Install the Agents control plane (Jangar) and run an AgentRun end-to-end on kind.

Agents Helm Chart

The Agents chart deploys the Jangar control plane plus the full Agents CRD suite. It is production-minded by default (RBAC, optional PDB/HPA/NetworkPolicy) while still supporting fast local iteration.

What you get

Jangar API and controllers for Agents, Orchestration, Tools, Signals, Schedules, Artifacts, Workspaces
Native runtime that executes AgentRuns as Jobs/Pods in the cluster
A full CRD suite with examples you can apply immediately

Requirements

Kubernetes 1.25+
Helm 3.12+
kind, kubectl, helm installed locally

Kind quickstart (end-to-end)

This flow spins up a kind cluster, installs Postgres, deploys the chart, and runs a smoke AgentRun so you can verify the control plane is operational.

scripts/agents/kind-e2e.sh

The first run may take a while because the script builds a local Jangar image. If the services/jangar/.output bundle is missing, it will run bun install and bun run build to generate it before building the container.

Expected output:

deployment/agents becomes ready
agentrun/agents-workflow-smoke reaches Succeeded

Verify manually:

kubectl -n agents get agentruns
kubectl -n agents describe agentrun agents-workflow-smoke
kubectl -n agents get jobs

Access the control plane UI locally:

kubectl -n agents port-forward svc/agents 8080:80

Then open http://127.0.0.1:8080.

What the script does

The scripts/agents/kind-e2e.sh script performs the following steps:

Creates a kind cluster (or reuses the existing one).
Builds a local Jangar image and loads it into kind.
Installs Postgres via the Bitnami chart in the agents namespace.
Creates the jangar-db-app Secret with the database URL.
Installs the Agents chart using charts/agents/values-kind.yaml.
Applies the smoke AgentProvider/Agent/ImplementationSpec/AgentRun examples.
Waits for the AgentRun to succeed.

Codex PR end-to-end (optional)

This flow runs a Codex agent inside the cluster and creates a GitHub PR. It requires a Codex auth session and a GitHub token with repo write access.

Create the secrets:

kubectl -n agents create secret generic codex-auth \
  --from-file=auth.json="$HOME/.codex/auth.json"

kubectl -n agents create secret generic codex-github-token \
  --from-literal=GITHUB_TOKEN="$(gh auth token)" \
  --from-literal=GH_TOKEN="$(gh auth token)"

Mount the auth secret so the controller can set up a writable Codex home directory:

helm upgrade agents charts/agents --namespace agents --reuse-values \
  --set controller.authSecret.name=codex-auth \
  --set controller.authSecret.key=auth.json

Apply the Codex provider + agent, then run an AgentRun with an inline ImplementationSpec:

kubectl -n agents apply -f charts/agents/examples/agentprovider-native-workflow.yaml
kubectl -n agents apply -f charts/agents/examples/agent-native-workflow.yaml

cat <<'EOF' | kubectl -n agents apply -f -
apiVersion: agents.proompteng.ai/v1alpha1
kind: ImplementationSpec
metadata:
  name: codex-e2e-proof
spec:
  source:
    provider: manual
    externalId: manual:codex-e2e-proof
    url: https://example.com/agents/codex-e2e-proof
  summary: "Codex E2E proof"
  text: |
    Create a new file docs/agents/codex-e2e-proof.md with the single line:
    "Codex E2E test ran on 2026-02-03."
  acceptanceCriteria:
    - File exists with the note
---
apiVersion: agents.proompteng.ai/v1alpha1
kind: AgentRun
metadata:
  name: codex-e2e-pr
spec:
  agentRef:
    name: codex-native-workflow
  implementationSpecRef:
    name: codex-e2e-proof
  runtime:
    type: workflow
    config:
      ttlSecondsAfterFinished: 900
  workflow:
    steps:
      - name: implement
        parameters:
          stage: implement
  workload:
    image: jangar-local:kind
    resources:
      requests:
        cpu: 250m
        memory: 512Mi
  secrets:
    - codex-github-token
  parameters:
    repository: proompteng/lab
    base: main
    issueTitle: "Codex E2E proof"
    issueUrl: "https://github.com/proompteng/lab"
EOF

Wait for the run and confirm the PR:

kubectl -n agents wait --for=condition=Succeeded agentrun/codex-e2e-pr --timeout=1800s
kubectl -n agents get jobs
kubectl -n agents logs job/codex-e2e-pr-step-1-attempt-1

Configuration knobs

The kind script supports the following environment overrides:

CLUSTER_NAME (default: agents)
NAMESPACE (default: agents)
POSTGRES_RELEASE (default: agents-postgres)
POSTGRES_USER (default: agents)
POSTGRES_PASSWORD (default: agents)
POSTGRES_DB (default: agents)
CHART_PATH (default: charts/agents)
VALUES_FILE (default: charts/agents/values-kind.yaml)
SECRET_NAME (default: jangar-db-app)
SECRET_KEY (default: uri)
KUBECTL_CONTEXT (default: kind-<cluster>)
IMAGE_REPOSITORY (default: jangar-local)
IMAGE_TAG (default: kind)
BUILD_IMAGE (default: 1, set to 0 to skip the Docker build)

Production checklist

For production deployments, keep the same chart but switch to production values:

Use an external Postgres with database.secretRef.
Prefer image digests (values-prod.yaml already does this).
Enable podDisruptionBudget and networkPolicy as needed.
Consider autoscaling.enabled for the control plane.
Scope controllers via controller.namespaces or enable cluster scope explicitly.

Troubleshooting

If the AgentRun does not complete:

kubectl -n agents logs deployment/agents
kubectl -n agents get agentruns
kubectl -n agents describe agentrun agents-workflow-smoke
kubectl -n agents get jobs

Common fixes:

Ensure the cluster has enough CPU/memory for Jobs.
Re-run the script to reinstall the chart and smoke resources.
If using a custom image registry, ensure image pull secrets are configured.

Clean up

helm uninstall agents -n agents
helm uninstall agents-postgres -n agents
kind delete cluster --name agents

Agents Helm Chart

On this page