Uninstall

Uninstalling is the inverse of Install: the same install.yaml (or the deploy/ kustomization) that stood the subsystem up also tears it down, so one kubectl delete removes the controller, its RBAC and NetworkPolicy, the Custom Resource Definitions, and the namespace.

Caution

Deleting the CRDs deletes every Agent, Station, AgentDefinition, and ModelCatalog in the cluster (CRDs are cluster-scoped, not confined to ai-agents), which cascades their Jobs and pods. Any run in flight is killed. There is no separate “drain” step — make sure nothing important is still running first.

Uninstall

kubectl delete -f https://github.com/re-cinq/ai-agent-subsystem/releases/latest/download/install.yaml

To remove a specific version you installed by digest, delete with that release’s manifest:

kubectl delete -f https://github.com/re-cinq/ai-agent-subsystem/releases/download/v0.14.0/install.yaml

From a source checkout

If you installed with kubectl apply -k deploy, delete the same kustomization:

kubectl delete -k deploy

With Helm

helm uninstall removes the namespace and everything the chart created, but Helm never deletes CRDs. Delete them yourself afterwards:

helm uninstall ai-agent-subsystem --namespace default
kubectl delete crd agents.agents.re-cinq.com stations.agents.re-cinq.com \
  agentdefinitions.agents.re-cinq.com modelcatalogs.agents.re-cinq.com

What gets removed

• The namespace ai-agents and everything in it: the agent-controller Deployment, the controller and caller RBAC, the run-pod NetworkPolicy, and any agent-secrets Secret you created out of band. Namespace termination drains the contained objects, so this can take a moment.
• The four CRDs (agents, stations, agentdefinitions, modelcatalogs under agents.re-cinq.com) and, with them, all custom resources of those kinds across every namespace. Agents own their Jobs through owner references, so deleting an Agent garbage-collects its Job and pod.

A batch Agent carries the finalizer agents.re-cinq.com/batch-cleanup from submit until the controller has cancelled and deleted its provider-side batch, and only the controller removes it. Delete batch Agents, or let them finish, before removing the controller. If one is stuck Terminating once the controller is gone, remove the finalizer by hand:

kubectl -n ai-agents patch agent <name> --type=json -p '[{"op":"remove","path":"/metadata/finalizers"}]'

The provider-side batch is then left behind and still billed.

Credentials you created out of band are namespaced, so they go with the namespace. Nothing on the node is touched — the subsystem mounts nothing from the host (see Prerequisites).

Keep the CRDs, remove only the controller

To take the operator down for an upgrade or maintenance window without deleting in-flight runs and their history, remove just the namespaced workload and leave the CRDs (and their custom resources) standing:

kubectl -n ai-agents delete deploy/agent-controller

Re-applying install.yaml brings the controller back, and it reconciles whatever Agents it finds — Kubernetes is the only state.

Verify removal

kubectl get ns ai-agents                       # "NotFound" once it finishes terminating
kubectl get crd | grep agents.re-cinq.com      # no output

An empty result from both means the subsystem is fully gone.