The Right Way To Build AI Agents — With Nvidia's Adel El Hallak and ServiceNow's Joe Davis

TL;DR

• ServiceNow claims a 90% automation rate for its own support desk — Joe Davis says ServiceNow now resolves most L1 IT requests automatically, with some resolution times dropping by as much as 99% when an AI can act in minutes instead of waiting days for a human queue.

• The production agent stack is multi-model, not one giant brain — Adel El Hallak describes ServiceNow’s deep-research setup as at least seven agents: an orchestrator, a planner, and several specialized sub-agents fine-tuned on Nvidia’s open NeMoTron models for tasks like critique, fact gathering, and forecasting.

• OpenAI or Anthropic may orchestrate, but Nvidia provides the runtime and safety rails — the partnership centers on Open Shell, Nvidia’s open-source secure runtime that sandboxes agents, enforces permissions at runtime, and starts from a “deny by default” posture.

• The enterprise risk is the ‘lethal trifecta’ — Hallak says CISOs get nervous when an agent has all three of these at once: access to the open internet, access to the company knowledge base, and access to a coding terminal.

• Harness engineering is becoming as important as model quality — both guests argue that an agent’s performance now depends heavily on the tooling around the model: file access, code interpreters, MCP tools, skills, integrations, governance, and orchestration loops.

• The next frontier is broader enterprise adoption — and eventually physical AI — Davis expects the next few years to be about deploying agents into complex business workflows like HR and CRM, while Hallak thinks the same governance layer will eventually extend from humans and agents to robots and physical assets.

The Breakdown