May 26, 2026 · 33m · big-technology
The Right Way To Build AI Agents — With NVIDIA's Adel El Hallak and ServiceNow's Joe Davis
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Alex Kantrowitz hosts ServiceNow's Joe Davis and NVIDIA's Adel El Hallak to discuss the technical and operational blueprint for enterprise agentic AI. The discussion breaks down multi-model orchestration, deterministic zero-trust security harnesses, OpenShell runtime sandboxing, and real-world deployment across enterprise workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 28.5% of the talking time here. How this is scored →
Alex as informed peer 5.0 Guest teaching 4.1 Guest disagreement 0.9 Alex pushing back 2.6
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 22:33 El Hallak forcefully defends harness engineering terminology
When Kantrowitz suggests replacing the term 'harness' with 'orchestrator', El Hallak directly rejects the substitution and insists on the distinct technical role of harness engineering.
Hardest push from Alex ▶ 19:09 Host challenges guests with stories of agents breaking guardrailsKantrowitz directly challenges the guests' safety claims by bringing up real-world instances where agents bypassed constraints and contacted developers outside authorized channels.
Biggest teaching moment ▶ 2:38 El Hallak breaks down behind-the-scenes multi-agent orchestrationEl Hallak educates the host on how user-facing single-model interfaces actually orchestrate multiple specialized planners, critique agents, and research models behind the scenes.
Alex holds their own ▶ 15:55 Host presses guests on probabilistic nature versus deterministic boundariesKantrowitz leverages an insightful analogy from Mark Cuban to press the guests on how probabilistic LLMs can be reliably prevented from committing consequential errors in production.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Multi-Model Agent Blueprints and Enterprise Benchmarks | 5 | 5 | 1 | 3 | Kantrowitz asks why agent systems require multiple models rather than single frontier brand names like Claude or ChatGPT, prompting El Hallak to explain multi-agent orchestration architectures. The host also probes why ServiceNow partnered with NVIDIA over foundational model providers, following up with an astute observation about benchmark hacking. | |
| The Rise of OpenClaw and Unbounded Autonomy | 4 | 4 | 1 | 2 | The conversation explores OpenClaw and autonomous digital assistants. Kantrowitz articulates the core enterprise anxiety around uncontrolled autonomous agents before asking how such bots can safely be deployed in enterprise settings. | |
| Governing Enterprise Agents with OpenShell and AI Control Tower | 4 | 4 | 0 | 1 | Davis and El Hallak explain the mechanics of OpenShell sandboxing and the AI Control Tower governance layer. Kantrowitz asks brief clarifying questions regarding runtime policy enforcement. | |
| Implementing Zero-Trust and Deterministic Controls in AI Systems | 7 | 5 | 2 | 6 | Kantrowitz mounts strong pushback citing Mark Cuban's critique that AI models do not comprehend the consequences of their actions and questioning how probabilistic models can respect deterministic rules. The guests explain zero-trust, deny-by-default runtime sandboxing and deterministic harnessing. | |
| Harness Engineering and Encapsulating Reusable Agent Skills | 4 | 5 | 2 | 3 | El Hallak details skill encapsulation and harness engineering. Kantrowitz pushes back on terminology, asking why the industry uses 'harness' instead of 'orchestrator', leading to a spirited defense of harness engineering from both guests. | |
| Real-World Deployment: The L-1 AI IT Specialist | 5 | 3 | 0 | 1 | The discussion turns to real-world deployment with ServiceNow's L-1 AI IT specialist. Kantrowitz demonstrates domain knowledge by citing Bill McDermott's reported 90% IT ticket automation rate from GTC. | |
| Expanding Agentic AI to HR and Enterprise Operations | 6 | 3 | 0 | 2 | The panel discusses expanding agentic workflows to HR and call centers before exchanging future predictions. Kantrowitz actively contributes his own thesis regarding on-demand visual communication and video generation in enterprises. |
Statements from this episode (16)
NVIDIA's El Hallak: AI agents require multi-model architectures, not single models
“Agents are not a single model, right? We fundamentally believe that agents are made up of a, of an array of different models. Some proprietary models, some open source models that have been customized or post-trained, as they say, right? And so that's ultimate…”
NVIDIA's deep research agent blueprint uses at least seven specialized agents
“That deep research blueprint or agent is actually made up of no less than seven agents”
NVIDIA sees the best agent orchestration results using Claude Opus and GPT
“We've seen some of the best results with using either Opus from Anthropic or GPT for OpenAI for the orchestrator.”
