---
title: "Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym"
description: "AI agents in chip design often produce plausible but incorrect artifacts, making reliability a key barrier to production adoption. NVIDIA NeMo Gym, an open library for evaluating models and agents using environments, solves this by diagnosing failure modes in multistage workflows through verification of final outputs against reference artifacts. Applied to the Siemens Fuse EDA AI Agent for standard-cell characterization, NeMo Gym uses a verifier that checks coverage, validity, relative validation, integrity, and accuracy of generated Liberty files against known references. After refining the agent’s skills and MCP tool descriptions based on trajectory analysis of failures, the composite benchmark score improved from approximately 55% to 95%, while agent turns fell from 155 to 53 and output tokens dropped from ~81K to 29K. This method generalizes to other engineering domains by emphasizing evaluation of final artifacts and domain-specific verifiers."
---

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

[NVIDIA](https://yomu.fyi/company/nvidia-developer-blog) · Tanya Lenz · Oct 14, 2026

**Type:** Problem & solution

## Summary

AI agents in chip design often produce plausible but incorrect artifacts, making reliability a key barrier to production adoption. NVIDIA NeMo Gym, an open library for evaluating models and agents using environments, solves this by diagnosing failure modes in multistage workflows through verification of final outputs against reference artifacts. Applied to the Siemens Fuse EDA AI Agent for standard-cell characterization, NeMo Gym uses a verifier that checks coverage, validity, relative validation, integrity, and accuracy of generated Liberty files against known references. After refining the agent’s skills and MCP tool descriptions based on trajectory analysis of failures, the composite benchmark score improved from approximately 55% to 95%, while agent turns fell from 155 to 53 and output tokens dropped from ~81K to 29K. This method generalizes to other engineering domains by emphasizing evaluation of final artifacts and domain-specific verifiers.

## Context

Standard-cell characterization generates timing, power, and noise models for logic cells and must be repeated when foundries issue new process design kits (PDKs). The Solido Characterization Suite agent automates this but suffers from undetected errors that propagate through stages. This work uses the GF180MCU PDK with 229 standard cells across 36 PVT corners, with a two-cell subset used for tuning. The Siemens Fuse EDA AI Agent orchestrates tools through MCP interfaces and skills in characterization workflows.

## Approach / What changed

The approach builds a custom NeMo Gym agentic environment that defines task, tools, state, and a domain-specific verifier for Liberty file generation. The verifier computes a composite reward score based on coverage, validity, relative validation, accuracy (fraction of values within tolerance), and integrity. Rollouts are run with Claude Opus 4.8 in Claude Code, and trajectories are inspected to diagnose failures. Skills and MCP tool descriptions are iteratively refined based on failure analysis, and the process repeats under identical evaluation conditions.

## Takeaways

- Evaluating agents on final Liberty file artifacts catches failures that intermediate checks miss, such as missing cells, incorrect thresholds, and timing values outside tolerance.
- Refining MCP tool descriptions with explicit JSON schemas and simplifying overlapping file and shell interfaces greatly improves tool selection and reduces agent errors.
- Using NeMo Gym’s iterative optimization, the composite verification score rose from ~55% to ~95%, agent turns dropped 66% (155 to 53), and output tokens dropped 64% (~81K to 29K).

## FAQ

### Why is evaluating final artifacts better than checking intermediate steps?

Because undetected errors early in a multistage workflow can propagate silently and produce plausible final outputs. Final Liberty verification reveals missing cells, incorrect thresholds, and out-of-tolerance values that intermediate checks may not catch.

### How does NeMo Gym compute the composite reward score?

The reward multiplies indicator gates for coverage, validity, relative validation, and integrity with the fraction of output values within tolerance. The reward is zero if any gate fails.

### How many rollouts were conducted and what agent harness was used?

The team conducted approximately 25 rollouts using Claude Opus 4.8 running in Claude Code, with each run taking about 1.5–2 hours on a two-cell subset across 36 PVT corners.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Chip Design](https://yomu.fyi/topic/chip-design), [EDA](https://yomu.fyi/topic/eda), [MCP](https://yomu.fyi/topic/mcp), [NVIDIA](https://yomu.fyi/topic/nvidia)

- Source: [NVIDIA](https://developer.nvidia.com/blog/optimizing-agent-reliability-and-tool-calling-efficiency-for-chip-design-using-nvidia-nemo-gym)
- Source URL: https://developer.nvidia.com/blog/optimizing-agent-reliability-and-tool-calling-efficiency-for-chip-design-using-nvidia-nemo-gym
- Ingested by Yomu: 2026-10-14T18:00:15.498Z

[Read original post](https://developer.nvidia.com/blog/optimizing-agent-reliability-and-tool-calling-efficiency-for-chip-design-using-nvidia-nemo-gym)