Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym

NVIDIA and Siemens describe using the open NeMo Gym library to evaluate the Siemens Fuse EDA AI Agent on standard-cell characterization, which generates timing and power models whenever a foundry issues a new process design kit. A domain verifier scored final Liberty files for coverage, validity, accuracy and integrity, giving zero reward if any gate failed.

After about 25 rollouts with Claude Opus 4.8 in Claude Code as the fixed model, refinements to skills and MCP tool descriptions raised the composite score from roughly 55% to 95%, cut agent turns from 155 to 53, and reduced output tokens from about 81K to 29K.

This is the most practical agent post we've read this week. The key insight — check the finished artifact, not whether each tool call succeeded — is the difference between an agent that looks busy and one that's actually right.

We especially like the integrity check, which catches agents reusing, truncating or reformatting existing values. That's a polite way of saying agents cheat when they can, and good verifiers assume it.

Notice that the model never changed. All of the gain came from clearer tool descriptions and better skills. That's encouraging for teams who can't train models but can write better interfaces.

The caveat: the tuning ran on two cells, not the full 229-cell library, and on a single benchmark. In chip design, plausible-but-wrong outputs are expensive, so I'm watching for production results across new process kits, not just development numbers.