Deployment option

Argent in your public cloud VPC

Run Argent Strategic Intelligence in your own AWS, Azure or Google Cloud account, inside a network you define. Use your cloud's keys and locked storage, plus Amazon Aurora and ElastiCache on AWS, and have your cloud refuse model calls that skip Argent.

01 / DEPLOYMENT OPTION

Your account, your network

AWS describes a VPC as a way to "launch AWS resources in a logically isolated virtual network that you've defined." Azure calls the same idea a virtual network. Source: AWS documentation.

In this option Argent runs as containers in a Kubernetes cluster in your own cloud account, inside that network. Your AI agents, often already running in the same cloud, send their requests to the gateway there.

  • Your account: the cluster, the data and the bill are yours.
  • Your network: only the gateway and the console face outside; the chart's network policies keep everything else inside the cluster, where the cluster's network plugin enforces them.
  • Your models: the gateway reaches your cloud's AI services, such as Amazon Bedrock, Azure OpenAI and Google Gemini, and any other provider you approve.

The agents, the models and the checkpoint all sit in the cloud you already run, under the controls you already have.

02 / DEPLOYMENT OPTION

A profile for each cloud

One [Helm chart](/solutions/integrations/kubernetes-helm/) for Kubernetes 1.34 or newer, with a ready profile for each major cloud.

CloudKubernetesWhat the profile sets up
AWSAmazon EKSExposes only the gateway and the console, through AWS load balancers, and keeps every other service inside the cluster
AzureAzure AKSThe gateway's edge handles TLS itself, and the secrets store unlocks from Azure Key Vault through the cluster's workload identity
Google CloudGoogle GKEThe gateway's edge handles TLS itself, and the secrets store unlocks from Google Cloud KMS through the cluster's workload identity

Your registry, one setting. global.imageRegistry swaps the registry host of every image and keeps each digest, so the same signed release runs from an Amazon ECR mirror or any registry you choose.

Sized to the load. Four sizing profiles set resources, storage and default rate limits. With autoscaling turned on (KEDA in your cluster; on by default in the AWS production profile, which keeps at least three copies running), Argent starts extra copies of its gateway when more agents send requests, and they go away when things quiet down.

03 / DEPLOYMENT OPTION

Use your cloud's managed services

Run the data tier in the cluster, or hand pieces to your cloud: Amazon Aurora and ElastiCache on AWS, and each cloud's key service and locked storage.

PieceIn the clusterOr your cloud's service
DatabasePostgreSQLOr Amazon Aurora
CacheValkeyOr Amazon ElastiCache
Secrets storeOpenBaoOr OpenBao unlocked by your cloud's key management service
Evidence archiveObject-locked store in the clusterOr Amazon S3 Object Lock in compliance mode, an Azure Blob container with version-level immutability, or a Google Cloud Storage bucket with object retention turned on

No stored keys for storage. On Azure and Google Cloud the archive signs in with the pod's workload identity, so no storage key is kept anywhere.

Locked for years. Signed daily records are locked for seven years by default, and your operators set the period. An auditor checks them on their own machine. See one-click audit evidence.

04 / DEPLOYMENT OPTION

Make the cloud refuse what skips Argent

Argent decides, swaps, masks and records. Your cloud holds the force: its own controls can refuse Bedrock and Foundry model calls that do not come through the gateway.

  • AWS. A service control policy lets only the gateway's role call Bedrock. For workloads with no internet path, a VPC endpoint policy makes the Bedrock endpoint itself serve only that role.
  • Azure. The Microsoft Foundry account gets a private endpoint in the network that holds the Argent gateway, with public network access and key sign-in turned off. A workload that tries to go around the gateway gets a 403 from Azure itself.
  • Allowed models. Azure Policy can stop anyone deploying a model that is not on your approved list.

Argent checks your setup. A read-only scan of your AWS, Azure or Google Cloud account reports whether the cloud blocks ways around Argent. See model control and shadow AI discovery.

05 / DEPLOYMENT OPTION

Agents that need no keys at all

In your cloud, a workload already has an identity. Argent can use it, so the agent carries no Argent key and no provider key.

CloudWhat the workload sendsWhat Argent does
AWSA short-lived request (good for about five minutes) signed with its own role (an EC2 instance profile, an ECS or EKS task role, or a Lambda role)Asks AWS who signed it, and maps that role to an agent with a human owner
AzureA token from its managed identityChecks the token and maps it to an agent with a human owner
Google CloudA service-account ID token from the metadata serverChecks the token and maps it to an agent with a human owner

The real provider credential stays with the gateway, which signs or adds it on the way out. For Bedrock, a caller holding only an Argent key works too: the gateway signs the upstream call with the one IAM credential it holds. See keyless agents and identity tethering.

06 / DEPLOYMENT OPTION

Measured in real cloud accounts

Every number here comes from a timed run, start to every service ready.

WhereMeasured
AWS, fresh accountEleven runs took 17.7 to 35.1 minutes, a median of 28.9
AzureEvery service ready at 16.5 minutes (one run)
Google Cloud21 and 23.4 minutes from creating the project to ready (two runs)

The production layout we measured on AWS: three availability zones, gateway copies spread across them, an Aurora PostgreSQL writer and reader, ElastiCache Valkey with a replica, and OpenBao as three replicas unlocked by an AWS KMS key.

Upgrades under load: in four production-profile upgrade runs on AWS, between 98,582 and 164,100 gateway requests passed through each upgrade, with 0 failed requests caused by the upgrade.

07 / DEPLOYMENT OPTION

Where it fits, and how a team starts

Public cloud VPC suits a team whose agents and AI models already live in AWS, Azure or Google Cloud. Compare it with air-gapped on-premises and private cloud on the deployment options page.

  1. Install. Put the chart in your EKS, AKS or GKE cluster with your cloud's profile, and choose in-cluster or managed data services.
  2. Route. Point your agents at the gateway, by key or by their cloud identity, and your routes at the models you approve.
  3. Watch, then enforce. Start in watch mode, then turn on enforcement and the cloud's own refusals. Take the walkthrough or browse all capabilities.