When an answer is not enough: NVIDIA gives the agent an engineer’s tools
NVIDIA is connecting agents to the physics, simulation and verification tools engineers actually use. For the region, the value is measured by verification accuracy and control of design data, not by the model’s eloquence.
Listen to this article
Read by Anchor
A language model may write code that looks sound, but a chip does not work on appearances. A small timing or verification error may remain hidden until the code is put through a simulator and an inspection tool. This is where the shift announced by NVIDIA on July 26, 2026, takes place. The engineering agent is no longer expected only to produce an answer. It can now call physics, simulation and numerical solving tools, then return with results that can be tested.
The company expanded the NVIDIA Agent Toolkit by rebuilding PhysicsNeMo libraries and adding updated CUDA-X libraries as callable tools and skills. According to NVIDIA’s official announcement, the capabilities include customisable physics models, iterative and direct solvers for sparse linear systems through the cuISS and cuDSS libraries, and quantum chemistry simulation through cuEST. In practical terms, a developer can connect an agent to a specialist tool instead of asking the model to guess a physical result from text.
The real shift begins when the tool becomes part of the reasoning process.
This is clear in chip design. Register transfer level (RTL) code describes the behaviour and timing of a circuit, and generating Verilog from a single prompt is not enough. NVIDIA’s technical blog explains that the ACE-RTL agent operates in a cycle of generation, testing, reflection and correction. It produces code, runs simulations and checks, reads the failures, and retains the context of previous attempts so that it does not repeat the same mistake.
In tests published by the company across nine categories of CVDP problems, Nemotron 3 Ultra within the agent achieved an average success rate of 97.1%, compared with 95.2% for the Kimi K2.6 model and 92.1% for the GLM 5.2 model. It also used an average of 6,629 tokens per cycle, about 28% fewer than GLM and about 71% fewer than Kimi. These are NVIDIA’s results in a defined test environment, not a guarantee of similar performance in every project. But they support a more precise proposition than a claim of full autonomy: feedback from verification tools may make the model more useful than running it on its own.
The measure of the agent then becomes the number of errors it resolves, not the number of lines it writes.
NVIDIA says Cadence, Siemens and Synopsys are integrating these components into existing design and verification workflows. Among the examples in the announcement, Siemens recorded more than a tenfold increase in library characterisation speed while reducing token cost by more than tenfold. Samsung used the cuLitho library to achieve up to twenty times higher performance in computational lithography. These remain figures announced by the participating companies, but they show where the competition lies: connecting the model to the working tools engineers already rely on, from design through approval of the result.
For the region, the issue is not acquiring another conversational assistant. Research centres, industrial design teams and simulation laboratories in the Middle East and North Africa handle sensitive engineering data and operational knowledge that are difficult to send without restriction to a closed service. NVIDIA says Nemotron 3 Ultra is an open model that can be customised with an organisation’s data and deployed locally or on its premises. This gives teams greater control over data residency, but does not create sovereignty automatically, because hardware, libraries, design tools and the support chain remain elements of technical dependence.
Sovereignty here is an architectural decision, not a quality conferred by the model’s name.
A serious test should therefore begin with one workflow, not a broad rollout. Choose a time-intensive task such as coverage closure, thermal analysis or circuit inspection. Compare the agent with the baseline on verification accuracy, completion time, token cost, and where specifications and logs are stored. Keep approval of the result with a responsible engineer. If the system cannot demonstrate its effect within the actual inspection tools, the eloquence of its answer does not mean that the design has improved.