What AWS published
- The AWS Machine Learning Blog published a how-to for a multi-account agent architecture using Amazon Bedrock AgentCore Gateway and Model Context Protocol (MCP). It is a tutorial, not a product launch (AWS Machine Learning Blog). The stated goal: let each team's data stay in its own line-of-business (LOB) account, with only the data a request needs flowing out at query time, so datasets do not leave their owning account. AWS frames the alternative — replicating data or untangling cross-account IAM — as the problem being avoided (AWS Machine Learning Blog). [1]
How the architecture is described
- Three layers: a central platform account, distributed LOB accounts, and AgentCore Gateway as the integration layer (AWS Machine Learning Blog). The platform account runs the agent on AgentCore Runtime, which AWS describes as a serverless, framework-agnostic environment with session isolation in dedicated microVMs, consumption-based pricing, and built-in authentication. LLM inference runs in the platform account through Amazon Bedrock (AWS Machine Learning Blog). The walkthrough uses a single agent, though AWS says the same pattern supports multiple agents. LOB teams expose data and tools as MCP servers; the platform account's Gateway registers each as a target and gives agents one endpoint for tool discovery with semantic search (AWS Machine Learning Blog). [1]
Governance and controls AWS describes
- Authentication uses AgentCore Identity with Okta; fine-grained authorization uses Policy in Amazon Bedrock AgentCore (Cedar), enforced at the Gateway layer outside the agent's code, as does Amazon Bedrock Guardrails for content safety (AWS Machine Learning Blog). The post also covers observability, plus the option to place Gateway in front of several dedicated inference accounts as an Inference Gateway that routes across model providers and applies per-team rate limits (AWS Machine Learning Blog). AWS notes HTTP targets — AgentCore Runtime agents, agent-to-agent services, and other HTTP endpoints — are supported as additional target types (AWS Machine Learning Blog). All of these are AWS's own descriptions; the evidence contains no independent testing or third-party benchmarks. [1]
Sources
- Build a multi-account AI agent with AgentCore Gateway and MCP
AWS Machine Learning Blog · Primary source ·