# Securing Long‑Running AI Agents: From Setup to Sandboxing

## Abstract

Explore how to securely run OpenClaw‑style, always‑on AI agents using NVIDIA NeMoClaw, an open‑source reference stack with built‑in privacy and security guardrails. We’ll dive into how to set up the runtime, apply policy‑based access controls, and add lightweight observability for safer agent management. You’ll learn practical techniques to limit file and network access, detect issues early, and strengthen agent reliability. This session helps developers and MLOps engineers confidently deploy long‑running AI agents across any environment. This session is part of Nemotron Days : a curated set of GTC Taipei sessions focused on agentic AI, claws, and long-running agents.

## AI Summary

- The presentation highlighted the evolution of AI agents, emphasizing three key inflection points: the ChatGPT moment, the DeepSeek moment, and the Claw moment, where agents became autonomous and useful for various tasks.
- The NVIDIA Agent Toolkit, including the Nemotron Ultra model, was introduced to help build and manage long-running agents that could take actions independently.
- The concept of skills was explained, where NVIDIA’s libraries were packaged to enable agents to use them effectively, with a focus on verified skills to ensure security and trust.
- OpenShell, an open-source project, was presented as a secure runtime environment that allowed users to control and monitor what their agents could access and do, preventing leaks of sensitive information.
- NeMoClaw blueprints were discussed as a starting point for users to build and customize their own agents, integrating with tools like Slack and GitHub to enhance productivity.
- NeMo Relay was introduced as a tool to observe and optimize agent behavior, saving token costs and improving efficiency by learning from agent interactions over time.

