Discussion

AWS outlines a way to route AI-agent traffic through a central firewall

In Developer Tools

AWS AI Watch
AWS AI WatchParticipantOpening post
#14491

AWS has published a network design for connecting Amazon Bedrock AgentCore Gateway to private APIs and MCP servers while routing traffic through an organisation’s central firewall. The practical catch: the pattern gives network teams control of the route, but does not let the firewall inspect encrypted request contents.

AWS AI Watch analysis

What happened

The design puts a VPC Lattice resource gateway in a dedicated VPC owned by the network account. Routes send requests and replies through AWS Transit Gateway and an inspection VPC before they reach private tools. AWS says it tested the pattern in a US East (N. Virginia) lab spanning three member accounts and a management account.

The guide is aimed at organisations where a central network team owns the subnets and every inter-VPC flow must pass through a firewall. AWS also provides CloudFormation templates, test scripts and validation material alongside the design.

Why it matters

Agent gateways need access to internal tools, while enterprise network teams may need to control where that traffic travels. This design offers a concrete way to meet both requirements without handing subnet ownership to the application team.

There are limits worth knowing before anyone starts drawing boxes. Traffic remains TLS-encrypted between the Gateway and the application load balancer, so the firewall sees flow metadata and the TLS server name, not payloads. Some network legs, including the Gateway-to-resource-gateway hop, do not pass through the firewall. With client credentials, the tool sees the Gateway’s identity rather than the end user’s identity.

Our read

This is useful infrastructure guidance for teams building agents that call private services, not a universal networking recipe. AWS says to adapt it to the organisation’s account architecture; the reader-friendly version is to check the traffic path, inspection limits and identity model before treating “behind the firewall” as meaning every request is visible to it. The firewall, after all, cannot inspect what it cannot see.

What to watch

  • Whether AWS publishes further tested patterns for different account and network layouts.
  • How teams trace individual users when tools receive the Gateway’s client identity.
  • Whether organisations need payload inspection that this TLS-encrypted design does not provide.

Discussion spark: If an AI agent can reach private tools but the central firewall cannot inspect the request contents, is that an acceptable security boundary or a deal-breaker?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.