, is our thought on it. You know, and over time, if it needs more, then you're mindful and you add those on. So rather than starting from, I give it access to everything and I take it away, you give it access to nothing and you build it up. So, and then third is back to what we were talking about with the monitoring. You know, when it's in operation, you need a way to really effectively monitor it. You need an alert system in case anything goes wrong. You need to keep track of your tokens. You know, tokenomics is still so important. And most importantly, you need to have a way to turn this thing off very quickly if something goes wrong.

SPEAKER_01 21:12

Is this more of a site, not a process, a linear process, but a cycle where things have to be pulled back into a sandbox environment to then go back out to deployment?

SPEAKER_00 21:19

And that would be the way, right? It this idea of you know, if you treat the sandbox almost as a dev environment, I would do a very, very slow rollout. So to me, I would rather than then release something that has all of this tool calling, I would release it very one little piece at a time. What a little piece at a time where, okay, we've tested this extensively in the sandbox. We're very comfortable with the policy, we're very comfortable with what it has access to. And now we can do we can do a prod release of this.

Earn Trust One Rollout at a Time

SPEAKER_01 21:52

So an agent is out of the sandbox. It's now interacting with and driving workflows, maybe touching sensitive data, stuff like that. Is this where local or open model alignment or choice comes into play? I'm really glad you asked that.

SPEAKER_00 22:08

We we saw a a very you know interesting thing happen out in the ecosystem where a lot of open source models were disappearing. And so NVIDIA really, really leaned in on that. We have always been big proponents of the open source community. We really, really lean into that. And so as we saw these open source models disappear, we decided we would do our own open source models in place. And this, again, Mematron 3 family of models, very, very powerful. What's really amazing about them is they're using some really interesting architectures that allow them to run on a lot smaller footprint than some of the big Frontier models. And what this means is that you can do things where you can have a very powerful reasoning age on a very powerful autonomous agent, and you can have it running locally. For our GTC event, we had our Nematron 3 supermodel running on one of our Djax Sparks. So that has 128 gigabytes of VRAM. You know, that's a really powerful model to run in uh in a in a small space. Yeah. And what's nice about that is like everything is fully contained. So you you kind of you own your model, you own the prompts to it, you own all of that process. You don't have to do anything external.

SPEAKER_01 23:28

So how do you how do you think about how should we think about bouncing back from different models or pivoting back and forth? Or is that even part of the equation here?

SPEAKER_00 23:36

Oh, it

The Model Gets Smarter. Then What?

SPEAKER_00 23:37

totally is. That comes back to a concept that we like to call the data flywheel. Okay. And the data flywheel is this idea that that yes, like a model should evolve over time. And a model really should evolve based on its user. It's this idea that it grows, it adapts, it evolves. And what we we actually have a whole wide product suite that that allows you to do some of these pieces, our Nemo frameworks. We have a lot of tools for doing fine-tuning with some of these models and doing custom models and again building synthetic data sets and lots and lots and lots of stuff around that kind of data flywheel concept. So the data flywheel really to kind of spell it out, it's this idea that you know, as the app is running and we're looking at users, we're looking at logs, we're looking at the data and the way people are using it, that data is being used to go back and fine-tune the model underneath, make that model smarter, sometimes make that model smaller, which is amazing, where we can have a fine-tuned model replace a bigger model and actually be more effective and accurate. Yeah.

SPEAKER_01 24:45

And it's like that cycle keeps spinning. Models evolve, and NVIDIA's data flywheel is one way you can keep up with that. Every user interaction feeding back into a smarter, sometimes smaller model underneath. That's the loop. But if that loop is going to run constantly with agents learning, acting, and optimizing in real time, then the next question is what kind of infrastructure has to exist underneath it?

The Factory Behind the Future

SPEAKER_01 25:07

I mean, you talk about tokenomics, and uh a lot of what you talked about is has started getting me to think about how maybe enterprise infrastructure might be changing on the horizon. I think it's the most clear in terms of when you put in the security setting. But what has all everything we've discussed so far, what are some of the infrastructure implications that that enterprise leaders should be thinking about now to enable this future where you have these agents running at you know 247, 365?

