How an AI Agent Can Query Data Spread Across AWS Accounts Without Copying It
The usual fix for an agent that needs many teams' data is consolidation: copy it into one place, or wire up cross-account permissions into a tangle. AWS describes an alternative built on Model Context Protocol (MCP) and Amazon Bedrock AgentCore Gateway. Each line-of-business team exposes its own data and tools as an MCP server, and the platform account's gateway gives the agent a single endpoint for discovering and invoking tools across the registered teams. The data-minimizing rule sits at query time: only the specific data a request needs flows out, so the underlying datasets never leave the accounts that own them. Ownership and isolation stay where the teams already put them; the agent gets reach without a warehouse.
AI-generated discussion · Different perspectives
The real tradeoff is who holds the keys. Funnelling discovery and invocation through one gateway endpoint in the platform account makes policy enforcement a single choke point — and a single place where an over-broad permission hurts everyone.
Maybe, but the query-time rule does more work than the topology. If only the fields a request actually needs leave their owning account, then the worst case is bounded by what someone asked for, not by what the gateway can reach.