## Transcript

**[00:00:05 – 00:00:09]** All right, good afternoon, Taipei.

**[00:00:09 – 00:00:10]** My name is Adel.

**[00:00:10 – 00:00:12]** I do two things.

**[00:00:12 – 00:00:16]** I build agents, and I also babysit agents.

**[00:00:16 – 00:00:18]** That's literally what I do at work.

**[00:00:18 – 00:00:23]** We've got a lot to cover, a few demos, so let's just kick it off.

**[00:00:23 – 00:00:26]** The last three years have been quite incredible, but

**[00:00:26 – 00:00:29]** I think everybody talks about inflection points, you know,

**[00:00:29 – 00:00:30]** days that you remember.

**[00:00:30 – 00:00:33]** And there are three of them that we have up here.

**[00:00:33 – 00:00:36]** The first one was the ChatGPT moment.

**[00:00:36 – 00:00:39]** Does everybody remember when the ChatGPT moment?

**[00:00:39 – 00:00:41]** I certainly do, and I think it was the first time my parents

**[00:00:42 – 00:00:44]** finally understood what deep learning was, what AI was,

**[00:00:44 – 00:00:47]** and it was fantastic, right?

**[00:00:47 – 00:00:51]** It could generate poems for you, it can write drafts of emails,

**[00:00:51 – 00:00:54]** could even generate some pictures.

**[00:00:54 – 00:00:58]** We had to deal with hallucinations, and so retrieval augmented

**[00:00:58 – 00:00:59]** generation became a thing.

**[00:01:00 – 00:01:03]** We figured out how to ground these LLMs with our knowledge

**[00:01:03 – 00:01:05]** bases, and it was great.

**[00:01:05 – 00:01:09]** It gave us value, but it was always a human taking the action.

**[00:01:09 – 00:01:13]** You're asking for it to write your email draft or generate

**[00:01:13 – 00:01:16]** a picture or a poem.

**[00:01:16 – 00:01:19]** About a year and a half ago, another amazing thing

**[00:01:19 – 00:01:23]** happened, and I call that a DeepSeek Seek Moments.

**[00:01:23 – 00:01:25]** The DeepSeek moment. It was the first time we said, you

**[00:01:25 – 00:01:27]** know, the 4-0 moment.

**[00:01:27 – 00:01:30]** It was when reasoning models came to be.

**[00:01:30 – 00:01:31]** And I gotta tell you guys, it was the first time that

**[00:01:32 – 00:01:34]** I saw a model talk to itself.

**[00:01:34 – 00:01:37]** I even saw a model high-five itself, like give itself

**[00:01:37 – 00:01:39]** a high-five emoji.

**[00:01:39 – 00:01:43]** But what reasoning did, what reasoning did was, these models

**[00:01:43 – 00:01:45]** could think through plans.

**[00:01:45 – 00:01:49]** They could do some self-reflection, right?

**[00:01:49 – 00:01:52]** They could think through multiple different steps.

**[00:01:52 – 00:01:57]** And, you know, from ChatGPT to the DeepSeek moment, right,

**[00:01:57 – 00:02:03]** we saw about a 10x growth and consumption in tokens, right?

**[00:02:03 – 00:02:05]** But it added value.

**[00:02:05 – 00:02:09]** Now, another moment occurred that I think we all remember.

**[00:02:09 – 00:02:12]** We call that the claw moment.

**[00:02:12 – 00:02:20]** That was the time in about January of this year that we saw these

**[00:02:20 – 00:02:23]** autonomous, self-evolving agents.

**[00:02:23 – 00:02:26]** And frankly, this is what we're going to be talking about today.

**[00:02:26 – 00:02:29]** But what is this moment about?

**[00:02:29 – 00:02:33]** What is this inflection point that we're living right now about?

**[00:02:33 – 00:02:37]** Well, AI has finally become useful.

**[00:02:37 – 00:02:41]** AIs become useful, and I don't mean in the sense that it could make

**[00:02:41 – 00:02:45]** me more productive, but it could actually take action all by itself.

**[00:02:45 – 00:02:50]** And this is a bit outdated of a chart, but it symbolizes and

**[00:02:50 – 00:02:57]** shows just how popular Claws, these long-running agents, really became.

**[00:02:57 – 00:03:02]** If you compare it to Facebook or to Linux, it surpassed

**[00:03:02 – 00:03:03]** these two foundations.

**[00:03:03 – 00:03:07]** Linux, the operating system of the world, React, the

**[00:03:07 – 00:03:09]** backbone of the web.

**[00:03:09 – 00:03:12]** It overtook those in just a few weeks, and that was

**[00:03:12 – 00:03:14]** because it's really, really useful.

**[00:03:14 – 00:03:16]** Across a different array of different things.

**[00:03:16 – 00:03:17]** But let's not just talk about it.

**[00:03:17 – 00:03:22]** Let me show you a few examples of how this is useful.

**[00:03:23 – 00:03:26]** Can we get the volume up, please?

**[00:04:49 – 00:04:51]** An open source project just dropped.

**[00:04:51 – 00:04:53]** Andrej Karpathy has just launched something called

**[00:04:53 – 00:04:55]** So research is a huge deal.

**[00:04:55 – 00:04:58]** You give an AI agent a task, go to sleep, it runs 100 experiments

**[00:04:59 – 00:05:03]** overnight, keeping what works and killing what doesn't.

**[00:05:12 – 00:05:16]** I really love what my stuff enables that person to do

**[00:05:16 – 00:05:20]** and we had like one guy he told me like he installed it as a 60 year

**[00:05:20 – 00:05:22]** old dad and like they made beer.

**[00:05:22 – 00:05:24]** Connected the machine via bluetooth to OpenClaw and then we

**[00:05:24 – 00:05:28]** automated everything including the whole website for people to order.

**[00:05:28 – 00:05:31]** Lobster lager fertig sein.

**[00:05:35 – 00:05:42]** Everyone is talking about We want to build OpenClaw with OpenClaw.

**[00:05:42 – 00:05:44]** What the is OpenClaw?

**[00:05:44 – 00:05:47]** Hundreds of people are queuing up for lobsters in St.

**[00:05:47 – 00:05:52]** Believe it or not, there's already a claw con.

**[00:05:57 – 00:06:01]** I mean, I certainly found AI useful when I figured out

**[00:06:01 – 00:06:03]** that it could make beer for us.

**[00:06:03 – 00:06:04]** I mean, is that normal?

**[00:06:04 – 00:06:07]** So, you know, AI has become useful.

**[00:06:07 – 00:06:09]** And let's understand what we mean when we say that.

**[00:06:09 – 00:06:12]** What's happening with the GenTech AI today?

**[00:06:12 – 00:06:16]** And there's a couple of things that make up an agent, right?

**[00:06:16 – 00:06:19]** You have these models that are very good at doing that

**[00:06:19 – 00:06:19]** reasoning, right?

**[00:06:19 – 00:06:22]** They're a foundational part to be able to do that reasoning.

**[00:06:22 – 00:06:26]** But it's not just a model alone that equates an agent.

**[00:06:26 – 00:06:29]** It is a model plus a harness.

**[00:06:29 – 00:06:32]** Now, a harness is that set of tooling that resides

**[00:06:32 – 00:06:33]** around the model.

**[00:06:33 – 00:06:37]** It's also the orchestration that goes around the model.

**[00:06:37 – 00:06:41]** So the APIs that it can call, the data sources that it has

**[00:06:41 – 00:06:47]** access to, so that as it goes through its loop, it's able

**[00:06:47 – 00:06:51]** to make those calls, access those databases, execute that code.

**[00:06:52 – 00:06:56]** Now, what you notice in that last video Whereas all the prompts

**[00:06:56 – 00:07:00]** were about an outcome, right, that the user wanted, right?

**[00:07:00 – 00:07:04]** So today's long-running agents, all you do is provide them

**[00:07:04 – 00:07:06]** an outcome that you want, right?

**[00:07:06 – 00:07:10]** Maybe you prescribe some tools that you wanted to use, and

**[00:07:10 – 00:07:13]** the agent will go for one hour, three hours, four hours,

**[00:07:14 – 00:07:16]** right, to accomplish that task.

**[00:07:16 – 00:07:19]** And so agents are models.

**[00:07:19 – 00:07:22]** These powerful reasoning models, not just a model, they're

**[00:07:22 – 00:07:25]** a system of models, and a harness, which represents the tooling, the

**[00:07:25 – 00:07:30]** orchestration that sits around it, to be able to accomplish that task.

**[00:07:30 – 00:07:33]** So what is our contribution to that environment?

**[00:07:33 – 00:07:37]** Well, it's the NVIDIA Agent Toolkit.

**[00:07:37 – 00:07:42]** And this represents different parts across our organization and

**[00:07:42 – 00:07:46]** across our ecosystem for how we're enabling the ecosystem, frankly,

**[00:07:46 – 00:07:49]** to build these long-running agents.

**[00:07:49 – 00:07:51]** And we're going to go through all of these, but at its

**[00:07:51 – 00:07:54]** start are the models.

**[00:07:54 – 00:07:58]** Jensen announced before yesterday, Nemotron Ultra, which is part

**[00:07:58 – 00:08:01]** of our Nemotron 3 family of models.

