Agent ID Architecture
Last Modified: 2026-11-19

Agent ID Architecture

Agent ID is built on a control-plane/data-plane split. The RSA-hosted administration console (control plane) is where administrators connect data sources and identity providers, review discovered identities, and author roles, scopes, and policies. Each gateway runtime (data plane) is where that published configuration is actually enforced, however it is deployed—on cloud or on-premises. A secure, outbound-only management channel lets the cloud-hosted console publish configuration to, and monitor, a gateway runtime without that runtime needing to be reachable from the Internet.

Two product modules share this same architecture: Discover reads from connected data sources to find agents, non-human identities (NHIs), and shadow assets, and surfaces posture findings, risk scores, and ownership; Secure places one or more gateways between agents and the resources they call, so every tool call can be authenticated, authorized by role, evaluated against policy, and optionally escalated for human approval.

Core Components

ComponentRole
Console (control plane)The RSA-hosted administrator UI and API where administrators connect identity providers and discovery data sources, review the registry, and author roles, scopes, and policies. It also hosts the Discovery Engine, Posture & Risk Engine, Publish management, and the audit log.
Gateway (data plane)The deployed runtime that enforces published configuration for every agent call routed through it—can be deployed to the cloud or on-premises, and connects back to the console over a secure, outbound-only management channel.
Identity providerAuthenticates administrators and end users (OIDC) for console sign-in, and supplies the groups that roles map to. Each gateway also trusts an identity provider, which it uses for agent registration (Dynamic Client Registration where supported, or an equivalent custom path).
Connectors / discovery data sourcesIdentity, cloud, security, developer, secrets, and gateway sources that feed the Discovery Engine with the identities, agents, and shadow assets it tracks.
RegistryThe system of record for agents/NHIs, resources, and owners, built from both discovered and registered identities.
Role and policy engineCompiles roles and Cedar policies and distributes them to gateways through Publish management, and the gateway evaluates them for every call.
Posture & risk engineRuns detections against discovered identities to produce posture findings and a per-identity risk score.
Audit logTamper-evident record of administrative and operational activity, isolated per tenant and per gateway.
Agents / MCP clientsThe AI agents and non-human identities that call resources through a gateway, acting with exactly the access of the person or role they are governed by.
Resources

MCP servers or REST APIs (OpenAPI) that an agent reaches through a gateway.

Request and Approval Flows

At a high level, two flows run through this architecture:

  • Discovery Flow

    Connectors pull identities and signals from connected data sources into the Discovery Engine, which populates the Registry and raises shadow assets and posture findings for an administrator or analyst to triage (For more information, see Triaging Shadow Assets and Triaging Posture Findings).

  • Call-brokering Flow

    An agent calls a resource through a gateway; the gateway authenticates the caller, filters visible tools by the caller's role and scopes, evaluates applicable policies, and either allows the call, blocks it, or escalates it to a human approver (MCP elicitation, CIBA backchannel, or email redirect) before completing or rejecting it. Every decision is written to the audit log (For more information, see Agent ID Policies and Human-in-the-Loop Approvals).

Deployment Topologies

The console (control plane) is always RSA-hosted. Each gateway (data plane) is deployed independently and can be cloud-hosted or on-premises. A single tenant can mix topologies across its gateways. For more information, see Deployment Model. An on-premises gateway stays reachable to the console only through the secure, outbound-only management channel, so it does not need to be otherwise exposed to the internet.

Supported Platforms

Deploy the gateway on Kubernetes with Helm.

Requirements

  • Kubernetes 1.29 or later, such as k3s, RKE2, OpenShift, kubeadm, AKS, or EKS

  • Helm 3.8 or later

  • A default StorageClass, for the in-cluster PostgreSQL and OpenBao

  • About 1 vCPU and 2.5 GiB of memory available for the gateway

Network Requirements

  • Management connection: Allow outbound TCP to the management host and port named in your installation bundle (the Agent ID console's mTLS address (shown in the bundle's README), admin_url in values.yaml, port 8443 by default). The gateway uses mutual TLS for this connection, so exclude it from TLS inspection.

  • Web proxy: This release does not support a web proxy. The gateway runtime has no proxy setting, so the connection must go out directly.

  • Outbound HTTPS from the gateway: Allow outbound HTTPS to the MCP servers that the gateway fronts and to the identity provider bound to the gateway.

  • Registry access:

    • The machine that copies the image needs HTTPS access to RSA's registry and to your own registry.

    • The nodes pull images from your registry.

    • The nodes also pull images from Docker Hub and Quay (PostgreSQL, Valkey, OpenBao), unless you mirror those images too or use your own PostgreSQL, Redis, and Vault.

  • Inbound access: MCP clients reach the gateway at its public URL through your ingress (Envoy Gateway, or your own ingress controller). The host needs a TLS certificate and a DNS record. If you use your own ingress controller, turn off response buffering and set read and send timeouts of at least 330 seconds. A tool call that waits for approval keeps its stream open for up to 5 minutes.

High Availability

  • Run at least 2 gateway replicas, spread across nodes. The bundle installs 1 replica by default. Set the number with replicas in values.yaml.

  • Use a managed PostgreSQL and a managed Redis instead of the bundle's in-cluster instances. The in-cluster PostgreSQL and Valkey are single instances.

  • For production, replace the default OpenBao seal, whose key is kept in a Kubernetes Secret, with a cloud KMS key or a Vault Transit engine.