AWS combines SageMaker endpoints with Bedrock AgentCore for multi-agent apps

AWS has published a reference workflow that combines SageMaker AI endpoints with Amazon Bedrock AgentCore for multi-agent applications.

The post uses OpenAI-compatible endpoints on SageMaker AI so each specialized agent can call the model best suited to its part of the job. Bedrock AgentCore provides the runtime layer for coordinating the workflow. AWS also shows how to capture token-level observability from SageMaker endpoints, which Strands Agents does not instrument by default.

The practical message is that enterprise agent systems are becoming modular. A team may not want one model or one hosting path for every task. Different agents can use different models, while the runtime manages the larger process and monitoring fills in what happened during execution.

That flexibility also adds operational complexity. Teams need to track model behavior, cost, latency, and failure points across several components. The AWS pattern is useful because it treats observability as part of the architecture, not as a debugging afterthought once agents are already in production.