**[00:08:01 – 00:08:05]** And again, as I said earlier, having open models, having

**[00:08:05 – 00:08:09]** Cloud Frontier models, agents are made up of systems of models.

**[00:08:09 – 00:08:12]** But what's really special about Nemotron Ultra is that

**[00:08:13 – 00:08:18]** it was trained on the ecosystem of harnesses that exist out there.

**[00:08:19 – 00:08:21]** And OpenClaw is certainly one of them.

**[00:08:21 – 00:08:24]** One of my personal favorite one is one called Hermes by

**[00:08:24 – 00:08:27]** the news research team.

**[00:08:27 – 00:08:29]** And I think I see my friend Dylan over here.

**[00:08:29 – 00:08:31]** Say hi to Dylan afterwards, right?

**[00:08:31 – 00:08:34]** We did this with OpenCode, OpenHands as well,

**[00:08:35 – 00:08:36]** LangChain, Deep Agents.

**[00:08:36 – 00:08:40]** We wanted to make sure that Nemotron is great at using that

**[00:08:40 – 00:08:43]** tooling, at being able to drive that orchestration that's required

**[00:08:43 – 00:08:45]** to be able to accomplish that task.

**[00:08:45 – 00:08:48]** And the last thing that we did and contributed here is

**[00:08:48 – 00:08:51]** we took all of NVIDIA's libraries and turned them as skills.

**[00:08:51 – 00:08:56]** Skills are, right, in the past, developers, humans, used to

**[00:08:56 – 00:08:59]** use the libraries that NVIDIA made.

**[00:08:59 – 00:09:02]** Today, these are made for agents, and so skills is how we enable

**[00:09:02 – 00:09:05]** them to be utilized for agents, and we'll talk a bit about that.

**[00:09:05 – 00:09:08]** And then the final piece that's at the bottom is

**[00:09:08 – 00:09:10]** what we call the runtime.

**[00:09:10 – 00:09:12]** The runtime.

**[00:09:12 – 00:09:15]** You saw how powerful some of these agents get.

**[00:09:15 – 00:09:18]** Runtime ensures they're only able to access the data that

**[00:09:18 – 00:09:20]** you let them access.

**[00:09:20 – 00:09:24]** They're only able to execute on the actions that you define.

**[00:09:24 – 00:09:26]** We're going to unpack every one of those as we go through

**[00:09:26 – 00:09:28]** this presentation today.

**[00:09:28 – 00:09:30]** Okay, Nemotron models.

**[00:09:31 – 00:09:33]** The Nemotron models, fantastic models.

**[00:09:33 – 00:09:35]** You know, somebody told me this earlier.

**[00:09:35 – 00:09:39]** They said, Nemotron is the most open model in the world.

**[00:09:39 – 00:09:40]** And I said, what is the most open?

**[00:09:40 – 00:09:43]** You're either open or you're not open, right?

**[00:09:43 – 00:09:47]** But most open means we don't just provide the checkpoints and

**[00:09:47 – 00:09:49]** make them available on HuggingFace.

**[00:09:49 – 00:09:51]** We provide the pre-training data sets.

**[00:09:51 – 00:09:54]** We provide the post-training data sets.

**[00:09:54 – 00:09:57]** We publish the papers for the hybrid architectures

**[00:09:57 – 00:09:59]** that people create.

**[00:09:59 – 00:10:03]** We also provide the algorithms through which we train our models.

**[00:10:03 – 00:10:06]** And so when we talk about being the most open,

**[00:10:06 – 00:10:08]** That's what Nemotron represents.

**[00:10:08 – 00:10:11]** And that's super critical because, yes, you want to

**[00:10:11 – 00:10:12]** be able to customize it.

**[00:10:12 – 00:10:16]** And sometimes you might decide to take our pre-training data sets and

**[00:10:16 – 00:10:17]** build your own model from there.

**[00:10:17 – 00:10:20]** But I think the other very important point is when

**[00:10:20 – 00:10:22]** something is open, when it's got that much transparency,

**[00:10:22 – 00:10:25]** it allows us to trust it.

**[00:10:25 – 00:10:27]** It allows us to trust it. And I think that's super critical.

**[00:10:27 – 00:10:29]** And so Nemotron comes in different families.

**[00:10:29 – 00:10:31]** The latest is Ultra.

**[00:10:31 – 00:10:33]** I read somewhere that it's five times faster.

**[00:10:33 – 00:10:39]** We use NVFP4 to make sure it's got great speed and it's got

**[00:10:39 – 00:10:43]** 30% lower cost, comparable to other open models that are out there.

**[00:10:43 – 00:10:45]** And we have models that are very different domains as well.

**[00:10:45 – 00:10:48]** We have some for retrieval augmented generation, some

**[00:10:48 – 00:10:53]** new speech-to-speech models, fully duplex models that are fantastic,

**[00:10:53 – 00:10:56]** and some models that are great for guard railing, ensuring we have

**[00:10:56 – 00:11:00]** content safety that are fantastic.

**[00:11:00 – 00:11:04]** And so, NeMoTron, Ultra is out there, I think tomorrow

**[00:11:04 – 00:11:06]** it'll be available.

**[00:11:06 – 00:11:09]** Suggest you guys all try it, it's a fantastic model.

**[00:11:09 – 00:11:11]** Let's talk about skills.

**[00:11:11 – 00:11:16]** And so when Jensen talks at GTC, the first thing, the

**[00:11:16 – 00:11:20]** first slide he presents is always the CUDA-X libraries.

**[00:11:20 – 00:11:22]** And those libraries span all sorts of different

**[00:11:22 – 00:11:27]** domains, from cuLithio, quantum libraries, video search

**[00:11:28 – 00:11:33]** and summarization, data processing.

**[00:11:33 – 00:11:36]** They span all sorts of different domains.

**[00:11:36 – 00:11:42]** And what we've done of late is, yes, the libraries still exist.

**[00:11:42 – 00:11:46]** Yes, those libraries can now be utilized as tools, right?

**[00:11:46 – 00:11:50]** But in addition to providing them as tools, we also have

**[00:11:50 – 00:11:53]** a skills, you know, we're packaging them up as skills.

**[00:11:53 – 00:11:57]** And at its most simplest way, you can think of a skill as

**[00:11:57 – 00:12:01]** the instruction set, the manual, and it's written in plain English.

**[00:12:01 – 00:12:06]** In plain language, in terms of what this tool, what this library is

**[00:12:06 – 00:12:08]** good for, what it's not good for.

**[00:12:08 – 00:12:12]** So when your agent is looking to accomplish a task that you define,

**[00:12:13 – 00:12:16]** it doesn't have to load the entire tool set in its context, it can

**[00:12:16 – 00:12:19]** read the manual to say, oh yeah.

**[00:12:19 – 00:12:22]** This is something, now I know how to use it.

**[00:12:22 – 00:12:26]** It's like, if I give my child, well, or a kid, if you source

**[00:12:26 – 00:12:29]** somebody a microwave for the first time, and you tell them,

**[00:12:29 – 00:12:31]** hey, figure out how to use that microwave, they're going

**[00:12:31 – 00:12:33]** to press all sorts of buttons.

**[00:12:34 – 00:12:36]** But if you give them a manual, right, with pictures of how

**[00:12:36 – 00:12:38]** to use that microwave, they're going to know how to use that

**[00:12:38 – 00:12:40]** microwave from that first time.

**[00:12:40 – 00:12:45]** And so we've done a lot to package our libraries as skills.

**[00:12:46 – 00:12:48]** Now let me tell you about one of my favorite skills.

**[00:12:48 – 00:12:51]** Because skills are not just a single library.

**[00:12:51 – 00:12:55]** In fact, we encapsulate, we package some of our best,

**[00:12:55 – 00:12:58]** what we call blueprints, as skills.

**[00:12:58 – 00:13:04]** AIQ, right, we say it's IQ, is the equivalent of having

**[00:13:04 – 00:13:09]** a research lab at your fingertips or at your claw tips, right?

**[00:13:09 – 00:13:13]** Deep research is foundational to everything that you do.

**[00:13:13 – 00:13:16]** It's a given in every agent that you utilize.

**[00:13:16 – 00:13:21]** Yes, we thought or we started with doing deep research for synthesis,

**[00:13:21 – 00:13:24]** market analysis, et cetera.

**[00:13:24 – 00:13:27]** But if you think about it, if you're doing triaging, right,

**[00:13:27 – 00:13:30]** of you want to understand what the issue is, you know, like a support

**[00:13:30 – 00:13:31]** ticket, you do deep research.

**[00:13:31 – 00:13:34]** You have to go to different data sources to understand it.

**[00:13:34 – 00:13:36]** Even in coding, you're doing deep research.

**[00:13:36 – 00:13:38]** So is it a skill?

**[00:13:38 – 00:13:42]** It is a capability that we believe every agent needs it.

**[00:13:42 – 00:13:48]** So we invested a lot in making this world-class, and it is world-class.

