Deployment View

Installation Artifact

Testbench is shipped as a single bundled manifest (install.yaml) published with each GitHub release. Applying it deploys:

The Testbench operator (Deployment + RBAC + leader election lease) into the testbench-operator-system namespace
The Experiment custom resource definition (testbench.agentic-layer.ai/v1alpha1), cluster-scoped
Five TestWorkflowTemplate resources (setup-template, run-template, evaluate-template, publish-template, visualize-template) in the testkube namespace
A grafana-testkube-dashboard ConfigMap in the monitoring namespace, picked up by the Grafana sidecar

Customisation is performed via Kustomize overlays against the released install.yaml (for example, to override the image tag or relocate the dashboard ConfigMap). The previously published Helm chart is deprecated.

Runtime Topology

Namespace Contents

Namespace	Contents
`testbench-operator-system`	Operator controller manager reconciling `Experiment` resources.
`testkube`	Testkube controller, the five `TestWorkflowTemplate` resources, the generated `TestWorkflow` and dataset `ConfigMap` per `Experiment`, and any `TestTrigger` the operator creates for event-driven re-runs.
Target agent namespace	The `Agent` resource referenced by `spec.agentRef`. The operator resolves its A2A endpoint and injects it into the Run phase.
`monitoring` (or wherever Grafana lives)	OTLP collector receiving published metrics and the Grafana dashboard ConfigMap.

testbench-operator-system

Operator controller manager reconciling Experiment resources.

testkube

Testkube controller, the five TestWorkflowTemplate resources, the generated TestWorkflow and dataset ConfigMap per Experiment, and any TestTrigger the operator creates for event-driven re-runs.

Target agent namespace

The Agent resource referenced by spec.agentRef. The operator resolves its A2A endpoint and injects it into the Run phase.

monitoring (or wherever Grafana lives)

OTLP collector receiving published metrics and the Grafana dashboard ConfigMap.

Container Image

All five phases share a single Docker image (ghcr.io/agentic-layer/testbench/testworkflows). Each TestWorkflowTemplate selects its phase via command arguments, keeping the build and distribution surface to a single image.

Each template exposes its image parameter, so individual phases can be overridden with custom images when needed (typically only the evaluate phase, for custom metric frameworks).

Workflow Composition

For each Experiment, the operator generates one TestWorkflow in the testkube namespace that composes the five templates via Testkube’s use mechanism. All steps share a single emptyDir volume at /app/data, which is how intermediate JSON files are passed between phases. Workflow pods are stateless; persistence (artifacts, execution history) is provided by Testkube.