Configure the AI Verify (v2) step

Add the AI Verify (v2) step inside a Container Step Group to automatically verify your Kubernetes deployments using a combination of statistical and agentic analysis along with crisp insights.

The AI Verify (v2) step deploys lightweight data collection plugins into your Kubernetes cluster to collect metrics and logs, strips personally identifiable information before data leaves your cluster, and produces a pass or fail verdict with actionable analysis. It runs exclusively inside a Container Step Group, which provides the Kubernetes infrastructure the plugins need to execute.

Feature availability

AI Verify is behind the feature flags CDS_CV_AI_VERIFY_NG and CDS_CV_HEALTH_SOURCES_ENABLED. Contact Harness Support to enable these features.

---

What you will learn from this topic

• Container Step Group setup: How to create a Container Step Group and configure its Kubernetes infrastructure so the AI Verify step can run inside it.

• Step parameters: The data collection window, timeout, health sources, and optional resource limits that control how the AI Verify step collects and analyzes data.

• YAML structure: The complete YAML for integrating AI Verify into a Blue-Green deployment pipeline, including a recommended baseline warm-up pattern.

• Execution results: How to read the analysis verdict, agent timeline, and scatter plot in the execution view.

---

Before you begin

• Health sources: You need at least one health source configured for this project. You can create health sources in advance or directly from the AI Verify step configuration. Go to AI-assisted health source configuration to create health sources.

• Kubernetes connector: You need a Harness Kubernetes connector with permissions to create and delete pods in the target namespace.

• Kubernetes namespace: The target namespace must already exist in your cluster. The step runs one pod with one container per health source inside that pod.

Contact Harness Support

If the AI Verify step does not appear in the step palette after the feature flags are enabled, contact Harness Support or your account administrator.

---

Add the AI Verify (v2) step

AI Verify (v2) must be placed inside a Container Step Group. You cannot add it as a standalone step outside a step group.

1

Add a Container Step Group

Open your pipeline and navigate to the Execution section of your deployment stage. After your deployment step, click Add Step Group. In the step group configuration panel, enter a name for the step group.

Enable container-based execution

You must turn on the Enable container-based execution toggle in the step group configuration. Without this, the step group does not expose the infrastructure section and the AI Verify step does not appear in the step palette inside it.

2

Configure the step group infrastructure

In the step group's Container Runtime Infrastructure section, configure where the plugin pod will run:

1. Select Kubernetes Direct as the infrastructure type.

2. Select the Kubernetes Cluster Connector for the cluster where you want the plugin pod deployed.

3. Enter the Namespace where the pod will run. This namespace must already exist.

3

Add and configure the AI Verify step

Inside the Container Step Group, click Add Step. Search for AI Verification and select the AI Verify (v2) tile (with the purple icon). The step configuration panel opens with the Step Parameters and Advanced tabs.

---

Configure step parameters

The Step Parameters tab contains all the settings for the AI Verify step.

AI Verify (v2) step parameters panel showing Name, Data Collection Window, Timeout, Health Sources, and Optional Configuration with resource limits

Parameter	Description	Required
Name	Name of the AI Verify step, displayed in the pipeline execution view.	Yes
Description	Optional free-text description for the step.	No
Data Collection Window	Duration for collecting logs and metrics; select from the dropdown or enter a value such as 5m or 2h.	Yes
Timeout	Maximum time for the entire step. Must exceed the data collection window to allow for analysis processing.	Yes
Health Sources	One or more AI Verify health sources to analyze, referenced by identifier.	Yes
Limit Memory	Maximum memory per plugin container, for example 256Mi. Under Optional Configuration.	No
Limit CPU	Maximum CPU per plugin container, for example 250m. Under Optional Configuration.	No

Name

Enter a name that describes what the step verifies, for example AI Verify Payment Service or Verify Production Deployment. This name appears in the pipeline execution view.