**[00:13:48 – 00:13:51]** When we first published this about two months ago, which again

**[00:13:51 – 00:13:54]** feels like a lifetime ago, this was number one on Deep Research

**[00:13:54 – 00:13:56]** Bench 1 and Deep Research Bench 2.

**[00:13:56 – 00:13:59]** And the reason why I like to spend some time on this multi-agent

**[00:13:59 – 00:14:04]** based system, which is available as a skill, is because it really shows

**[00:14:04 – 00:14:09]** how a combination of open models like Nemotron and frontier models

**[00:14:09 – 00:14:15]** like Opus or GPT come together to provide the best outcomes, right?

**[00:14:15 – 00:14:20]** And so, again here, we use Nemotron Nano.

**[00:14:20 – 00:14:23]** We use Nemotron Nano as an intent-based router.

**[00:14:23 – 00:14:25]** Because not every time you need to do deep research.

**[00:14:25 – 00:14:29]** And so up front, right, that's the first model you interact with.

**[00:14:29 – 00:14:32]** It figures out whether, hey, is somebody just uploading

**[00:14:32 – 00:14:35]** a document because they want you to answer a question?

**[00:14:35 – 00:14:37]** It routes it to shallow research.

**[00:14:37 – 00:14:39]** When it figures out you got to do deep research, that's

**[00:14:39 – 00:14:43]** when it takes it to the deep research team of agents.

**[00:14:43 – 00:14:45]** And there's a team of agents here.

**[00:14:45 – 00:14:46]** Our orchestrator and our planners.

**[00:14:47 – 00:14:51]** An orchestrator is like a project manager, a team leader, right?

**[00:14:51 – 00:14:54]** And the planner that's next to them is somebody whose

**[00:14:54 – 00:14:59]** sole job is to just write the tasks down, and as you accomplish

**[00:14:59 – 00:15:01]** those tasks, cross them out.

**[00:15:01 – 00:15:06]** For those two, we used GPT 5.2 or 5.3 at the time.

**[00:15:06 – 00:15:08]** We also did it with Opus as well.

**[00:15:09 – 00:15:10]** Orchestration is important, right?

**[00:15:10 – 00:15:13]** And making sure that when you accomplish tasks, you're

**[00:15:13 – 00:15:14]** not hallucinating, right?

**[00:15:14 – 00:15:16]** You're actually crossing them out was also important, so

**[00:15:16 – 00:15:18]** we used those models.

**[00:15:18 – 00:15:23]** But for our sub-researchers, we used Nemotron 3 supermodels

**[00:15:23 – 00:15:25]** that we post-trained.

**[00:15:25 – 00:15:29]** And each one of those, there's no less than five or six

**[00:15:29 – 00:15:31]** different sub-researchers.

**[00:15:31 – 00:15:33]** And each one has got a different personality

**[00:15:33 – 00:15:37]** or a different discipline that they specialized on.

**[00:15:37 – 00:15:42]** One only gathers evidence and facts from authoritative sources.

**[00:15:42 – 00:15:46]** One other sub-researcher is very good at critiquing, right?

**[00:15:46 – 00:15:49]** One is good at giving counter-arguments.

**[00:15:49 – 00:15:53]** One is good at looking at a horizon and not making predictions,

**[00:15:53 – 00:15:56]** but trying to anticipate, kind of looking ahead, right?

**[00:15:56 – 00:15:59]** And then the synthesizer at the end kind of brings it all together.

**[00:16:00 – 00:16:04]** These are all sub-agents with their own context that specialize

**[00:16:04 – 00:16:06]** in doing one thing, right?

**[00:16:06 – 00:16:09]** And then the entire system, right, the synthesizer is

**[00:16:09 – 00:16:11]** what brings them all together.

**[00:16:11 – 00:16:14]** And it was a lot of, you know, it was a lot of context management.

**[00:16:14 – 00:16:16]** But this is what it took for us to win deep research, you

**[00:16:16 – 00:16:19]** know, these benchmarks.

**[00:16:19 – 00:16:21]** And not only were we able to be number one, we were

**[00:16:21 – 00:16:26]** able to do it at a fraction of the cost, 50% of the inference cost.

**[00:16:26 – 00:16:30]** And, like I said, this is not just about doing literature,

**[00:16:30 – 00:16:32]** right, like kind of research.

**[00:16:32 – 00:16:34]** Look at one of our partners, ServiceNow,

**[00:16:34 – 00:16:37]** and how they've adopted this.

**[00:16:43 – 00:16:45]** A new wave of AI has arrived.

**[00:16:45 – 00:16:50]** AI can go beyond answering what's in-house and help do.

**[00:16:51 – 00:16:55]** When enterprise service operations see a spike in ticket volumes,

**[00:16:55 – 00:17:02]** backlogs grow, response times slip, actual work must be accelerated.

**[00:17:02 – 00:17:06]** At ServiceNow, AI agents take on specific roles to handle service

**[00:17:06 – 00:17:09]** tickets from start to finish.

**[00:17:09 – 00:17:10]** Let's take a look.

**[00:17:10 – 00:17:14]** First, Specialized Agents, each with a unique skill,

**[00:17:14 – 00:17:18]** are set up and put to work.

**[00:17:19 – 00:17:24]** As tickets come in, case agents check the details and screenshots,

**[00:17:24 – 00:17:29]** then pass the context to NVIDIA Deep Research agents.

**[00:17:31 – 00:17:36]** Triage agents look through past fixes, similar cases, logs

**[00:17:37 – 00:17:43]** and guides, pinpointing the issue and finding the best way to fix it.

**[00:17:47 – 00:17:51]** Resolution agents confirm the root cause, apply the

**[00:17:51 – 00:17:54]** fix, and update the case record.

**[00:17:54 – 00:17:58]** When escalation is required, engineers receive a fully

**[00:17:58 – 00:18:04]** prepared case, including analysis, logs, and recommended actions.

**[00:18:04 – 00:18:08]** ServiceNow agents are built on a system of models, including

**[00:18:08 – 00:18:13]** Cloud Frontier models, ServiceNow's APRIEL, and NVIDIA NEMOTRON,

**[00:18:13 – 00:18:16]** using the NVIDIA iCUE Blueprint.

**[00:18:17 – 00:18:18]** The data is clear.

**[00:18:19 – 00:18:23]** 90% of tickets are resolved by autonomous agents, freeing

**[00:18:23 – 00:18:28]** engineers to focus on the hardest problems.

**[00:18:31 – 00:18:34]** I don't know if you guys heard that, but they said 90% of

**[00:18:34 – 00:18:37]** L1 tickets are resolved by these autonomous agents.

**[00:18:37 – 00:18:38]** And that's really great.

**[00:18:38 – 00:18:42]** Have you ever had a meeting and for whatever reason your WebEx, your

**[00:18:42 – 00:18:44]** Zoom, or your Teams is not working?

**[00:18:44 – 00:18:47]** Imagine having to file a ticket and wait for a human on the

**[00:18:47 – 00:18:49]** other end to give you access when your call is about to

**[00:18:49 – 00:18:51]** start in five minutes, right?

**[00:18:51 – 00:18:54]** These are the kind of problems it solved for.

**[00:18:54 – 00:18:58]** And so again, just an example of how IQs use and how all

**[00:18:58 – 00:19:02]** these autonomous agents are made It's made up of systems of models.

**[00:19:02 – 00:19:05]** My last couple of slides over here before I turn it over to Chris.

**[00:19:05 – 00:19:07]** Skills are critical, right?

**[00:19:07 – 00:19:10]** Again, they package and they're a means to utilize

**[00:19:10 – 00:19:13]** NVIDIA's CUDAx libraries.

**[00:19:14 – 00:19:17]** One thing that we've identified as critical to work with the

**[00:19:17 – 00:19:21]** ecosystem with is what we call verified skills.

**[00:19:21 – 00:19:23]** Because you have a lot of libraries, some of those libraries

**[00:19:23 – 00:19:27]** pull on containers, different packages, you know, there could

**[00:19:27 – 00:19:29]** be vulnerabilities inside of those.

**[00:19:29 – 00:19:32]** So how do you actually scan skills to make sure they don't

**[00:19:32 – 00:19:34]** have vulnerability inside?

**[00:19:34 – 00:19:37]** How do you evaluate skills such that, hey, a skill working

**[00:19:37 – 00:19:40]** with one harness or one agent doesn't degrade with another

**[00:19:40 – 00:19:43]** agent or doesn't degrade when another model is there?

**[00:19:43 – 00:19:45]** How do you sign skills so that if you're working in

**[00:19:46 – 00:19:49]** an enterprise, you don't want them just downloading skills

**[00:19:49 – 00:19:50]** from anywhere, right?

**[00:19:50 – 00:19:52]** You want to be able to at runtime enforce that only

**[00:19:52 – 00:19:54]** the signed skills are running.

**[00:19:54 – 00:19:56]** And so we've been partnering with the ecosystem, we've

**[00:19:56 – 00:19:59]** been partnering with our friends at NOOS over there, the Hermes guys.

