Nvidia expands Agent Toolkit with PhysicsNeMo and CUDA-X libraries for autonomous engineering agents

6:45 AM · July 28, 2026

Nvidia announced on July 26, 2026 an expansion of its Agent Toolkit, a set of libraries meant to give AI agents the specific technical skills needed to do real engineering work rather than just reason about it in the abstract. The update folds in a re-architected version of PhysicsNeMo, Nvidia's physics AI library, now packaged as agent friendly skills for training and deploying custom physics models, alongside updated CUDA-X libraries including cuISS for accelerating sparse linear systems, cuDSS for electronic design automation and scientific simulation, and cuEST for quantum chemistry simulations at scale. “Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design,” said Timothy Costa, Nvidia's vice president and general manager of computational engineering. The toolkit is already being adopted by major electronic design automation and engineering software vendors, including Cadence's AuraStack AI Super Agent and Millennium M2000, Synopsys's AgentEngineer platform, and Siemens's Fuse EDA AI Agent, alongside Samsung, TSMC, ChipAgents, Silvaco and Keysight. Nvidia also highlighted its Nemotron 3 Ultra model as a leading open model for agentic register-transfer level coding, the low level hardware description work chip designers rely on. The expansion continues Nvidia's push to position its hardware and software stack as the foundation not just for training and running AI models, but for AI agents that can autonomously carry out physics heavy engineering tasks across chip design, industrial design and scientific computing.