AgentSafe in front of a Service, in your cluster.

For caller egress, upstream ingress and mesh identity rules, see gateway deployment strategies.

The gateway runs as a Deployment in front of one upstream Service, from the Helm chart in charts/agentsafe. It is the same image the release builds for Docker, and the same runtime as the executable; the chart deploys it with a Service, a ServiceAccount, a ConfigMap, a NetworkPolicy, a PodDisruptionBudget and, when asked, an autoscaler and an on-call Role. The deployment pattern is the gateway: one Deployment per protected Service, in that Service's namespace, the smallest production-grade shape. The trusted executor's own kit, with two namespaces and a StatefulSet, stays in deploy/ for the proposal ingress.

Availability: the chart is packaged and pushed to oci://ghcr.io/decionis/charts/agentsafe by the release workflow from v0.2.0 on. Before that, install from the checkout: helm install agentsafe ./charts/agentsafe ....

Install

The Decionis key is a Secret you create; the chart references it by name and never holds a value.

kubectl create namespace payments
kubectl -n payments create secret generic decionis --from-literal=api-key="$DECIONIS_API_KEY"

helm install agentsafe oci://ghcr.io/decionis/charts/agentsafe --version <chart version> \
  --namespace payments \
  --set decionis.tenantId=<your organization id> \
  --set upstream.service=payments --set upstream.port=8080

A values file says the same, and is where the routes go:

decionis:
  tenantId: 00000000-0000-4000-8000-000000000000
  apiKeySecretRef:
    name: decionis
    key: api-key

gateway:
  mode: shadow
  failurePolicy: FailClosed
  routes:
    - path: /payments/**
      action: payment.create
      methods: [POST]
    - path: /refunds/**
      action: refund.create
      methods: [POST]

upstream:
  service: payments
  port: 8080

image:
  digest: sha256:<the release's image digest>

helm install agentsafe oci://ghcr.io/decionis/charts/agentsafe --version <chart version> -n payments -f values.yaml. The chart's values.yaml documents every value; its schema refuses a mode, a policy or a route the runtime would refuse.

Point the callers at the gateway

Callers reach http://agentsafe.payments.svc:8080 instead of the upstream. In shadow nothing changes for them: every request goes through unchanged and the gateway's log records what Decionis would have decided. The NetworkPolicy admits ingress from networkPolicy.ingressFrom (every pod in the namespace when empty); tightening it, and denying the callers a direct route to the upstream, is what makes the gateway a boundary rather than an option. The security page says how.

When the callers cannot be pointed anywhere, the same boundary is held from beside them: the transparent interceptor is a kustomize Component (deploy/intercept/kubernetes) that redirects a labelled Deployment's outbound 80 and 443 into an agentsafe intercept sidecar, reports every destination the workload reaches, and governs the ones you name under an authority the workload trusts. It needs a namespace under the baseline Pod Security profile, and it says what that authority costs.

Watch it

kubectl -n payments rollout status deployment/agentsafe
kubectl -n payments logs deployment/agentsafe -f
kubectl -n payments port-forward svc/agentsafe 8080:8080
curl -s http://127.0.0.1:8080/_agentsafe/status

Readiness is /_agentsafe/readyz, liveness /_agentsafe/healthz, metrics /_agentsafe/metrics (scrape annotations are on the pods; metrics.tokenSecretRef guards the route with a bearer token). Every event, the chained evidence included, is one JSON line on standard output.

Send your first governed action

From inside the cluster:

kubectl -n payments run curl --rm -it --image=curlimages/curl --restart=Never -- \
  curl -i -X POST http://agentsafe:8080/payments -H 'content-type: application/json' -d '{"amount": 500}'

In shadow the request reaches the upstream and the log line says "state":"SHADOW" with what Decionis would have decided. helm upgrade ... --set gateway.mode=enforcement makes the decisions binding: an ALLOW is forwarded once, an ESCALATE is held (202), a BLOCK is refused (403), and an unreachable authority is 503 with AUTHORITY_UNAVAILABLE, never a BLOCK.

Upgrade, roll back, remove

helm upgrade rolls the pods with no replica unavailable; a changed configuration rolls them through the ConfigMap checksum. helm rollback returns to the previous values and image. helm uninstall removes everything the chart created and leaves the Secret you created.

Production

Production is the checklist: pin the image by digest, enforcement with fail-closed, evidence retention, the NetworkPolicy tightened, alerts on agentsafe_authority_errors_total. High availability says what replicas, the budget and the affinity give, and what a held escalation needs.