Tech • AI • Robotics • Game
Super AI Agents for Science, Manufacturing and Engineering | GTC Berlin 2026 Edition
TL;DR
Nvidia is positioning a full agent-building stack that combines models, enterprise data, simulation tools and secure runtime controls to let companies deploy specialized AI agents for research, factory operations and engineering design.
KEY POINTS
Five-part agent architecture
Nvidia describes AI agents as systems built from five core components: a model for perception, reasoning and planning; a harness to coordinate tasks; a knowledge foundation linked to structured and unstructured company data; tools for simulation, experimentation and action; and Open Shell to control what the agent can access and execute. The framework is aimed at turning domain expertise into customized agents rather than relying on a general chatbot alone.
Choice of model and enterprise customization
Companies can begin with a frontier closed model or customize an open model, then connect it to Nvidia’s Agent Toolkit. The emphasis is on grounding the agent in proprietary business knowledge and coupling it to software and instrument workflows that can produce measurable outputs.
Drug design for B-cell lymphoma
At Lyra Sciences, the target is a therapy for B-cell lymphoma. Its Lyra Iris agent, built on Nvidia NeMo Triton, converts that scientific objective into an experimental plan, while agents using the BioNeMo toolkit call protein design and RNA optimization models to run millions of virtual experiments in a computational dry lab.
Closed-loop lab automation
The most promising drug candidates then move into an AI-driven science factory where autonomous robots and instruments, developed with Nvidia Omniverse and Cosmos, physically make and test them. Experimental results feed back into Iris, creating a repeated design-test-refine loop intended to speed discovery.
Factory planning at Siemens
At Siemens’ electronics factory in Erlangen, AI agents are used to prepare new products for manufacturing. Process Preparation X agents use Nvidia Cosmos to generate virtual prototypes and synthetic data, allowing inspection models to be trained and validated before production begins.
Production validation with digital twins
Intelligence Center X agents use orders, supplies and line configurations to recommend production plans. A Digital Twin Composer built on Nvidia Omniverse libraries then validates those plans with physics simulation and real-time factory data before any change reaches the factory floor.
Automotive redesign with Dassault Systèmes
In a Dassault Systèmes workflow, an engineer adapting a Volkswagen Golf door for the Tiguan uses Aura and Leo, powered by Nvidia Nemotron, to reason over engineering requirements. The harness preserves context across planning, design and simulation so specialized sub-agents can work as a coordinated system.
Engineering software tied into the loop
Aura uses ENOVIA to plan against vehicle dimensions, weight limits and other constraints. Leo selects a compatible door design and reshapes it in CATIA, while a simulation agent uses Simulia, accelerated by Nvidia CUDA-X libraries, to test the change, identify a vibration problem and return results for further refinement before final human approval.
CONCLUSION
Across biotech, manufacturing and automotive engineering, Nvidia is making the case that useful AI agents depend on secure access, domain data and simulation-backed tools as much as on the underlying model. The strategy is to embed AI inside operational loops where software recommendations can be tested, validated and improved before real-world deployment.
Ask a question