NVIDIA Agent Toolkit: Building Customized AI Agents with Models, Tools, and Secure Runtime

An overview of how AI agents combine a model, harness, knowledge foundation, tools, and a secure runtime, using NVIDIA's agent toolkit to turn enterprise knowledge into customized agents. The content illustrates the architecture through three applied scenarios: therapeutic design, electronics production planning, and automotive engineering design adaptation. It emphasizes the roles of open or frontier models, accelerated tools, and simulation in coordinating agent workflows.

■ Agent architecture
- Model for perception, reasoning, and planning
- Harness for work coordination
- Knowledge foundation from structured and unstructured data
- Tools for simulation, experimentation, and action
- NVIDIA Open Shell for access and execution control

■ Example workflows
- Lyra Sciences: B-cell lymphoma therapy design with NeMo Triton, BioNeMo, Omniverse, and Cosmos
- Siemens Electronic Factory Erlangen: process preparation and production planning with Cosmos, factory operations blueprint, and Omniverse
- Dassault Systèmes with Volkswagen: door design adaptation using Nemotron, ENOVIA, CATIA, Simulia, and CUDA-X

Suited to technical decision-makers, AI practitioners, and manufacturing or engineering teams evaluating agent platforms. Viewers gain a high-level view of how models, domain data, tools, and secure execution fit into agent workflows and can identify components to assess for their own agent initiatives. It provides architecture and scenario-level coverage rather than pricing, implementation steps, or benchmark results.