Self-verifying, long-running EDA AI agents that engineers can trust - Custom IC

Summary

Siemens and NVIDIA announced long-term AI agents for EDA that continuously validate their own decisions; these agents, while guiding multi-tool semiconductor and PCB workflows, constantly verify each decision with deterministic, physics-based engines.

Why it matters

This approach promises real-time validation to reduce design errors and accelerate tapeout schedules; it addresses silicon's high cost and tight market windows. At the same time, it signals a shift to AI-native EDA with enterprise-level security, GPU acceleration, and multi-vendor orchestration.

Highlights

• Continuous physics-based validation at every step of the workflow.
• A unified orchestration layer (MCP architecture) coordinating dozens of specialized EDA tools.
• Integration of NVIDIA's NeMo Gym, OpenShell, and Nemotron models; for self-improvement, security, and token efficiency.

Critical take

The article does not provide concrete performance data or real-world case studies; it leaves open the question of whether these agents can manage the full complexity of production designs without creating new bottlenecks.