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

Siemens and NVIDIA Boost Chip Design AI Reliability with NeMo Gym

Siemens EDA has significantly improved the performance of its AI agent for chip design by using NVIDIA's NeMo Gym, an open evaluation environment. The collaboration focused on refining the Siemens Fuse EDA AI Agent, which automates standard-cell characterization, a foundational but error-prone step in semiconductor manufacturing. By implementing a rigorous, end-to-end verification process, the teams increased the agent's composite benchmark score from approximately 55% to 95%, establishing a more reliable pathway for AI-driven automation in complex engineering workflows.

End-to-End Verification Drives Agent Improvement

The core challenge in agentic chip design is that minor errors in early stages can silently propagate, leading to plausible but incorrect final artifacts. To address this, the project used NVIDIA NeMo Gym to create a controlled evaluation loop that inspected the final Liberty (.lib) files against verified references. This approach allowed developers to trace failures back to their root cause and refine the agent’s skills and tool descriptions. The iterative optimization process yielded substantial efficiency gains on top of the reliability improvements.

• Composite verification score increased from ~55% to ~95%.
• Agent turns, or interactions, were reduced by approximately 66% (from 155 to 53).
• Output tokens fell by around 64% (from 81K to 29K), indicating more concise and effective agent operations.

A Blueprint for Industrial AI Agents

This methodology provides a practical template for developing trustworthy agents in other specialized fields beyond Electronic Design Automation (EDA), such as Computer-Aided Engineering (CAE). The key insight is the necessity of evaluating the agent's performance based on the integrity of the final engineering artifact, rather than simply confirming the success of individual, intermediate tool calls. By building domain-specific verifiers that check for completeness, accuracy, and validity, organizations can systematically enhance agent reliability and prepare them for production environments.

The collaboration demonstrates that for agentic systems in complex engineering tasks, reliability hinges on evaluating the integrity of the final, complete artifact, not just the success of intermediate tool calls.