NVIDIA expanded its Agent Toolkit on July 26, 2026 to include NVIDIA PhysicsNeMo and CUDA-X libraries as agent-callable tools, letting AI agents invoke physics simulations and accelerated solvers directly as part of chip and systems engineering workflows.

What's new

NVIDIA announced an expansion of its NVIDIA Agent Toolkit on July 26, 2026, adding NVIDIA PhysicsNeMo and CUDA-X libraries as agent-ready tools. In its own announcement: "NVIDIA today announced an expansion of NVIDIA Agent Toolkit for engineering, now adding NVIDIA PhysicsNeMo™ and CUDA-X™ libraries as agent-ready tools and skills built to transform how the world designs and develops products."

PhysicsNeMo, NVIDIA's physics-AI framework for training and deploying simulation-capable models, has been re-architected into a set of agent-friendly libraries specifically for engineering workflows — meaning AI agents can now invoke physics simulations as callable tools rather than requiring a human to run them separately and feed results back in. Alongside it, CUDA-X libraries bring accelerated numerical solvers and quantum chemistry capabilities into the same agentic toolchain, giving agents direct access to GPU-accelerated computation for tasks like chip verification, packaging analysis, and systems-level design.

Timothy Costa, NVIDIA's vice president and general manager of computational engineering, framed the release around a shift in what AI agents can now do in engineering contexts: "Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design. With NVIDIA Agent Toolkit, developers can build agentic engineers that reason using physics, run complex simulations and generate high-fidelity data to become a new engine for innovation in chip and system design."

Context

The release extends NVIDIA's push to position agentic AI as a core part of its own chip design and engineering pipeline, following earlier moves like deploying its own Vera CPU to speed internal chip design and open-sourcing GPU-accelerated simulation frameworks for other technical domains (including a medical-physics simulation toolkit). By packaging PhysicsNeMo and CUDA-X as agent-callable tools rather than standalone libraries, NVIDIA is explicitly targeting the emerging category of "agentic engineers" — AI systems that plan and execute multi-step technical workflows (running a simulation, evaluating the result, adjusting a design, re-simulating) rather than answering single-shot queries.

Why it matters

This is a concrete example of physical-simulation software becoming agent-native infrastructure rather than a separate tool a human operates alongside an AI assistant. For chip designers and systems engineers, it lowers the barrier to building agents that can autonomously iterate on physics-grounded design problems — running a simulation, checking results against constraints, and adjusting parameters — instead of just generating code or text. It also reinforces NVIDIA's strategy of expanding its moat beyond hardware into the software layer that makes its hardware indispensable for agentic workflows specifically, at a moment when chip design and engineering simulation are increasingly cited as one of the more commercially credible near-term applications of agentic AI.