AWS Publishes Four-Stack Bedrock Blueprint for Observable Agentic Retrieval

AWS has published a deployable pattern for enterprise retrieval agents that must choose among knowledge bases, retrieve more than once, and leave a record of what happened. The reference implementation uses four CloudFormation stacks to combine Bedrock Managed Knowledge Bases, AgentCore, MCP-based tools, dashboards and evaluations in one setup.

The design splits retrieval into two jobs. An agent running in Bedrock AgentCore selects the tool for the knowledge base matching a question’s topic. That tool reaches the selected knowledge base through AgentCore Gateway and the Model Context Protocol, or MCP, which connects models to tools. Inside the knowledge base, AgenticRetrieveStream breaks a question into sub-queries, retrieves iteratively, and returns a grounded answer with citations. The agent can then decide whether to retrieve again before responding.

The managed-service boundary

That comparison defines the blueprint’s limit as well as its convenience. The Gateway exposes each knowledge base through a native connector as an MCP retrieval tool, without a Lambda function or extra container. But the architecture depends on capabilities AWS says are available on its Managed Knowledge Bases rather than customer-managed ones.

Evaluation arrives with the runtime

The four stacks create knowledge bases and their data sources, configure the Gateway, build and host an OpenTelemetry-instrumented agent runtime, and create two CloudWatch dashboards. AWS groups the telemetry into seven layers: knowledge-base health, ingestion, retrieval quality, Gateway activity, the agent’s span tree, token use and evaluation scores.

Those layers do not all use the same data path. The runtime automatically emits spans, token use and metrics, while the sample’s driver notebook publishes custom metrics for retrieval quality, token use and evaluation. A cited answer alone does not show whether the agent selected an appropriate tool or how many tokens a session consumed.

A deliberately narrow routing test

The sample uses separate knowledge bases for a synthetic financial filing and a publicly available report on tornadoes. Each is exposed as its own tool, letting the agent make a topic choice while keeping monitoring signals separate. One knowledge base with two data sources would remove the routing choice the example is intended to demonstrate.

Deployment still requires permissions across Bedrock, AgentCore, IAM, CloudWatch, X-Ray, ECR, CodeBuild, S3, Lambda and CloudFormation, plus model access and CloudWatch Transaction Search. AWS says the runtime stack’s CodeBuild container build takes roughly eight to 10 minutes.