Thu Oct 15
The Silicon Beneath the Agent
As agentic AI moves onto safety-certified industrial silicon and RTOS stacks, buyers must locate exactly where the certification boundary ends and the uncertified agent begins.
The Silicon Beneath the Agent
Two announcements this month point at the same unresolved question from different directions. ForwardEdge ASIC, a Lockheed Martin subsidiary, has selected a new chip platform for mission-critical autonomous systems, and Inova Semiconductors is building a reference physical AI platform for robotics with MIPS, with Green Hills Software supplying the safety-critical software layer underneath, according to Manufacturing Dive. Separately, BlackBerry’s QNX, a real-time operating system built for safety-critical deployments, is positioning itself as a growth engine across automotive, industrial automation, and robotics as those sectors absorb more AI workload at the edge, per TradingView.
Both stories describe the same shift: agentic AI inference is moving off the cloud and directly onto embedded silicon and real-time operating systems that already carry safety certification pedigree in aerospace and automotive contexts. That pedigree is the selling point. It is also the open question compliance teams need to press on before signing a procurement contract.
Functional safety certification under frameworks like IEC 62443 and DO-178C was built to validate deterministic software running on known hardware boundaries. An agentic model making autonomous decisions about equipment movement is not deterministic in the same sense, and it is not clear that a certified RTOS or chip automatically extends its certification envelope to cover a non-deterministic AI layer running on top of it. Vendors are not being asked this question loudly enough, and the marketing language, tends to blur the line between “built on certified silicon” and “the AI decision layer is itself certified.”
The contrast with ABB’s new EyeMotion platform makes the gap sharper. ABB is explicitly lowering the skill threshold required to train, deploy, and optimize AI-powered robotics applications, per Automate.org. That is good for throughput and bad for the assumption, baked into most existing safety cases, that a qualified engineer configured the system and understands its failure modes. Meanwhile the market for human-robot interaction sensors, the hardware layer that detects and reacts to human proximity, is expanding specifically because collaborative robot safety mandates demand it, according to IndexBox. In other words, as the software layer gets more autonomous and easier to deploy, more of the safety burden is quietly being pushed down to sensors and silicon, with the expectation that hardware-level certification will catch what the software layer doesn’t guarantee.
That expectation needs to be tested, not assumed. Industrial buyers evaluating agentic AI platforms should require vendors to document, in writing, precisely where the certified boundary of the stack ends and where the uncertified decision-making layer begins. If a vendor cannot draw that line cleanly, the safety case for the deployment has a hole in it, regardless of which chip or RTOS sits underneath. The procurement question for 2026 is not whether the silicon is safety-rated. It is whether the agent running on it was ever in scope of that rating at all.
Board record
This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.
| Seat | Reviewer | Finding |
|---|---|---|
| Chair · Editorial Judgment | Claude | cleared. The core argument—that functional safety certification of underlying silicon/RTOS does not automatically extend to non-deterministic AI decision layers—is logically sound and the procurement recommend |
| Source & Claim Verification | Qwen · hosted fallback (qwen-2.5-72b-instruct) | cleared. Most factual claims are supported by citations, but a few lines about the nature of functional safety certification and the contrast with ABB’s platform could benefit from additional sources. |
| Regulatory & Framework Fidelity | Mistral | cleared. The briefing accurately identifies DO-178C and IEC 62443 concerns but fails to address NERC CIP, ISO 42001, or NIST AI RMF implications for the described agentic AI deployments. |
| Technical Accuracy | Llama | cleared. The technical distinction between deterministic safety certifications (DO-178C/IEC 62443) and the non-deterministic nature of LLM-based agentic inference is accurately characterized. |
| Bias, Balance & Hype Control | Gemini | cleared. The briefing effectively identifies and critiques vendor hype by highlighting the distinction between ‘built on certified silicon’ and ‘AI decision layer is itself certified,’ directly addressing the |
| Novelty & Non-Duplication | Grok | held. The certification-boundary procurement angle is a serviceable synthesis of otherwise standard wire items on edge silicon/RTOS and ABB tooling, not a pure rewrite, though the underlying safety-envelope |
| Validation | DeepSeek | cleared. The briefing’s central claim about the certification gap is validated by the sources, which confirm vendors are marketing certified hardware/OS layers for AI without clarifying if the AI decision laye |
Sources cited: 19. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.