Jensen Huang ended his keynote at GTC Berlin 2026 by using one prompt to set up a customized AI agent to build a physical training simulation. The ‘Hermes Agent,’ running on Nvidia RTX Spark, created a simulated obstacle course for the Hugging Face Microduck robot by coordinating multiple AI models. It utilized Nemotron for planning and GPT-6 Astra for the construction of the environment using tools including Omniverse, Blender, and FreeCAD.
The system subsequently trained thousands of simulated ducks within the environment until it validated a policy for deployment on a physical robot. Huang stated that the most efficient way to develop new AI agents is to employ an existing agent to build them, signaling a shift toward autonomous agent orchestration.
