Siemens uses NeMo Gym to fix chip-design agent
Siemens says NVIDIA's NeMo Gym raised its standard-cell characterization agent's benchmark score from about 55% to about 95%, mainly by checking the agent's final outputs instead of its individual tool calls.
NVIDIA has published a case study on how Siemens used NeMo Gym, NVIDIA's open library for evaluating and improving agents, to make its EDA agent more reliable. NeMo Gym runs agents inside controlled test environments. The agent in question is the Solido Characterization Suite agent, which works through the Siemens Fuse EDA AI Agent. Fuse gives coding tools such as Claude Code, Codex and Cursor access to Model Context Protocol (MCP) servers and skills.
The agent automates standard-cell characterization. This is the job of rebuilding timing, power and noise models for logic cells whenever a foundry releases a new process design kit. The workflow covers tool setup, SPICE simulation, distributed compute, Liberty (.lib) file generation and validation. According to the post, errors often produce plausible-looking output rather than clean failures, so a setup mistake can carry through every later stage unnoticed.
The team built a NeMo Gym environment with a domain-specific verifier that compares the final Liberty files against verified reference files. It scores five things: coverage, validity, relative validation, accuracy within tolerance, and integrity, which checks that the agent didn't simply reuse or reformat existing values. Checking only the final files revealed missing cells and corners, wrong slew thresholds and out-of-tolerance values that checks on intermediate steps had missed.
The team then went through the agent's rollout logs and rewrote the skill and MCP tool descriptions. No new model was involved. Siemens reports the composite score rose from roughly 55% to roughly 95%, with fewer turns and fewer output tokens; turns reportedly fell by 66%.
The evidence base is small. The results reportedly rest on about 25 rollouts on a two-cell subset of the benchmark, and NVIDIA published the write-up on its own blog, partly to promote its own products. Independent replication has not been reported.