Skip to content

The Opinionated Safety Stack

This architecture defines a Winning Default for each plane, ensuring clear ownership and a single source of truth for autonomous governance.

1. The Stack Map

Plane Winning Default Responsibility
Compute GKE (Autopilot) Isolated execution of agent pods using gVisor (Sandbox).
Identity Workload Identity Securely mapping Kubernetes service accounts to GCP IAM.
Secrets Secret Manager Hardened storage and automatic rotation of RSA keys.
Signing Cloud KMS Hardware-backed asymmetric signing (RSASSA-PSS-2048-SHA256).
Observability Cloud Logging Centralized Source of Truth for all agentic audit trails.
Inference Vertex AI (Gemini) LLM analysis with safety filter enforcement.
Sanitization Model Armor Proactive redaction of secrets/PII before consensus.
Governance Safety Gate Deterministic enforcement of resource and cost boundaries.

2. Why This Stack?

  • Zero-Trust Identity: By using Workload Identity, we eliminate the need for long-lived service account keys, reducing the blast radius of a compromised agent pod.
  • Unified Observability: Rather than splitting telemetry across multiple vendors, we use Cloud Logging as the primary sink. This simplifies the SafetyAnalyzer logic and ensures audit trails are compliant with Google Cloud's data residency standards.
  • Deterministic Guardrails: The Safety Gate sits at the end of the pipeline, ensuring that even if the AI (Vertex AI) suggests a high-risk action, the infrastructure boundaries are physically enforced.

3. Control Points & Ownership Boundaries

A governable platform is defined by clear ownership of its safety boundaries.

Control Point Mechanism Ownership Responsibility
Provenance Workload Identity Platform Team Verifying that pods originate from the trusted GKE pool.
Authorization RSA-PSS Quorum Security Team Cryptographic attestation of multi-agent agreement.
Sanitization Model Armor Privacy Team Proactive enforcement of data-residency and PII policies.
Boundary Safety Gate SRE Team Fiscal and resource quota enforcement.

4. Platform vs. Agentic Layer

This architecture makes a sharp distinction between the Platform Plane (GCP/GKE) and the Incident Intelligence Layer (The Agent Framework).

  • The Platform provides the capabilities (compute, logs, identity).
  • The Agentic Layer provides the governance (consensus, sanitization, safety gates).

Separating these layers ensures that the platform remains stable even as the AI models and safety policies evolve.