Build a multi-account AI agent with AgentCore Gateway and MCP

This article demonstrates how to build a multi-account AI agent architecture that keeps data distributed across line-of-business accounts while enabling centralized agent reasoning through Amazon Bedrock AgentCore Gateway and Model Context Protocol (MCP).

• Deploy agents centrally on AgentCore Runtime in a platform account with LLM inference through Amazon Bedrock
• Each LOB team exposes data and tools as MCP servers running on AgentCore Runtime in their own accounts
• AgentCore Gateway aggregates MCP servers behind a single endpoint with semantic tool discovery and unified authentication
• Use OAuth 2.0 machine-to-machine credentials and AgentCore Identity for cross-account authentication
• Enforce fine-grained authorization with Policy in AgentCore (Cedar) at the Gateway layer based on user JWT claims
• Apply Amazon Bedrock Guardrails for PII redaction and content filtering on user input and agent output
• Enable cost attribution by tagging agent execution roles and tracking tool invocations per LOB
• Monitor agent quality continuously with AgentCore Evaluations and safely roll out versions with A/B testing

This hub-and-spoke pattern enables enterprises to build AI agents that reason over distributed data without centralizing or copying it, while maintaining security, governance, and clear cost boundaries across accounts.