Data Collection Window

The data collection window controls how long the plugin pod collects data from your observability platforms. The window begins when the step starts, so your deployment should be fully rolled out before the AI Verify step runs. Common values are 5m for most deployments, 3m for quick verifications, and 10m for services that require longer observation periods.

Timeout

The timeout must exceed the data collection window because the step includes both the data collection phase and the analysis processing phase. If you set the data collection window to 5m, set the timeout to at least 10m to allow enough time for plugin startup, data collection, and analysis. If the step exceeds the timeout, it fails regardless of analysis state.

Recommended timing

A data collection window of 5m with a timeout of 10m works well for most deployments. Add a Shell Script step before the AI Verify step group to pause for one to two minutes after your deployment completes. This gives the new pods time to emit healthy baseline logs before the collection window opens.

Health sources

Select the health sources you want to analyze. The step group runs a single pod with one plugin container per health source. Each container runs its own analysis independently and produces its own verdict. You can mix log-type and metric-type health sources in the same step.

Click + Add to open the health source selector and choose from health sources created at the project, organization, or account level. Each entry in the list shows:

• Health source name

• A LOG or METRIC badge indicating the data type

• Edit and Delete icons

Health sources are referenced by their identifier, not their display name. Go to AI-assisted health source configuration to view and copy health source identifiers.

Optional configuration: Resources

The Optional Configuration section exposes resource limits for the plugin containers. These default to 256Mi memory and 250m CPU, which handle typical workloads. Expand this section to override the defaults.

• Limit Memory: Maximum memory per plugin container. Increase to 512Mi or higher if you are analyzing high-volume data sources or see out-of-memory errors in pod logs.

• Limit CPU: Maximum CPU per plugin container. Increase to 500m or 1000m if you see CPU throttling. Both fields accept Kubernetes resource notation.

---

Write the YAML configuration

You can configure the AI Verify (v2) step in YAML. The step type is AIVerifyNG and must be inside a stepGroup with a stepGroupInfra block that defines the Kubernetes infrastructure.

The example below shows a complete Blue-Green deployment pipeline that includes a baseline warm-up step before the AI Verify step group. The warm-up step pauses briefly after the stage deployment completes to give the new pods time to emit healthy baseline data before the collection window opens.

Blue-Green pipeline with AI Verify (v2): full YAML

pipeline:
  name: BlueGreen_AI_Verify_Pipeline
  identifier: BlueGreen_AI_Verify_Pipeline
  projectIdentifier: your_project
  orgIdentifier: default
  stages:
    - stage:
        name: BlueGreen Deployment Stage
        identifier: BlueGreenDeploymentStage
        type: Deployment
        spec:
          deploymentType: Kubernetes
          service:
            serviceRef: my_service
          environment:
            environmentRef: my_environment
            deployToAll: false
            infrastructureDefinitions:
              - identifier: my_k8s_infrastructure
          execution:
            steps:
              - step:
                  name: Stage Deployment
                  identifier: stageDeployment
                  type: K8sBlueGreenDeploy
                  timeout: 10m
                  spec:
                    skipDryRun: false
                    pruningEnabled: false

              - step:
                  type: ShellScript
                  name: Warm Up Baseline
                  identifier: warmUpBaseline
                  timeout: 5m
                  spec:
                    shell: Bash
                    onDelegate: true
                    source:
                      type: Inline
                      spec:
                        script: |
                          echo "Waiting for the stage pod to emit healthy baseline logs"
                          sleep 120
                    environmentVariables: []
                    outputVariables: []

              - stepGroup:
                  name: AI Verify
                  identifier: aiVerifyGroup
                  steps:
                    - step:
                        type: AIVerifyNG
                        name: AIVerifyV2
                        identifier: AIVerifyV2
                        spec:
                          dataCollectionWindow: 5m
                          healthSources:
                            - healthSourceRef: my_health_source
                          resources:
                            limits:
                              cpu: 250m
                              memory: 256Mi
                        timeout: 10m
                        failureStrategies:
                          - onFailure:
                              errors:
                                - AllErrors
                              action:
                                type: MarkAsFailure
                  stepGroupInfra:
                    type: KubernetesDirect
                    spec:
                      connectorRef: my_k8s_connector
                      namespace: my_verify_namespace
                  failureStrategies:
                    - onFailure:
                        errors:
                          - AllErrors
                        action:
                          type: MarkAsFailure

