AWS has introduced a method to combine Amazon SageMaker AI with Amazon Bedrock AgentCore to create multi-agent workflows that leverage different models for specialized tasks. The approach allows users to integrate OpenAI-compatible endpoints on SageMaker AI with Bedrock AgentCore runtime, enabling cost optimization, data residency, and model flexibility. The solution involves deploying Qwen 3.5 9B on SageMaker AI and integrating it into a multi-agent system alongside models on Bedrock, with results flowing back through an orchestrator agent.
The architecture connects three model-hosting paths through a single Amazon Bedrock AgentCore container. The orchestrator agent, running on Bedrock, classifies user intent and routes tasks to either a budget agent or a financial analysis agent. The budget agent uses Claude Sonnet 4.6, while the financial analysis agent uses Qwen 3.5 9B through a SageMaker AI real-time endpoint. This setup enables specialized agents to collaborate on complex tasks, with each using the model best suited to its function.
The source text describes the integration mechanics, including how to achieve token-level observability from SageMaker endpoints, which Strands Agents does not provide by default. The solution highlights the need for custom OpenTelemetry spans to monitor token usage for models on SageMaker AI, as the default instrumentation does not recognize them as generative AI calls. This approach ensures that users can track cost, latency, and performance for the Qwen model.
Source: awsml
