Skip to content
VZU

Insights · 9 min read

The on-prem agent runtime.

On-prem is a first-class deployment mode. Why the regulated operator chooses an on-prem agent runtime over a hosted SaaS, and what the runtime has to do to earn the call.

Brief VZU
/>

The default in 2024 was cloud. The default in 2026, for the regulated operator, is on-prem. The reason is the same as it has always been: the audit trail is the regulator's form, and the regulator's form must run on the operator's hardware.

Cloud AI is a fine choice when the operator does not have a regulator that cares where the data lives. Cloud AI is the wrong choice when the operator is a public company (SEC Reg FD, OSC 51-102, ESMA MAR, FCA DTR 2), a healthcare provider (HIPAA), a financial institution (PCI-DSS, SOX), a defense contractor (ITAR), a government agency (FedRAMP, GC Cloud), or a critical infrastructure operator (NERC CIP). In each of these cases, the data must run on the operator's hardware, behind the operator's firewall, with the operator's audit trail.

The on-prem agent runtime is the runtime that ships on the operator's hardware. It is the same code as the cloud runtime. The difference is where the data lives and who owns the audit trail.

What the runtime has to do to earn the on-prem call

Five things. Each is non-negotiable for a regulated operator.

  1. Audit trail on from the first character. Every tool call, every model invocation, every file read, every API call, every operator override — every action lands in the audit trail. The trail is signed, replicated, and queryable by the regulator.
  2. MNPI boundary as a mechanical wall, not a procedural one. The runtime enforces the boundary by class of data, not by convention. An agent cannot cross a wall it does not have the role to cross, regardless of what prompt the operator types.
  3. RBAC at the agent level, not the user level. A user has roles. An agent has roles. A tool has roles. A model has roles. The runtime checks every action call against the role map of the agent, the user, the tool, and the model.
  4. Bring-your-own LLM. The operator picks the LLM. The runtime runs the orchestration. The data never leaves the perimeter because the LLM runs on the operator's hardware (or in the operator's VPC).
  5. Air-gapped option. The runtime can run with no internet dependency. The audit trail is signed, replicated, and regulator-grade without any cloud dependency. The brief is the contract. The work is the work.

The VZU on-prem posture

VZU ships cloud, hybrid, on-prem, and air-gapped. The same code runs in all four modes. The audit trail, the RBAC, the MNPI boundary, the SOC 2 / HIPAA / ISO 27001 / PIPEDA posture are enforced by the runtime, not by a vendor.

The 12 agents on the VZU OS — Atlas, Vega, Kai, Nova, Pixel, Echo, Sketch, Launch, Sentinel, Oracle, Quill, Hunter — are runtime citizens. Each agent has a toolset, a memory schema, an audit row, a role, a class of data it can touch. The runtime enforces each. The operator owns each.

The deployment modes

Mode 1

Cloud

VZU-managed, in our VPC. Data is segregated. No training. SOC 2 in pursuit.

Mode 2

Hybrid

Runtime in the operator's VPC. Operator-owned audit trail. Bring-your-own LLM.

Mode 3

On-prem

Runtime on the operator's hardware. Operator-owned everything. Same code.

Mode 4

Air-gapped

No internet dependency. Signed, replicated audit trail. Regulator-grade.

The bottom line

The next decade of enterprise AI is on-prem. The runtime is the operator's. The audit trail is the regulator's form. The brief is the contract. The work is the work.

Brief VZU on your runtime