NVIDIA VP: Same AI Model, 30% to 100% (Must Know)
I sat down with Adel El Hallak, VP of Software Product Management for the NVIDIA Agent Toolkit, at NVIDIA GTC Berlin 2026. The same AI model went from 30% to 100% on a benchmark just by changing the harness around it. Adel explains how, and walks through what NVIDIA is building for agents: the NVIDIA Agent Toolkit, the Agent Knowledge Foundation versus plain RAG, Kumo models that make predictions from tables, why every agent demo breaks at some point, NeMo Relay for tracing what agents do, OpenShell as the one non-negotiable runtime, skills and routers (NeMo Switchyard), and the tools he uses himself. Chapters 0:00 Cold open 0:28 Introduction 0:48 Adel's background: from IBM to NVIDIA 2:04 What is the NVIDIA Agent Toolkit? 3:37 Agent Toolkit vs LangChain, CrewAI and Agno 5:10 Blueprints: open reference agents 6:51 Agent Knowledge Foundation vs RAG 10:00 Kumo models: prediction on tables 11:43 Ontologies and GraphRAG 12:11 When a demo breaks: measuring agents 15:57 Harness engineering: 30% to 100% 18:02 Tools the team uses: OpenShell and NeMo 21:16 How to improve agents: skills and routers 23:47 Adel's personal agent stack 26:01 Bonus: how do you say Adel? 26:26 Thanks for watching Sources for the slides and diagrams shown - 30% → 100% (NVIDIA developer blog, 21 Aug 2026): https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/ - Kumo models (NVIDIA): https://developer.nvidia.com/topics/ai/agent-knowledge-foundation/kumo - Jensen's five parts of an agent (NVIDIA GTC Berlin 2026 keynote): https://www.youtube.com/watch?v=PvFjbPYEOZI - Blueprints (NVIDIA): https://build.nvidia.com/blueprints - NeMo Relay (NVIDIA): https://developer.nvidia.com/blog/tracing-agent-harness-behavior-with-nvidia-nemo-relay/ - NeMo Switchyard (NVIDIA): https://developer.nvidia.com/blog/route-ai-agent-workloads-across-models-with-nvidia-nemo-switchyard/ Filmed in person at NVIDIA GTC Berlin, 22 October 2026. #NVIDIAGTC #AIAgents #NVIDIA