ServiceNow runs fine-tuned models on internal GPU clusters using NVIDIA software
“We have a range of models that we use for different use cases from the frontier models to fine tune models that we're running on NVIDIA software inside of our own GPU clusters inside of ServiceNow.”
ServiceNow and NVIDIA publish enterprise benchmarks for frontier model labs to train on
“One of the things we haven't talked about is we work together to actually publish benchmarks that the frontier models then use to train on to make those use cases that we jointly care about that are really focused on enterprise use cases better for our enterpr…”
NVIDIA's El Hallak: OpenClaw is the fastest-growing GitHub project in history
“And OpenClaw became the fastest growing GitHub project on GitHub. Like, you look at Linux, it surpassed Linux, it surpassed React in a few weeks, right?”
Internet, knowledge base, and terminal access create a 'lethal trifecta' for agents
“Our CISOs, our compliance officers talk about a lethal trifecta in the enterprise context. And that lethal trifecta is when you are intermixing access to the unfettered internet, access to your knowledge base, and a coding terminal. Two of those three, no prob…”
El Hallak: NVIDIA OpenShell Governs Agent Permissions and Routing at Runtime
“And we have a, an open source secure runtime that's just between the infrastructure and the agent. And what it does is it defines what an agent can do at runtime. What it has access to, what it can read to, what it can read from, what it can write to what AP, …”
NVIDIA OpenShell AI sandboxes strictly enforce deny-by-default access permissions
“The default is no. When you spin up OpenShell, the default at runtime for an agent running in a sandbox is an, you are explicitly giving it access to very specific, you know, processes or actions that it wants to take.”
Enterprise AI requires wrapping probabilistic LLMs in strict, deterministic security harnesses
“There is an LLM that's reasoning and planning at the core. But there is a harness around it that provides a lot of determinism. And so that determinism is adding in governance, security, trust, integrations, permissions, and that's really what's going to make …”
El Hallak: ServiceNow Is Making 20 Autonomous Agents Available Alongside Project Arc
“Kind of a super exciting part about our partnership with ServiceNow is they talk about 20 autonomous agents that they've, you know, they're making available. These are in addition to these, you know, to project arc.”
ServiceNow reports AI reduces IT support resolution times by up to 99%
“And so we are seeing resolution times be reduced by as much as 99%. So if an AI can do it in five minutes, you don't have to wait days for a human to do it, and so resolution times really are drastically reduced when these things work.”
ServiceNow automated 90 percent of its internal IT support requests this year
“Yeah, we've automated 90% of our support requests this year, so most of our issues are just immediately resolved and somebody doesn't have to wait.”
ServiceNow builds autonomous AI agents mapped directly to existing human job roles
“Look, the way to think about what is relevant is to think about what jobs do people do today? And then to look at augmenting that and providing autonomous AI that does the same job. That's the way we're looking at this. So when we look at, Hey, how should we d…”
NVIDIA's El Hallak: Future enterprise systems will govern humans, agents, and robots
“I think in a couple years time, we'll be sitting here talking about how AICT governs, helps govern humans, agents, and robots.”
ServiceNow's Davis: Only a fraction of real enterprises have meaningful AI adoption
“There's still a fraction of real enterprises that have meaningful AI adoption throughout the enterprise, and so when I just think about the next couple of years, it's going to be about adoption and deployment in really, really complex business scenarios.”