What it is
An agent system that writes, sends and decides needs rules it cannot talk its way around. Zen Aegis runs its own company on the platform it deploys for clients, so the controls had to hold on its own work first. The figures below were counted on August 1, 2026.
The decision
Nothing about prompts, outputs, credentials or client content leaves the boundary. Each client is isolated at the infrastructure layer, with its own host, its own secrets vault, its own tracing and its own set of tool servers.
The grounding gate checks on three surfaces, and the third cancels at delivery, because the first two can be routed around.
What was built
- Grounding Gate: fail-closed enforcement that no AI assertion ships without a live-validated source; every agent is subject to it, with no exemptions.
- Secrets: credentials scoped per agent across five isolated vault projects, loaded by a boot step that keeps the last known good set if a fetch is partial.
- Tenancy: 94 row-level security policies across 31 tables that treat AI agents as first-class principals, with three explicit deny policies that make decision logs and signed agreements immutable.
- Self-verification ban: a database constraint means no agent can be recorded as the verifier of its own work.
- Autonomy bands: what an agent may do on its own is graded by how reversible it is, green, yellow or red, and enforced as constraints.
- A pre-flight check on every agent session before it starts work.
- A security and tenant architecture memo for client security teams: three-layer fail-closed tenant isolation, secrets handling, AI-layer permissions and auditability by receipts.
Results
- 2,292 AI assertions blocked by the grounding gate, 1,819 of them caught only at the final gate.
- 482 secrets across 56 paths in five isolated vault projects.
- 2,361 records held under the self-verification constraint and 1,891 decisions graded by autonomy band.
- The memo states the deployment posture plainly: it deploys inside the client's tenant and nothing leaves the client's environment.
Stories from this work
Keeping the prompts inside the wallsThe easy way to build AI systems for clients is to put every client on one shared stack and send everything through outside services.Standing on principle
A gate that will not let an unsourced claim out the doorAI agents write fluently whether or not they are right.Standing on principle
At a glance
| Project type | AI governance and security architecture |
|---|---|
| Industry | Software |
| Seat | Founder, Zen Aegis |
| Ran through | Zen Aegis |
| Dates | 2026 |
| Related pages | Zen Aegis |
| Credits | Common Ground Robby Prochnow, Founder, Zen Aegis: designed and built the controls and wrote the memo |