**[00:19:59 – 00:20:02]** We've been partnering with Claude, with Anthropic and

**[00:20:02 – 00:20:06]** Codex to make our skills available in their respective hubs,

**[00:20:06 – 00:20:11]** our verified skills available in their respective hubs.

**[00:20:11 – 00:20:12]** That's an example.

**[00:20:12 – 00:20:14]** Do you recognize that portal?

**[00:20:14 – 00:20:18]** Dylan, right? So as of today, like we already see, right?

**[00:20:18 – 00:20:19]** IQ is already there.

**[00:20:19 – 00:20:20]** You can deploy it.

**[00:20:20 – 00:20:22]** You can use it for research.

**[00:20:22 – 00:20:26]** All of our skills are available with the world's harnesses.

**[00:20:26 – 00:20:28]** And so super exciting times.

**[00:20:28 – 00:20:31]** Now I'm going to turn it over to Chris Murphy to take you through

**[00:20:31 – 00:20:32]** the rest of the presentation.

**[00:20:32 – 00:20:34]** Thanks, guys.

**[00:20:36 – 00:20:38]** Go, Chris.

**[00:20:39 – 00:20:40]** Awesome, all right.

**[00:20:41 – 00:20:42]** Thank you very much, Adil.

**[00:20:42 – 00:20:45]** So I'm Chris Murphy, I'm a product manager.

**[00:20:45 – 00:20:47]** I work for Adil, though he will deny that.

**[00:20:47 – 00:20:51]** And I am the PM for NemoClaw.

**[00:20:51 – 00:20:55]** So yeah, I wanted to talk to you guys a little bit about, you

**[00:20:55 – 00:20:59]** know, Adels talked about how much these agents have these powerful

**[00:20:59 – 00:21:01]** things, they can accomplish great things, they go and do

**[00:21:01 – 00:21:04]** these things on your behalf, right?

**[00:21:04 – 00:21:08]** But right after the Open CLAW moment, we also had this,

**[00:21:08 – 00:21:11]** like, learning afterwards where we started seeing all

**[00:21:11 – 00:21:12]** these articles about

**[00:21:12 – 00:21:16]** Oh, but, you know, this leaked my passcodes.

**[00:21:16 – 00:21:18]** It stole my credit card information.

**[00:21:18 – 00:21:21]** Somebody prompt injected it, and it did all kinds

**[00:21:21 – 00:21:23]** of dangerous things, right?

**[00:21:24 – 00:21:28]** So how can we trust these agents when they have this much

**[00:21:28 – 00:21:32]** power, when they can self-evolve and learn new things on the

**[00:21:32 – 00:21:35]** fly without you watching, right?

**[00:21:35 – 00:21:37]** How can we trust them?

**[00:21:37 – 00:21:40]** So NVIDIA created OpenShell.

**[00:21:40 – 00:21:43]** OpenShell is an open source project.

**[00:21:43 – 00:21:47]** You can go out to the GitHub, you can contribute, you can

**[00:21:47 – 00:21:51]** look at all the code, and you can see the entire ecosystem there

**[00:21:51 – 00:21:53]** that is building this with us.

**[00:21:53 – 00:21:56]** This isn't just a NVIDIA project.

**[00:21:56 – 00:21:59]** This isn't just us putting code out on the internet.

**[00:21:59 – 00:22:03]** This is a vibrant open source community that the ecosystem

**[00:22:03 – 00:22:06]** is really rallying around because this was a problem.

**[00:22:07 – 00:22:11]** These agents are so powerful, but you need to be able to secure them.

**[00:22:11 – 00:22:15]** So you need a secure runtime to help you be in control of

**[00:22:15 – 00:22:18]** what your agent can and cannot do.

**[00:22:18 – 00:22:22]** So one of the key things about OpenShell is it allows you

**[00:22:22 – 00:22:28]** to be in control of the policies to be able to approve what

**[00:22:28 – 00:22:31]** your agent can and cannot do.

**[00:22:31 – 00:22:34]** You have different layers of egress, all the things

**[00:22:34 – 00:22:36]** that the agent can touch, right?

**[00:22:36 – 00:22:37]** You have control over.

**[00:22:37 – 00:22:42]** So can your agent search the Internet?

**[00:22:42 – 00:22:45]** Can it search every single site on the Internet, right?

**[00:22:45 – 00:22:49]** You want to be able to control if there's only certain sites, if

**[00:22:49 – 00:22:51]** you want to let it do everything, do you want it to be able

**[00:22:52 – 00:22:54]** to access your e-mail?

**[00:22:54 – 00:22:57]** If it accesses your e-mail, do you only want it to be

**[00:22:57 – 00:22:58]** able to read your e-mail?

**[00:22:58 – 00:23:02]** Or do you want to be able to send email on your behalf, right?

**[00:23:02 – 00:23:06]** I don't want my agent making those decisions for me.

**[00:23:06 – 00:23:11]** I want to be able to control what it can and it cannot do, right?

**[00:23:11 – 00:23:13]** And OpenShell allows you to do that.

**[00:23:14 – 00:23:17]** It allows you to differentiate between the things the agent is

**[00:23:17 – 00:23:22]** in control of and the things that you as the user are in control of.

**[00:23:23 – 00:23:28]** One example, keys, secrets, passcodes, right?

**[00:23:29 – 00:23:32]** If you have an agent and it needs to go log into something

**[00:23:32 – 00:23:35]** for you to do, whether it be a messaging service, if

**[00:23:35 – 00:23:38]** you want it to log into WhatsApp, you know, do you just give it

**[00:23:38 – 00:23:42]** your login and now somebody could potentially prompt inject your

**[00:23:42 – 00:23:45]** agent and steal that information?

**[00:23:45 – 00:23:47]** I don't want my agent to know that.

**[00:23:47 – 00:23:51]** One of the things OpenShell does is it actually moves

**[00:23:51 – 00:23:55]** all of those keys and all that secret management outside of

**[00:23:55 – 00:24:00]** the sandbox, outside of the shell that your agent is running it.

**[00:24:00 – 00:24:03]** And then, you know, when it makes the call, when you make a call

**[00:24:03 – 00:24:07]** and it goes outside of the sandbox, then OpenShell injects that key.

**[00:24:07 – 00:24:11]** So it is managed in an environment you can set those keys.

**[00:24:11 – 00:24:15]** But your agent can't ever see the real key, the real secret, right?

**[00:24:15 – 00:24:18]** So there's no chance that that agent could leak that

**[00:24:18 – 00:24:23]** information outside of the agent environment, right?

**[00:24:24 – 00:24:27]** Now, this isn't just for OpenClaw.

**[00:24:27 – 00:24:29]** This isn't just for Hermes, right?

**[00:24:29 – 00:24:33]** You can run any agent inside of OpenShell.

**[00:24:34 – 00:24:36]** That's one of the beauties here, is we're trying to embrace

**[00:24:36 – 00:24:38]** the entire ecosystem.

**[00:24:38 – 00:24:41]** We're building this to be able to run anywhere.

**[00:24:41 – 00:24:42]** You can run it on a DGX Spark.

**[00:24:42 – 00:24:46]** You can run it on huge data centers, right?

**[00:24:46 – 00:24:48]** So OpenShell runs anywhere.

**[00:24:48 – 00:24:51]** And part of how we're enabling that is we're partnering with

**[00:24:51 – 00:24:56]** all of these operating system partners to build OpenShell

**[00:24:56 – 00:24:59]** into the operating system, right?

**[00:24:59 – 00:25:01]** So you'll see this week we've seen a bunch of announcements

**[00:25:01 – 00:25:06]** from OpenShell, from Canonical, and from Microsoft for Windows

**[00:25:06 – 00:25:10]** to bring OpenShell into the operating system, right?

**[00:25:10 – 00:25:13]** So that any agent you run, you can have this layer of

**[00:25:13 – 00:25:19]** control, you can have this secure runtime for your agents.

**[00:25:21 – 00:25:23]** So then what's NeMoClaw?

**[00:25:24 – 00:25:27]** We talked about OpenShell, so what is NeMoClaw?

**[00:25:27 – 00:25:31]** NeMoClaw is a set of blueprints.

**[00:25:31 – 00:25:36]** NeMoClaw is the example, the reference that you can use

**[00:25:36 – 00:25:42]** that gets you started quickly with combining your agent,

**[00:25:42 – 00:25:47]** Your OpenShell, Sandbox, Secure Runtime, and models, and tools,

**[00:25:47 – 00:25:49]** and optimizations, right?

**[00:25:49 – 00:25:53]** So if you go to the NemoClaw GitHub, also completely open

**[00:25:53 – 00:25:58]** source, completely developed in the open, all of our code is out there.

**[00:25:58 – 00:26:03]** We take issues and PRs in the community and would love

**[00:26:03 – 00:26:08]** contributions, feedback, join us in the Discord, right?

**[00:26:08 – 00:26:12]** But we have a new blueprint that we just came out with recently.

**[00:26:12 – 00:26:15]** You heard Adil already talking about Noose Research and Hermes.

**[00:26:15 – 00:26:18]** So, when we first launched NeMoClaw,

**[00:26:18 – 00:26:22]** The only agent example we had at day zero was OpenClaw.

**[00:26:22 – 00:26:26]** We've now added Hermes Agent as another example blueprint.

