Agentic Design Automation for Embedded Systems: The Promise, the Peril, and the Proofs

Agentic design automation is transitioning rapidly from demonstration to deployment in embedded systems, marking a paradigm shift driven by radical gains in cost, turnaround time, and optimization quality. I will begin by presenting a vision for multi-agent EDA flows spanning architecture through signoff—combining physics-based AI, dynamic training, and adaptive algorithms. These flows constitute novel ecosystems of distributed, semi-autonomous participants. Left to develop organically, they lack robust identity verification, scoped credentials, and auditable execution traces. This creates deep architectural exposures: poisoning attacks and backdoors that carry contamination directly to silicon; tool metadata and inter-agent messaging that laterally propagate malicious injections; generator and verifier agents—derived from common models—that unwittingly certify their own errors; and agents that optimize for checkable proxies rather than true design intent. Compounding these risks, shared contexts dissolve confidentiality, and continual adaptation leaves no fixed artifacts to formally certify. The resulting defects are fabricated, replicated, and unpatchable.

I will then examine methods for thwarting these threats through AI resilience anchored in physics-based ground truth, hardware/software co-design for attestation and provenance, and cryptographically secure computing. Ultimately, these defenses converge on a single foundational question: How do we establish verifiable trust in the agents that monitor other agents? I conclude with the technical challenges and opportunities ahead.