              - step:
                  name: Swap Primary and Stage Services
                  identifier: swapPrimaryAndStage
                  type: K8sBGSwapServices
                  timeout: 10m
                  spec: {}

            rollbackSteps:
              - step:
                  name: Swap Primary and Stage Services
                  identifier: rollbackSwapPrimaryAndStage
                  type: K8sBGSwapServices
                  timeout: 10m
                  spec: {}

        failureStrategies:
          - onFailure:
              errors:
                - AllErrors
              action:
                type: StageRollback

  delegateSelectors:
    - my_delegate

stepGroup fields:

• stepGroupInfra.type: Must be KubernetesDirect.

• stepGroupInfra.spec.connectorRef: Identifier of your Kubernetes connector.

• stepGroupInfra.spec.namespace: Namespace where the plugin pod runs.

• failureStrategies: Set at the step group level to control what happens if the verification fails. MarkAsFailure fails the stage without rolling back; StageRollback triggers the rollback steps.

AIVerifyNG step fields:

• spec.dataCollectionWindow: Duration string such as 5m or 2h.

• spec.healthSources: List of health source references. Each entry requires a healthSourceRef key. The step runs one plugin container per health source inside a single pod.

• spec.resources.limits.cpu: CPU limit per plugin container, for example 250m.

• spec.resources.limits.memory: Memory limit per plugin container, for example 256Mi.

• timeout: Maximum step duration. Place this at the step level, not inside spec.

---

View execution results

Once the pipeline runs, click the AI Verify step to open its execution details. The Analysis tab shows the verification results for each health source.

The verification status banner at the top shows whether the analysis passed or failed. When issues are detected, the interface shows a breakdown with Summary, Root Cause Hypothesis, Recommendation, Evidence, and Actionable Fix sections. These insights come from LLM contextualization of the statistical anomaly detection results.

Analysis summary panel showing the pass/fail verdict and LLM-generated insights including root cause and recommendations

Below the summary, the Analysis tab shows the full execution view with an agent timeline on the left, a scatter plot in the center, and execution metadata at the top.

Complete analysis tab view with agent timeline and scatter plot visualization

The agent timeline displays each processing phase with its duration. You will see Clustering Agent, Data Gatekeeper Agent, and other analysis components listed in sequence. Click the arrow next to any entry to expand the detailed logs for that phase.

The scatter plot visualizes analyzed data points, showing Critical Regression and Non-Blocking Regression classifications. Normal behavior clusters together while anomalies appear as outliers.

Scatter plot showing anomaly detection results with critical and non-blocking regression classifications

Click View Detailed Analysis to open a full-screen view with cluster-by-cluster breakdowns, individual log patterns, and metric anomaly timelines.

Events list tab showing all health events with Healthy, Warn, and Unhealthy indicators

The Events List tab provides an alternative view showing all events with health indicators (Healthy, Warn, Unhealthy) and event types such as NON_BLOCKING_REGRESSION or IGNORED_CLUSTER. The Details tab shows execution metadata such as start time, duration, and timeout. The Input and Output tabs show the step's configuration parameters and execution outputs respectively.

---

Next steps

You have added the AI Verify (v2) step to your pipeline. The step will now analyze your deployment automatically each time the pipeline runs.

• AI-assisted health source configuration: Create and manage the health sources that AI Verify references to collect logs and metrics.

• AI Verify (v2) overview: Understand how statistical anomaly detection and LLM contextualization work together to produce verification verdicts.

• Interpret log results: Read and act on the log analysis results from verification runs.