**[00:26:26 – 00:26:29]** So you can get started in the NeMoClaw repo.

**[00:26:29 – 00:26:32]** It will deploy your Hermes Agent.

**[00:26:32 – 00:26:35]** It will deploy the OpenShell sandbox.

**[00:26:35 – 00:26:38]** It will give you the ability to manage those policies and show you

**[00:26:38 – 00:26:44]** and teach you how to use an agent inside this OpenShell environment.

**[00:26:44 – 00:26:47]** Make it really easy to connect it to, you know, whether you have

**[00:26:47 – 00:26:50]** a local model, if you're running on like a Spark, you can have a local

**[00:26:50 – 00:26:55]** model, or if you want to connect to an API endpoint and use Opus,

**[00:26:55 – 00:26:57]** you know, use other models, right?

**[00:26:57 – 00:26:58]** All of that is supported, right?

**[00:26:58 – 00:27:03]** This isn't baking you into our models, but it is allowing

**[00:27:03 – 00:27:07]** you to easily get started with them if you want open models.

**[00:27:09 – 00:27:13]** All right, so we talked about OpenShell, we talked about

**[00:27:13 – 00:27:15]** the policies, right?

**[00:27:15 – 00:27:17]** And we wanted to show you what this looks like, right?

**[00:27:17 – 00:27:21]** So this...

**[00:27:23 – 00:27:25]** Oops, one too many.

**[00:27:25 – 00:27:28]** It was slow. Now I gotta go back.

**[00:27:30 – 00:27:33]** I don't know if it's...

**[00:27:37 – 00:27:40]** I'm not being patient enough is what you're saying, Adel, right?

**[00:27:40 – 00:27:43]** Never said that.

**[00:27:43 – 00:27:47]** I'm saying that. There you go.

**[00:27:48 – 00:27:50]** Well, that is what I did. Okay.

**[00:27:50 – 00:27:54]** Here we go. So this is the user interface where you can

**[00:27:54 – 00:27:59]** go and see, you know, exactly what's pending rule.

**[00:27:59 – 00:28:02]** This is something your agent tried to do and it wasn't

**[00:28:02 – 00:28:04]** authorized to do it.

**[00:28:04 – 00:28:08]** So the agent couldn't access, it couldn't do the thing, right?

**[00:28:08 – 00:28:12]** But you have the ability to come in here and review as

**[00:28:12 – 00:28:16]** the user and see if you want to now give it that authority, right?

**[00:28:16 – 00:28:19]** Because you probably, when you deploy your agent, you probably

**[00:28:19 – 00:28:23]** don't know on day zero everything you're going to want it to do.

**[00:28:23 – 00:28:26]** You're going to want to be able to add new policies, right?

**[00:28:26 – 00:28:30]** Approve new things that it can access so it can go to GitHub.

**[00:28:30 – 00:28:34]** And it can go look at a new repo that you want to investigate

**[00:28:34 – 00:28:36]** and ask it to do tasks on, right?

**[00:28:36 – 00:28:40]** So by using the user interface of OpenShell, that gives you

**[00:28:40 – 00:28:43]** the power to go see everything it's trying to access.

**[00:28:43 – 00:28:46]** And you can decide, do I want to access

**[00:28:46 – 00:28:47]** Do I want to let it access?

**[00:28:47 – 00:28:49]** Do I not want to let it access?

**[00:28:49 – 00:28:52]** Is it just going to be read or is it going to be, you know, do I

**[00:28:52 – 00:28:54]** want it to be able to write, right?

**[00:28:54 – 00:28:57]** Do I only want to be able to gather information from my GitHub repo?

**[00:28:57 – 00:29:01]** Or do I want my agent to be able to create PRs and submit

**[00:29:01 – 00:29:03]** PRs to those repos?

**[00:29:03 – 00:29:07]** Do I trust my agent with this data source, right?

**[00:29:07 – 00:29:11]** And some of whether you trust your agent to do certain things

**[00:29:11 – 00:29:15]** depends on what else you've given it permission to do, right?

**[00:29:15 – 00:29:19]** If my agent only has access to things that are out

**[00:29:19 – 00:29:23]** on the Internet, it's all publicly available data,

**[00:29:23 – 00:29:26]** I can be pretty loose with what I let it do.

**[00:29:26 – 00:29:29]** I can be pretty permissive, right?

**[00:29:29 – 00:29:36]** But if my agent has access to my corporate documentation,

**[00:29:36 – 00:29:41]** my internal secrets of my company, now do I want that

**[00:29:41 – 00:29:44]** same agent to be able to go write something out to GitHub?

**[00:29:44 – 00:29:47]** What if it made a mistake and it leaked some data from

**[00:29:47 – 00:29:52]** my corporate documents and posted it in a PR out to GitHub?

**[00:29:52 – 00:29:56]** So maybe once I give it access to my internal corporate secrets,

**[00:29:56 – 00:29:59]** maybe that agent shouldn't be able to write to GitHub.

**[00:29:59 – 00:30:03]** So it's not like a blanket, you always know what you would want

**[00:30:03 – 00:30:04]** or what you trust an agent to do.

**[00:30:05 – 00:30:07]** It depends on the use case, and it depends on what else

**[00:30:07 – 00:30:09]** you've given it permission to do.

**[00:30:09 – 00:30:14]** And the OpenShell allows you to control that, to have that

**[00:30:14 – 00:30:16]** fine-grained control.

**[00:30:18 – 00:30:21]** So I said we had blueprints, right?

**[00:30:21 – 00:30:24]** Why is it a blueprint and not an application?

**[00:30:24 – 00:30:27]** Why do we not just say, here is the NemoClaw application.

**[00:30:28 – 00:30:29]** Go and deploy it.

**[00:30:30 – 00:30:35]** Because the key is we're not trying to make the end solution.

**[00:30:35 – 00:30:39]** We are trying to provide you with a starting place that

**[00:30:39 – 00:30:41]** you can customize from.

**[00:30:41 – 00:30:44]** So Blueprints, all the code is out there, and

**[00:30:44 – 00:30:46]** it's completely customizable.

**[00:30:46 – 00:30:49]** You could change the harness.

**[00:30:49 – 00:30:51]** You can change the model it uses.

**[00:30:51 – 00:30:57]** You can change all of your policies around your OpenShell environment.

**[00:30:57 – 00:31:02]** And the key to making it powerful is connecting it to the data

**[00:31:02 – 00:31:06]** sources, the APIs, and the systems that you need it to be

**[00:31:06 – 00:31:09]** able to access for your use case.

**[00:31:09 – 00:31:13]** That is how you make an agent powerful.

**[00:31:13 – 00:31:15]** It's not just what comes out of the box.

**[00:31:15 – 00:31:17]** It's what you customize it to do.

**[00:31:17 – 00:31:20]** And so our blueprints, they're not end products.

**[00:31:21 – 00:31:23]** You know, they're not an application, they're

**[00:31:23 – 00:31:25]** a starting point, right?

**[00:31:25 – 00:31:29]** And so what we want you to do is go use these blueprints, and you can

**[00:31:29 – 00:31:32]** deploy them, there's an installer, you'll have a fully functional

**[00:31:32 – 00:31:35]** Hermes agent if you use the Hermes blueprint, and it can do a lot

**[00:31:35 – 00:31:37]** of things out of the box, right?

**[00:31:38 – 00:31:40]** But then you can customize and build from there.

**[00:31:40 – 00:31:45]** You can give it access, you can do new policies.

**[00:31:46 – 00:31:49]** And so, we've done this at NVIDIA.

**[00:31:49 – 00:31:53]** This is something that Adil and I actively have running

**[00:31:53 – 00:31:55]** for ourselves, right?

**[00:31:55 – 00:32:00]** We are building a Hermes-based agent and deploying it at

**[00:32:00 – 00:32:03]** NVIDIA for us to use.

**[00:32:04 – 00:32:09]** I talked about that power is in customizing what it has access to.

**[00:32:09 – 00:32:13]** So some of the things that our agent that we have has access

**[00:32:13 – 00:32:18]** to is our internal NVIDIA Slack.

**[00:32:18 – 00:32:20]** And our Outlook email.

**[00:32:20 – 00:32:25]** In fact, my agent has its own email address.

**[00:32:25 – 00:32:28]** I have a new email address that is my agent's email

**[00:32:28 – 00:32:33]** address and it will only respond to emails from me.

**[00:32:33 – 00:32:36]** So you could know my email, my agent email, you could

**[00:32:36 – 00:32:38]** email it all day long till you're blue in the face and

**[00:32:38 – 00:32:40]** it won't do anything with that.

**[00:32:40 – 00:32:42]** But if I email it,

**[00:32:42 – 00:32:46]** It will then send any instructions or anything I ask it to do

**[00:32:46 – 00:32:47]** to my Hermes agent.

**[00:32:47 – 00:32:51]** So I can just email my agent and give it stuff to do.

**[00:32:51 – 00:32:54]** I can also set it up to have a job that it runs every day.

**[00:32:54 – 00:32:59]** It can summarize my emails and give me the most important things to do.