SPEAKER_00 25:35

Well, you know, we get into that concept then, and that that's a a very broad topic, but this idea of you know the AI factory. Yeah. And that, you know, really to to do to do training and inference at scale, you do need you do need a lot of hardware if you're doing it. I mean, I say that, I mean, you are able to do fine-tuning training on a smaller profile, but to be really efficient with it, you need an AI factory. And over time, I think we'll see more, I mean, we are seeing it. We're seeing more and more of these AI factories spread because with these agents running all of the time, what it's doing too on the side is it's creating more and more and more data. And you'll know, you'll you'll notice one of the big things that Jensen spent a lot of time talking about in San Jose was this massive amount of data that we need to parse through, both structured and unstructured. And these agents are some of the creators of that. They're creating a lot of data as they're doing all of this stuff autonomously. And and we need ways to be able to convert that back into intelligence, right? And and continue to train these models.

SPEAKER_01 26:48

In your experience or from what you're seeing in the landscape, what are organizations maybe getting wrong about how to build agents or manage agents or even leverage agents?

The Mistakes Everyone Makes First

SPEAKER_01 26:58

Is there anything that you see where you're like, it's just not really how you should be looking at it?

SPEAKER_00 27:01

What I would say on that is that just like with any new tool, I think it's really, really important to get in and start using these tools right away. I know WWT is very AI forward. I know that your teams and particularly the software teams that I interface with are now using a lot of these AI tools. So I mean, that would be the first thing. But yeah, about that whole agentic lifecycle, one of the things that that NAT does really well is it helps with observability as well. And so to me, anytime I think about agentih workflows, it always comes back to observability, to observe these agents, to observe how they're working, to observe what is working and what's not working, and and and build on top of that. Another shout out I'll do here as well is my team builds a piece of software that we call the POC factory. And what it is, is it basically is a it's a way that you can build a custom blueprint and it it it goes in, it it actually is an entire AgenTic workflow. So it looks at all of our existing blueprints, it looks at a lot of our existing libraries, and at the front of the interface, you know, you upload something like an RFP. I know you guys work with that a lot, or you describe your use case, and it will go in and it will basically build a custom blueprint for you. And so we have released that internally. We're in the last phases of our security review. And when we release that externally, we're really hoping that WWT will be one of the first to take it on. You guys have been great about really leaning into this and being very cutting edge in the AI space. And honestly, it's just it's a pleasure working with you guys.

SPEAKER_01 28:52

Last question. Things change all the time. You know, you're never able to anticipate what's coming next. So I'm not gonna ask you to necessarily get out a crystal ball here, but where do you see a little bit of the the AI landscape moving over the next foreseeable future, three, six, twelve months? Nothing too far away, but you know, what should we be preparing for now that maybe we're not?

SPEAKER_00 29:13

So I think what's gonna be really, really interesting. I look at at my own personal workflow, and what what is pretty amazing now is that it's getting to this point where these AI agents are so capable and they're so powerful that at times I feel like I've got a little team of workers that work around the clock. And so I look at all of these things with with my job, with my team, with things we're building, and just this ability to be able to go out and go, okay, I'm gonna put an agent on that, and I'm gonna put an agent on this and an agent on this. And it's like I have a whole team in my pocket that is working on all of these things around the clock. And the productivity increase that we've seen, the amount of software we've been able to release, the amount, the the how fast the iterative process has gotten now because of these tools. I don't see that stopping anytime soon. And look, I'm just a standard nerd, but I'm amazed when I see people that are my my my brothers and my fathers are accountants. I have them using some of our agenti tools now to build out their spreadsheets, to build out pieces, to find something in their workflow that isn't efficient. And it's like, hey, you don't need a software guy like me anymore to write it for you. Now you can create that tool yourself. And I'm really excited to see in the wider space, in like I look at WWT and and the amazing customers that you guys work with. I'm really interested to see how people are able to analyze their workloads, identify their bottlenecks, and and build tools around it. I think that process is just gonna speed up and speed up and speed up.

SPEAKER_01 31:05

Yeah, it's an exciting future for sure. Well, Dave, self-described standard nerd, thanks for uh dropping by the studio here. Thank you for having me. We'll be talking soon in the future, I'm sure. Thank you for having me. This has been great. Okay, thanks to Dave for joining us here in

The Framework You'll Actually Need

SPEAKER_01 31:20

studio. The key lesson here three things have to be true before an agent goes into production. First, governance, a defined role, explicit boundaries, policies that live inside the system and not just in a document. Second, security, least privilege by default. Nothing gets access until you decide it needs access. And third, monitoring, the ability to see what an agent is doing and stop it fast if something goes wrong. This episode of the AI Proving Ground Podcast was co produced by Nas Baker and Kara Kuhn. Our audio and video engineers, John Knoblock. My name is Brian Felt. Thanks for listening. See you next time.