Siemens and NVIDIA bet on self-verifying AI agents

Siemens and NVIDIA expanded their partnership on Sunday around a phrase worth stealing: self-verifying agentic AI. The setting is chip and PCB design, which is not your business. The architecture is, because it answers the question every AI automation vendor dodges — how do you know the agent got it right? Siemens' answer isn't a better model. It's a deterministic engine that checks the agent's homework before anything ships.

What actually happened

Per Siemens' announcement, the company is adding NVIDIA technology to its Fuse EDA AI Agent system so long-running agents can continuously validate their decisions against deterministic, physics-based EDA engines. The stack: NVIDIA's Nemotron 3 Ultra reasoning model for the agent loop, the NeMo Gym library to tune agents for quality and token efficiency, OpenShell as a secured runtime, and CUDA-X libraries for the heavy math.

Siemens claims a 10x cut in library characterization turnaround and a 5x to 10x reduction in token costs, and notes that verification currently eats up to 70% of design time. Those are vendor numbers on unreleased capability — the features are slated for forthcoming releases of Siemens' EDA portfolio, with no ship dates given. Treat the figures as marketing and the design as engineering.

Why self-verifying agents matter for your business

Strip out the semiconductors and the pattern is the one small operators need most. An agent proposes; a deterministic system judges. The model does the part it's good at — reading messy inputs, planning, calling tools — and something dumb, fast, and repeatable decides whether the output is acceptable. You are not trusting the model. You are trusting the checker.

You already have checkers, and they're cheaper than Siemens'. A pricing agent's output goes through a margin-floor rule before it touches the catalog. An invoice agent's line items get reconciled against the PO total. A support agent's refund gets validated against your actual policy table. Each of those is fifty lines of ordinary code, and each converts "the AI usually gets it right" into a claim you can defend to a customer.

The tell for a serious AI vendor is whether they can name the verification step. If the answer is a bigger model, a confidence score, or a human in the loop who's really just clicking approve — that's a spot-check, not verification. This is the same case we made when Poetic raised $50M for deterministic execution. The industry keeps arriving at it because it's the only thing that works at volume.

Sources: Siemens (PR Newswire), Siemens Fuse EDA AI Agent.