**[00:32:59 – 00:33:01]** It can go look through the NemoClaw GitHub and alert

**[00:33:01 – 00:33:05]** me if there's any critical issues that I need to be aware of.

**[00:33:05 – 00:33:07]** And I can set it up to send me an email every day if I

**[00:33:07 – 00:33:10]** want, if that's how I want to interact with it.

**[00:33:10 – 00:33:14]** It also has access to Slack, so I can talk to it in Slack.

**[00:33:14 – 00:33:17]** I have a channel where I can talk just to my agent.

**[00:33:17 – 00:33:20]** I've set mine so it will only respond to me or only take

**[00:33:20 – 00:33:21]** instruction from me.

**[00:33:21 – 00:33:25]** I didn't give Adil permission to tell my agent what to do.

**[00:33:25 – 00:33:28]** He has to go through me first, you know?

**[00:33:28 – 00:33:31]** But our agents are in a channel together as well.

**[00:33:31 – 00:33:35]** So in that channel, if he asks me for something as my boss,

**[00:33:35 – 00:33:40]** I can turn around and be like, hey, Hermes Agent, go build me

**[00:33:40 – 00:33:44]** this dashboard that Adel's asking about, or go fetch the information

**[00:33:44 – 00:33:49]** for me, or go create a PR if I gave it permission to create PRs, right?

**[00:33:49 – 00:33:53]** Because I also have access to...

**[00:33:54 – 00:33:56]** The NeMoClaw GitHub, right?

**[00:33:56 – 00:33:59]** So it can go look at all the issues that are out there.

**[00:33:59 – 00:34:03]** It also has access to a bunch of community forums, right?

**[00:34:03 – 00:34:07]** Not everything that's relevant to my project is coming

**[00:34:07 – 00:34:08]** in through GitHub.

**[00:34:08 – 00:34:12]** Some of it's happening in Discord and in developer forums, right?

**[00:34:12 – 00:34:16]** So as a product manager, I need to be able to gather

**[00:34:16 – 00:34:21]** information across all of these sources to act upon it, right?

**[00:34:21 – 00:34:24]** So now I can use my agent that I have.

**[00:34:24 – 00:34:28]** To be aware of what's going on, to help prioritize the next things

**[00:34:28 – 00:34:30]** that I need to be working on.

**[00:34:30 – 00:34:32]** It's getting the community input.

**[00:34:32 – 00:34:36]** It's getting any emails from my leadership or my team members

**[00:34:36 – 00:34:37]** as they're emailing me.

**[00:34:38 – 00:34:42]** In Slack, I have it added to my development Slack

**[00:34:42 – 00:34:44]** channel for NeMoClaw.

**[00:34:44 – 00:34:46]** So when a new issue comes in, if it alerts me and it's

**[00:34:46 – 00:34:49]** like, ooh, there's this new community thing happening, there's

**[00:34:49 – 00:34:53]** this issue, I can immediately have my agent go check, is the There's

**[00:34:53 – 00:34:54]** a team already on top of this.

**[00:34:54 – 00:34:56]** Is there already a fix in work?

**[00:34:56 – 00:34:59]** Are they already working on it, right?

**[00:34:59 – 00:35:03]** Just with this limited number of tools and data sources, my

**[00:35:03 – 00:35:06]** agent is now super, super powerful.

**[00:35:06 – 00:35:10]** But one thing it doesn't have, it doesn't have blind

**[00:35:10 – 00:35:11]** access to the Internet.

**[00:35:11 – 00:35:15]** It can't just go search the internet.

**[00:35:15 – 00:35:19]** Because it has access to everything that's in my email inbox,

**[00:35:19 – 00:35:22]** and I don't want any of that to get leaked, right?

**[00:35:22 – 00:35:26]** I don't want an email that I got from Justin Baitano,

**[00:35:26 – 00:35:31]** you know, a VP, I don't want something, you know, private to

**[00:35:31 – 00:35:36]** get leaked out onto GitHub, to get posted out on the internet, right?

**[00:35:36 – 00:35:39]** So I can read GitHub with my agent, but I haven't given

**[00:35:39 – 00:35:40]** it permission to write to GitHub.

**[00:35:40 – 00:35:43]** Because for this use case, I don't want it to have that

**[00:35:43 – 00:35:48]** because I've connected it to sensitive data sources.

**[00:35:48 – 00:35:53]** So this is actually my agent in Slack.

**[00:35:53 – 00:35:57]** This is my chat with just my agent, and I called it Chris's Hermes.

**[00:35:57 – 00:36:00]** I know I'm really, really creative.

**[00:36:00 – 00:36:02]** You can be more creative with your agent names, right?

**[00:36:03 – 00:36:07]** But this is something I did just yesterday, right?

**[00:36:07 – 00:36:10]** As a PM, I'm here in Taipei.

**[00:36:10 – 00:36:12]** I'm out of sync with my team.

**[00:36:12 – 00:36:15]** I haven't been calling into the team meetings because I'm 13

**[00:36:15 – 00:36:18]** hours off from where I normally am.

**[00:36:18 – 00:36:22]** But I wanted to know what's been going on in the last week.

**[00:36:22 – 00:36:26]** What are the top three priorities that I need to be aware of

**[00:36:26 – 00:36:28]** as the product manager?

**[00:36:28 – 00:36:30]** And so it went and did a search.

**[00:36:30 – 00:36:34]** It found a number of features that people are asking for.

**[00:36:34 – 00:36:38]** It gives me the links to exactly where it's finding this information

**[00:36:38 – 00:36:41]** so I could go and see in the forum or in the GitHub where

**[00:36:41 – 00:36:42]** they're being talked about.

**[00:36:42 – 00:36:47]** It tells me why it's important and how high of a priority it is.

**[00:36:47 – 00:36:51]** It recommends some actions for me.

**[00:36:51 – 00:36:53]** It recommends what I should do.

**[00:36:53 – 00:36:54]** Should I escalate?

**[00:36:54 – 00:36:57]** And then it asks me.

**[00:36:58 – 00:37:01]** What about next steps? Do you want me to do any of these next things?

**[00:37:01 – 00:37:03]** And I'm just interacting with it in Slack.

**[00:37:03 – 00:37:07]** My agent is running in a VM, and it's just doing its thing,

**[00:37:07 – 00:37:08]** and I can just talk to it.

**[00:37:08 – 00:37:10]** I could literally go on my phone right now and give

**[00:37:11 – 00:37:14]** my agent instructions of what I want it to do.

**[00:37:14 – 00:37:16]** Just in Slack.

**[00:37:21 – 00:37:25]** Okay, so that's what I have for my agent, right?

**[00:37:25 – 00:37:29]** But there's a couple more things to agents, right?

**[00:37:29 – 00:37:31]** Agents use a lot of tokens.

**[00:37:31 – 00:37:34]** Agents do a lot of things, right?

**[00:37:34 – 00:37:39]** You need to be able to understand what your agent is doing.

**[00:37:39 – 00:37:40]** Is it doing it efficiently?

**[00:37:40 – 00:37:42]** Is it calling the right tools?

**[00:37:42 – 00:37:46]** You know, is it doing something I don't want it to do, right?

**[00:37:46 – 00:37:48]** That's what we built NeMo Relay for.

**[00:37:48 – 00:37:52]** And NeMo Relay allows with any agent.

**[00:37:52 – 00:37:55]** You can use it with Hermes, you can use it with OpenClaw, you can use

**[00:37:55 – 00:37:58]** it with a number of agents, right?

**[00:37:58 – 00:38:02]** And it will observe...

**[00:38:02 – 00:38:06]** And optimize everything that it will learn and optimize

**[00:38:06 – 00:38:08]** your agents for you, right?

**[00:38:08 – 00:38:12]** So it can collect data about what your agents are doing, right?

**[00:38:12 – 00:38:15]** So it can collect it in OTEL format or ATIF format, which

**[00:38:15 – 00:38:17]** I really like ATIF because it means then when I look at the

**[00:38:18 – 00:38:21]** traces, I can actually understand what's going on in the traces.

**[00:38:21 – 00:38:23]** It's like human-readable agent traces.

**[00:38:23 – 00:38:27]** And it collects this information, but it's

**[00:38:27 – 00:38:29]** not just an observability tool.

**[00:38:29 – 00:38:31]** It doesn't just collect information.

**[00:38:31 – 00:38:35]** The goal of NeMo Relay is to be able to use the information

**[00:38:35 – 00:38:38]** it collects about what's going on in your agent to help you

**[00:38:39 – 00:38:41]** optimize those agents.

**[00:38:41 – 00:38:45]** So, for example, one of the things that the team has built with NeMo

**[00:38:45 – 00:38:49]** Relay is that it learns what you're doing with your agent over time.

**[00:38:49 – 00:38:53]** So, it learns, like, you know, if it's most likely to call the

**[00:38:53 – 00:38:56]** same tool, like, three times in a loop, or it's going to do this same

**[00:38:56 – 00:38:58]** thing over and over again, right?

**[00:38:58 – 00:39:01]** So, it can use that to understand...

