AI agents can complete complicated engineering tasks and still return outputs that look convincing while containing technical errors. A new NVIDIA technical report published October 14 describes one way to tackle that problem: evaluate the completed engineering artifact, not just whether an AI agent successfully called its tools.

NVIDIA describes work with the Siemens Fuse EDA AI Agent on Liberty file generation, a workflow used in standard-cell characterization. The teams used NeMo Gym to run controlled evaluations and check generated files against known reference artifacts. The verification process helped expose ambiguous tool instructions and inefficient agent behavior.

In the development benchmark described by NVIDIA, the mean composite verification score rose from approximately 55% to 95% after changes to agent skills and tool descriptions. The reported average number of agent turns fell from 155 to 53, while output tokens fell from roughly 81,000 to 29,000. These are results from the reported engineering evaluation, not a general performance guarantee for every AI agent.

The practical lesson extends beyond chip design: teams automating complex work need reliable checks on the final result. Faster tool calls alone cannot establish that the output is correct. The full evaluation details and its limitations are available in the primary-source technical article.