OpenShell governs what the agent may do. Regulayer™ proves who currently authorizes it.
The NVIDIA agent stack decides how an agent runs. Regulayer™ adds one check at the moment an action takes effect: does a named person’s current authority cover this exact action? Then it seals the decision in a record an outsider can check.
The stack, in one request
Each layer answers its own question. Regulayer™ answers the last one.
Outside the modelRegulayer™ lives outside the model. Change the model and the check stays where it is.
Proposes the actionIn NVIDIA’s words“Input, retrieval, dialog, execution, and output rails run at different stages of an LLM interaction.” NeMo Guardrails documentation
Rails checkedIn NVIDIA’s words“OpenShell provides a secure runtime boundary for controlling how autonomous AI agents execute tasks.” NVIDIA, 28 Sep 2026
Inside the runtime policyWhat it addsA named person’s current authority, checked at the moment the AI acts. A sealed record an outsider can check.
CheckingWhat arrivesOnly an action a named person currently authorizes, with its sealed record.
WaitingWhere it plugs in
NVIDIA names where this check goes. Regulayer™ is designed to sit there.
“Trusted middleware outside the sandbox can connect identity services and add application-specific checks to the request path.”NVIDIA Technical Blog, “Add Runtime Controls to AI Agents with NVIDIA OpenShell”, 28 Sep 2026. Read it
In the request path. Regulayer™ answers before the action takes effect, in milliseconds in our live tests. If a named person’s authority does not cover it, the action does not run.
Outside the model and the sandbox. The check is designed to sit beside OpenShell, outside the agent, so no output of the agent can override it.
A record for someone who was not in the room. OpenShell records its policy decisions in an audit trail. Regulayer™ adds a sealed record of who authorized each action, which an inspector, an insurer or a partner can check.
Where it lands first · GxP
Human review decides. Regulayer™ makes that decision enforceable when the AI acts.
“Any output or recommendations from an AI agent must be reviewed and cleared by an authorized human representative.”US FDA, first warning letter on AI, 2 April 2026. The record in the Consequence Library
The quality unit names who may release. Regulayer™ checks that person’s current authority on every release the agent attempts, and stops the ones it does not cover.
Withdraw authority, and the same request stops. Every decision, allowed or stopped, is sealed as it happens.
What Regulayer™ adds to the stack
Two points.
Each one runs in the live demo.
Regulayer™ is a member of NVIDIA Inception. NVIDIA, the NVIDIA logo, Inception, NeMo, NIM, Nemotron and OpenShell are trademarks of NVIDIA Corporation. This page describes how Regulayer™ fits alongside NVIDIA software. It is not an NVIDIA endorsement. Patent pending.
