Vertex AI Agent Engine hardening

Production-readiness checklist for Regulus-on-ADK agents deployed to Vertex AI Agent Engine.

Pre-deploy

• ./gradlew regulusAdkDoctor regulusComplianceScan passes.
• Profiles match the regulatory scope agreed with legal / compliance.
• Residency allowlist explicit; no fallback to the empty case.
• CMEK key exists in the same region.
• Audit Kafka topic configured with retention matching the profile.
• Kill-switch persistent store provisioned (not in-memory).

Identity and access

• Service account for Agent Engine has only the necessary IAM roles.
• Operator identities for the kill-switch admin endpoint are SSO-backed.
• Cloud Audit Logs enabled at admin / read / write level for Vertex AI resources.

Network

• Egress restricted via VPC Service Controls where possible.
• Outbound DNS resolution for the LLM provider's endpoints — and only those — allowed.

Observability

• Audit topic monitored (lag + dead-letter rate).
• Prometheus scraping /actuator/prometheus.
• Tracing (OpenTelemetry) wired through to your APM.
• Synthetic transaction every N minutes; alert if it fails.

Disaster recovery

• Kill-switch store has a documented recovery time.
• Audit topic replicated to a second region (or compacted to Object Lock for archive).
• Runbook for "Vertex AI Agent Engine is down" — what happens to in-flight requests, what the customer experience is, how recovery proceeds.

Compliance evidence

• Coverage matrix regenerated and reviewed at each release.
• Quarterly access review on the audit topic + kill-switch store.
• Annual rehearsal of the SAR + erasure + DORA / NIS2 notification paths.