Bounded Agentic AI Workflow Engine for Amazon EKS

This Helm delivery option installs the Bounded Agentic AI Workflow Engine into a buyer-managed Amazon EKS cluster. The buyer provides the EKS cluster, RDS database, encrypted S3 artifact bucket, KMS key, networking, and add-ons. Application processing and governed inference remain in the buyer's AWS environment.

Model artifacts are accessed directly by the AWS runtime S3 provider through IRSA. No S3 filesystem mount is used. The Pods access Amazon S3, AWS KMS, and AWS Marketplace Metering through IAM Roles for Service Accounts (IRSA). Create one IAM policy limited to the buyer's artifact prefix and KMS key. Save the following as tacr-workload-policy.json, replacing the placeholders:

{ "Version": "2012-10-17", "Statement": [ { "Sid": "ArtifactBucketLocation", "Effect": "Allow", "Action": "s3:GetBucketLocation", "Resource": "arn:aws:s3:::<artifact-bucket>" }, { "Sid": "ArtifactBucketList", "Effect": "Allow", "Action": ["s3:ListBucket", "s3:ListBucketMultipartUploads"], "Resource": "arn:aws:s3:::<artifact-bucket>", "Condition": { "StringLike": { "s3:prefix": ["artifacts", "artifacts/*"] } } }, { "Sid": "ArtifactObjects", "Effect": "Allow", "Action": [ "s3:GetObject", "s3:PutObject", "s3:DeleteObject", "s3:AbortMultipartUpload", "s3:ListMultipartUploadParts" ], "Resource": "arn:aws:s3:::<artifact-bucket>/artifacts/*" }, { "Sid": "ArtifactEncryption", "Effect": "Allow", "Action": ["kms:Decrypt", "kms:Encrypt", "kms:GenerateDataKey", "kms:DescribeKey"], "Resource": "<kms-key-arn>" }, { "Sid": "MarketplaceMetering", "Effect": "Allow", "Action": "aws-marketplace:MeterUsage", "Resource": "*" } ] }

Create the policy and the IRSA-bound service account. The role trust is restricted to system:serviceaccount:tacr-system:tacr-workload:

export AWS_REGION="<eks-region>" export CLUSTER_NAME="<eks-cluster-name>" kubectl create namespace tacr-system \ --dry-run=client -o yaml | kubectl apply -f - export POLICY_ARN="$(aws iam create-policy \ --policy-name '<workload-policy-name>' \ --policy-document file://tacr-workload-policy.json \ --query 'Policy.Arn' --output text)" eksctl utils associate-iam-oidc-provider \ --cluster "$CLUSTER_NAME" --region "$AWS_REGION" --approve eksctl create iamserviceaccount \ --cluster "$CLUSTER_NAME" --region "$AWS_REGION" \ --namespace tacr-system --name tacr-workload \ --role-name '<workload-role-name>' \ --attach-policy-arn "$POLICY_ARN" \ --override-existing-serviceaccounts --approve

Confirm the role association:

kubectl -n tacr-system get serviceaccount tacr-workload \ -o jsonpath='{.metadata.annotations.eks\.amazonaws\.com/role-arn}{"\n"}'

The chart uses serviceAccount.create=false and the serviceAccount.name=${AWSMP_SERVICE_ACCOUNT} override to consume this service account.

Create the model repository token as a Kubernetes Secret outside Helm values:

read -rsp "Model repository token: " MODEL_TOKEN; echo printf %s "$MODEL_TOKEN" | kubectl -n tacr-system create secret generic tacr-hf-token \ --from-file=token=/dev/stdin --dry-run=client -o yaml | kubectl apply -f - unset MODEL_TOKEN

Render values.aws.yaml.tpl with buyer-owned resource identifiers, then install the supplied chart:

helm upgrade --install tacr ./tacr-platform-0.1.4.tgz \ --namespace tacr-system --values ./values.aws.yaml \ --atomic --timeout 20m

The service is ClusterIP by default. The Marketplace metering CronJob uses the same IRSA-bound service account and reports one cluster_hour record per installed cluster.