NVIDIA Pushes Agentic AI Into Engineering Workflows With NemoClaw and OpenShell

NVIDIA is positioning its NemoClaw agent blueprint, backed by the OpenShell secure runtime, as a building block for what it calls “autonomous AI engineers” in computer-aided engineering and electronic design automation. The message, delivered alongside a lineup of engineering software vendors at GTC Taipei during COMPUTEX, is that accelerated compute has already cut simulation runtimes dramatically, but the broader workflow around simulation still takes too long because teams must stitch together design, meshing, setup, debugging, and post-processing step by step.

In NVIDIA’s description, NemoClaw is meant to turn that end-to-end process into a long-running agent that can execute across the whole workflow without human handoff at each stage. The company says NemoClaw is an “open blueprint” for specialized agents that combine frontier-model capabilities with a security layer, and it can be integrated into different orchestration approaches enterprises use to deploy and coordinate agents. NVIDIA also says NemoClaw comes with components including a choice of “harness” for integrations, a model router, and NVIDIA NeMo libraries for customization.

NemoClaw’s deployment story is also aimed at enterprises that do not want tools that only run in public clouds. NVIDIA says users can deploy NemoClaw from its DGX Spark personal AI supercomputers and also through enterprise data centers and cloud service providers. At the core, OpenShell governs how an agent accesses files, networks, and tools, with policy-based controls intended to be enforced at each layer. NVIDIA’s OpenShell documentation describes the runtime as providing sandbox containers, a credential-storing gateway, inference proxying, and policy enforcement, without dictating what the agent does inside the sandbox.

The engineering software partners highlighted by NVIDIA focus on shrinking verification and simulation loops that can stall product timelines. Cadence, for example, is building an autonomous register-transfer level (RTL) engineer using NemoClaw that orchestrates Cadence Design Systems ChipStack for design and verification, with NVIDIA saying the workflow was demonstrated in a GTC Taipei keynote demo and is cutting RTL verification time from weeks to hours. RTL verification is the step in digital chip design where engineers check that a circuit’s behavior matches its specifications before manufacturing can proceed.

Dassault Systèmes is presented as productizing an agentic platform intended to run long-running and autonomous agents for design, simulation, and manufacturing operations in secured environments, powered by NVIDIA NemoClaw and OpenShell. Siemens is integrating NemoClaw and OpenShell into Fuse EDA AI Agent, which NVIDIA describes as a domain-scoped multi-tool agent for semiconductor, 3D integrated circuit, and printed circuit board system design. Synopsys is collaborating with NVIDIA to apply agents across end-to-end engineering workflows, with NVIDIA pointing to Synopsys’ Ansys Icepak for a NemoClaw-based demo that meshes, simulates, and optimizes GPU electronics cooling designs.

NVIDIA’s blog also extends the use cases beyond chip design into optics, physics model development, geometry processing, and thermal simulation for consumer devices. Flexcompute is applying OpenShell to its Tidy3D and PhotonForge agents for multiphysics co-packaged optics design, with NVIDIA saying the workflow explores thousands of design variants overnight. Luminary is described as using NemoClaw to reduce the time and complexity of training AI physics models by orchestrating data generation, model selection, and re-training loops.

Other examples in the showcase target engineering pipelines that span multiple simulation modalities. Neural Concept is described as chaining electromagnetic, structural, and noise-vibration-style simulations in a multi-step pipeline for electric motor design. NVIDIA also cites nTop, the geometry engine behind JetZero’s blended-wing-body aircraft program, using NemoClaw to compress what it characterizes as days of geometry iteration into hours. In thermal simulation, NVIDIA highlights PhysicsX partnering with the Microsoft Surface team to automate the lifecycle of electronics thermal simulations, from mesh sensitivity analysis and data generation through model training and continuous accuracy monitoring during design exploration.

Not all details are fully specified in NVIDIA’s public materials. While the company provides qualitative cycle-time reductions (such as weeks-to-hours and days-to-hours), it does not disclose the experimental baselines, dataset assumptions, compute configurations, or success thresholds used in the demonstrations. Similarly, the security approach is described in terms of sandboxing and policy-based enforcement, but the blog and documentation do not enumerate specific certifications or third-party validation for each partner’s production deployments.

Still, the breadth of the partner list suggests NVIDIA is trying to standardize how agents handle sensitive engineering workflows, rather than merely offering model access. If NemoClaw and OpenShell become common reference infrastructure across CAE and EDA platforms, they could reduce the “glue work” that has traditionally required specialized teams to connect tools and debug pipelines manually. The next thing to watch is whether partners move these agent workflows from demos into repeatable, measurable production deployments, and whether NVIDIA expands what it considers governed tool access for additional engineering toolchains.