Overview

Architecture

Monte Carlo is the agent trust platform. It autonomously observes agents and the data that powers them, finds what breaks, and proposes and verifies fixes, so your agents get better with every fix you accept and you can scale trust at the speed of AI.

Monte Carlo operates as a software-as-a-service with options to extend components of the platform into the customer environment for security and compliance requirements.

Integration types

Monte Carlo connects to two types of integrations:

Components

Cloud Service

Monte Carlo operates a cloud service to support the user interface, monitoring system, Bayes, dashboards, and notification routing. This term is used interchangeably with the Monte Carlo Platform.

Data Store

Object storage bucket for a small sample of individual values or data records from the customer environment that are associated with an alert/incident detected by Monte Carlo.

Collection agent

A collection agent is used as a bridge between the Monte Carlo Cloud Service and customer resources. Not to be confused with the AI agents that Monte Carlo monitors. Collection agents come in two kinds:

  • Cloud-native collection agents run on a serverless service: AWS Lambda, Azure Functions or GCP Cloud Run. Monte Carlo calls the collection agent.
  • The Generic Agent runs as a container on any platform that supports Docker, including Kubernetes on any cloud and on-premises infrastructure. It opens every connection itself, outbound to Monte Carlo, so it needs no inbound ports. See Generic Agent Deployment.

Agent Observability Data Platform

A self-hosted pipeline for agent traces: an OpenTelemetry Collector, ClickHouse and an LLM worker on a Kubernetes cluster in your AWS, Azure or GCP account. Your agents send traces to it, and Monte Carlo queries them through a collection agent. See Agent Observability Data Platform.

Deployments

Options Overview

Monte Carlo offers three different deployment models that vary based on where components of the architecture are deployed. Generally, for best-in-class support and minimal-to-no overhead, we recommend either Cloud or Cloud with Customer-hosted Data Store deployment models for data integrations. Agent integrations currently need a customer-hosted collection agent.

DeploymentData integrationsAgent integrations
CloudSupportedComing soon: Monte Carlo-hosted Agent Observability Data Platform
Hybrid: Customer-hosted Data StoreSupported on AWS, Azure and GCPComing soon: Monte Carlo-hosted Agent Observability Data Platform
Hybrid: Customer-hosted Agent & Data StoreSupported with a cloud-native collection agent (AWS, Azure, GCP) or the Generic Agent (any platform)Supported with the Agent Observability Data Platform on AWS, Azure and GCP. The collection agent that queries it can be a cloud-native collection agent on any of the three, or the Generic Agent on AWS (Azure and GCP coming soon).

Data integrations and agent integrations are deployed separately, and you can use a different deployment for each. For example, you can monitor your warehouse from Monte Carlo's cloud and run the Agent Observability Data Platform in your account. See Combining deployments.

Cloud Deployment

Monte Carlo operates completely as a software-as-a-service (SaaS). A customer-dedicated Object Storage is deployed in Monte Carlo's cloud for any customer data. The Monte Carlo cloud service connects to customer data integrations.

Hybrid: Cloud with Customer-hosted Data Store Deployment

Monte Carlo operates completely as a software-as-a-service (SaaS), with the exception of the Data Store being hosted by the customer. The Monte Carlo cloud service connects to customer data integrations.

This deployment model allows customers to never store record-level data with Monte Carlo while minimizing hosting maintenance and overhead.

Available for AWS, Google Cloud, and Azure.

Hybrid: Customer-hosted Agent & Data Store Deployment

The Monte Carlo Collection Agent and Data Store are deployed in the customer environment.

This deployment model allows the most control but requires the most maintenance and hosting for customers.

Data integrations with a cloud-native collection agent

Available for AWS, Google Cloud, and Azure. Monte Carlo calls the collection agent, and the collection agent connects to your integrations and data store.

Data integrations with the Generic Agent

Available on any platform that supports Docker, including Kubernetes on AWS, Azure, GCP, other clouds and on-premises infrastructure. The Generic Agent connects out to the Monte Carlo Agent Service and receives operations over a persistent HTTPS stream. Integration credentials stay in your secret manager. See Generic Agent Deployment.

Agent integrations with the Agent Observability Data Platform

Available for AWS, Google Cloud, and Azure. Your instrumented agents send traces to the platform's OpenTelemetry Collector, which writes them to ClickHouse. Monte Carlo queries ClickHouse through a collection agent that you deploy in the platform's network: a cloud-native collection agent or, on AWS, the Generic Agent. See Agent Observability Data Platform.

Combining deployments

Data integrations and agent integrations don't need the same deployment. Two common combinations:

Cloud deployment for data integrations, customer-hosted platform for agent integrations. Monte Carlo connects to your warehouse from its cloud, while traces stay in the Agent Observability Data Platform in your account.

Customer-hosted collection agents for both. One collection agent and data store serve your data integrations. A separate collection agent in the platform's network serves your agent integrations.

You can use more than one deployment in the same account. To see which integrations use which deployment, see this FAQ.