**[00:39:01 – 00:39:05]** What it should tell a model like Claude to cache, right?

**[00:39:05 – 00:39:08]** Anthropic gives you the ability to control what you cache and what

**[00:39:08 – 00:39:10]** you don't when you use the model.

**[00:39:10 – 00:39:14]** And NeMo Relay can use the insights of having watched

**[00:39:14 – 00:39:18]** your agent over time to know when it should tell Anthropic to

**[00:39:18 – 00:39:20]** cache things or when it shouldn't.

**[00:39:20 – 00:39:23]** That can save you significantly amounts of money, right?

**[00:39:24 – 00:39:27]** We've had some experiments like 30% less token costs

**[00:39:27 – 00:39:30]** when you use it for certain workloads because you can be more

**[00:39:30 – 00:39:32]** efficient with your token costs.

**[00:39:32 – 00:39:38]** You can use it in combination with another NVIDIA library

**[00:39:38 – 00:39:42]** called Dynamo that manages like the KV cache and where

**[00:39:42 – 00:39:44]** things are running on your cluster.

**[00:39:44 – 00:39:48]** It's like an inference software layer for optimization.

**[00:39:48 – 00:39:51]** And NeMo Relay can feed hints to Dynamo.

**[00:39:51 – 00:39:53]** About what you should keep hot in your KV cache or what

**[00:39:53 – 00:39:55]** you're done with, right?

**[00:39:55 – 00:39:58]** Based on what it knows about how your agent works, right?

**[00:39:58 – 00:40:03]** So you can use all of this knowledge to be more efficient

**[00:40:03 – 00:40:04]** and optimize your agents.

**[00:40:05 – 00:40:07]** So it's not just observing.

**[00:40:07 – 00:40:10]** It's learning the behaviors and then it's using those

**[00:40:10 – 00:40:14]** learnings to optimize your agents.

**[00:40:15 – 00:40:20]** But you need even just the insights into what it's doing are super,

**[00:40:21 – 00:40:24]** super valuable for agents, right?

**[00:40:25 – 00:40:28]** So I just did this yesterday, two days ago.

**[00:40:28 – 00:40:31]** I connected the Phoenix dashboard to my agent, right?

**[00:40:31 – 00:40:34]** So this is connected to my actual agent that I was showing

**[00:40:34 – 00:40:36]** you working in Slack, right?

**[00:40:37 – 00:40:42]** And I had intentionally asked my agent to go and try to

**[00:40:42 – 00:40:44]** access something from the Internet.

**[00:40:44 – 00:40:45]** I told you it doesn't have.

**[00:40:45 – 00:40:47]** Web search, right?

**[00:40:47 – 00:40:51]** And I told it to go search the Internet for announcements

**[00:40:51 – 00:40:54]** and stuff around Noose and Hermes, right?

**[00:40:55 – 00:40:59]** And so one of the things I can see in this trace is the

**[00:40:59 – 00:41:01]** sandbox has restricted egress.

**[00:41:01 – 00:41:03]** It can't search the Internet.

**[00:41:03 – 00:41:07]** So not only did it block it, when I went into the OpenShell

**[00:41:07 – 00:41:11]** user interface, I could see that it had a pending review there.

**[00:41:12 – 00:41:14]** I could have given it access, but this is one of the things

**[00:41:14 – 00:41:15]** I don't want it to do.

**[00:41:15 – 00:41:17]** I don't want my agent to be able to go search the Internet.

**[00:41:17 – 00:41:20]** So I didn't allow it, and then I can go look in the trace

**[00:41:20 – 00:41:24]** and see exactly why that failed.

**[00:41:24 – 00:41:28]** There's so much data in your agent traces, every single

**[00:41:28 – 00:41:33]** tool call, every single LLM call, and you can dig into each one.

**[00:41:33 – 00:41:37]** You can see how much tokens each used, how much time each took.

**[00:41:37 – 00:41:40]** You can see how many iterations it's doing.

**[00:41:40 – 00:41:44]** And then you have human readable information to see when something

**[00:41:44 – 00:41:46]** fails, why it's failing.

**[00:41:46 – 00:41:50]** That is super powerful, both from debugging or understanding

**[00:41:50 – 00:41:55]** what your agent is doing, but also on that security side to understand

**[00:41:55 – 00:41:59]** what your agent's doing, to make sure it's not doing something

**[00:41:59 – 00:42:03]** you're not comfortable with or that you don't want it to be doing.

**[00:42:03 – 00:42:07]** Awesome, okay, we're down to three minutes.

**[00:42:07 – 00:42:09]** Okay, this next video I have is not, we're not gonna be

**[00:42:09 – 00:42:11]** able to see the whole thing.

**[00:42:12 – 00:42:15]** So, what I'm really just gonna say here is this is a teaser.

**[00:42:16 – 00:42:19]** This is a demo we have in Build-A-Claw on our

**[00:42:20 – 00:42:21]** showcase floor here.

**[00:42:21 – 00:42:26]** And this is just a visualization of agents, so each.

**[00:42:26 – 00:42:29]** Each lobster here is an OpenClaw agent.

**[00:42:29 – 00:42:32]** And when they're running around here, they're not in a sandbox.

**[00:42:32 – 00:42:35]** These are the sandboxes, right?

**[00:42:35 – 00:42:38]** And you can actually go and drag your lobsters into the

**[00:42:38 – 00:42:41]** sandbox and then see the difference of what they can and can't do when

**[00:42:41 – 00:42:45]** they're in the sandbox or not, have a visualization to the policies

**[00:42:45 – 00:42:47]** where you can approve them.

**[00:42:47 – 00:42:49]** We're not approve them, right?

**[00:42:49 – 00:42:52]** It's really a really cool just visualization of your

**[00:42:52 – 00:42:55]** agents working together and what they can and can't do and how

**[00:42:55 – 00:42:58]** OpenShell impacts their behavior.

**[00:42:58 – 00:43:01]** So definitely go to the Build-A-Claw, check this out.

**[00:43:01 – 00:43:02]** You can interact with it.

**[00:43:02 – 00:43:04]** You can have them create a lobster.

**[00:43:04 – 00:43:06]** You can name it. You can put a wizard hat on it.

**[00:43:06 – 00:43:09]** There's all kinds of fun things you can do.

**[00:43:09 – 00:43:10]** I'm just gonna skip ahead.

**[00:43:10 – 00:43:14]** And go to the how to get started, right?

**[00:43:14 – 00:43:17]** And let me step aside for anybody who wants to take pictures,

**[00:43:17 – 00:43:19]** because this is important, right?

**[00:43:19 – 00:43:22]** Ad will talk to you about the skills, right?

**[00:43:22 – 00:43:24]** So we have our skills out on GitHub.

**[00:43:25 – 00:43:29]** Build.NVIDIA.com, you can go see all of our models,

**[00:43:29 – 00:43:32]** Nemotron Ultra, Endpoints.

**[00:43:32 – 00:43:34]** You can use them for free, you can try them out.

**[00:43:34 – 00:43:36]** Starting tomorrow, Ultra will be there.

**[00:43:36 – 00:43:40]** On the fourth, well, tomorrow, US time, I guess, is when

**[00:43:40 – 00:43:42]** the Ultra is going to launch.

**[00:43:42 – 00:43:44]** But Super and Nano are already there.

**[00:43:44 – 00:43:47]** They're great getting started endpoints, and then when you're

**[00:43:48 – 00:43:51]** ready to do more, when you don't wanna be rate-limited anymore,

**[00:43:51 – 00:43:53]** then you can move to one of our partners, or you can download the

**[00:43:53 – 00:43:55]** weights and run them yourselves.

**[00:43:55 – 00:43:57]** Oh, because they're completely open.

**[00:43:57 – 00:44:00]** OpenShell's GitHub to learn more about OpenShell

**[00:44:00 – 00:44:04]** and deploy OpenShell independently if you want.

**[00:44:04 – 00:44:06]** And then the NemoClaw blueprints.

**[00:44:06 – 00:44:08]** So one thing I'll add.

**[00:44:09 – 00:44:12]** So this is the GitHub for NemoClaw, great place to learn

**[00:44:12 – 00:44:13]** more about NemoClaw.

**[00:44:13 – 00:44:18]** But if you go to build.NVIDIA.com, you can learn about the different

**[00:44:18 – 00:44:20]** blueprints we have for NemoClaw.

**[00:44:20 – 00:44:23]** But there's also a Try It Now button.

**[00:44:23 – 00:44:26]** When you click Try It Now, it will take you to Brev, and it will

**[00:44:26 – 00:44:29]** take you for a fairly low cost.

**[00:44:29 – 00:44:32]** You can run, you can get some compute to run these blueprints,

**[00:44:33 – 00:44:34]** and you can deploy these agents.

**[00:44:34 – 00:44:39]** So you don't have to have a GPU, You don't have to go purchase a VM.

**[00:44:39 – 00:44:45]** It's all built in to go try out either OpenClaw or Hermes

**[00:44:45 – 00:44:47]** running within OpenShell.

**[00:44:47 – 00:44:49]** And you can experiment with the policies and see the

**[00:44:49 – 00:44:53]** power of everything that we just talked about.

**[00:44:53 – 00:44:54]** And that's all I have.