Options Comparison

The following matrix compares what each deployment model means from an infrastructure perspective (i.e., the resources you host), a network perspective (i.e., how to establish reachability between Monte Carlo and an integration), an object storage/data store perspective (i.e., where troubleshooting and temporary data is stored), and a maintenance perspective (e.g., any required upgrades or management).

OptionDeployment TypeInfrastructureNetworking*StorageMaintenance
CloudCloudNone. Monte Carlo hosts everything for you.You can connect by allowlisting IP addresses (all clouds), or PrivateLink between Monte Carlo's cloud and your cloud (AWS, Azure only).Troubleshooting data is stored in customer-dedicated object storage in Monte Carlo's cloud.None required.
Cloud w/ Customer-hosted Data Storage (AWS)HybridYou host an S3 bucket.Integrations: as for Cloud. Data store: Monte Carlo reaches the bucket through a VPC endpoint, and you can restrict the bucket to that endpoint.Troubleshooting data is stored in an S3 bucket hosted in customer AWS cloud.None required.
Cloud w/ Customer-hosted Data Storage (GCP)HybridYou host a GCS bucket.You can connect by allowlisting IP addresses (all clouds)Troubleshooting data is stored in the GCS bucket hosted in customer GCP cloud.None required.
Cloud w/ Customer-hosted Data Storage (Azure)HybridYou host an Azure blob storage account and containerYou can connect by allowlisting IP addresses (all clouds) or Private Link.Troubleshooting data is stored in the Azure blob storage account (container) hosted in customer Azure cloud.None required.
Customer-hosted Agent & Data Storage (AWS)HybridYou host a collection agent. This includes a Lambda Function, Storage Bucket, etc.You can connect by deploying the collection agent into your own VPC (or a new VPC) and allowlisting the dedicated IP address, or Peering / PrivateLinking between the collection agent and your resource on AWS.Troubleshooting data is stored in an S3 bucket hosted in customer AWS cloud.Upgrades and monitoring are required.
Customer-hosted Agent & Data Storage Deployment (GCP)HybridYou host a collection agent. This includes a Cloud Run Service, Storage Bucket, etc.You can connect directly or by allowlisting / using VPC service controls between the collection agent and your resource in GCP.Troubleshooting data is stored in the GCS bucket hosted in customer GCP cloud.Upgrades and monitoring are required.
Customer-hosted Agent & Data Storage Deployment (Azure)HybridYou host a collection agent. This includes a Function, Storage Account + Container, etc.You can connect directly or by allowlisting, using VNet outbound traffic configuration (e.g., PrivateLink) between the collection agent and your resource in Azure.Troubleshooting data is stored in the Azure storage account and container hosted in customer Azure cloud.Upgrades and monitoring are required.
Customer-hosted Agent & Data Store (Generic)HybridYou host the Generic Agent container on Kubernetes or Docker, plus object storage and a secret manager. Requires an Enterprise plan with the Advanced Networking add-on.Outbound only: the agent connects to Monte Carlo over HTTPS, so no inbound ports are opened. Optionally over AWS PrivateLink or Azure Private Link.Troubleshooting data is stored in S3, GCS, Azure Blob Storage or S3-compatible storage that you host.Upgrades and monitoring are required. Remote upgrades are not supported.
Agent Observability Data Platform (AWS, Azure, GCP)HybridYou host the platform on EKS, AKS or GKE, deployed with Terraform, plus the collection agent that queries it.The platform's endpoints are internal only. Your AI agents send traces to them from your network, and Monte Carlo queries ClickHouse through the collection agent, which runs in the platform's network.Traces are stored in ClickHouse in your account.Upgrades (Terraform module and Helm chart) and monitoring are required.

*Networking in this table refers to the connectivity between the Monte Carlo Platform and your integration, or the collection agent and your integration. You can learn more about connectivity and support here.


Note that Metadata, Metrics, Query logs, and Aggregated / anonymized statistics are always stored in Monte Carlo's Cloud. Troubleshooting data refers to a small sample of individual values or data records to make it easier for you to investigate, root cause, and resolve an incident or data issue. You can always opt out of storing any troubleshooting data, per integration, regardless of the option you select. See security docs and data sampling docs for more information.

Enterprise connectivity

All deployments work over the public internet with TLS. If your security requirements call for private connectivity, these options are available:

OptionApplies toGuide
IP allowlistingCloud and hybrid deployments, all cloudsNetwork Connectivity
AWS PrivateLinkCloud deployments, AWS data stores and collection agentsAWS PrivateLink. AWS data stores and collection agents use VPC endpoints by default on the V2 platform.
Azure Private LinkCloud deployments, Azure data stores and collection agentsAzure Private Link
Agent Service PrivateLink / Private LinkGeneric Agent on AWS or AzureAWS, Azure
On-premises private connectivityGeneric Agent on-premises, bridged through your cloud connectivityOn-Premises Private Connectivity
Proxies and traffic inspectionGeneric Agent behind an egress proxy or TLS inspectionProxies and Traffic Inspection

Did this page help you?