Building agentic workflows with SageMaker AI and Bedrock AgentCore

This article demonstrates how to build multi-agent workflows combining SageMaker AI's OpenAI-compatible endpoints with Bedrock AgentCore runtime for cost-optimized, flexible model deployment with full observability.

• Deploy Qwen 3.5 9B on SageMaker AI using vLLM and integrate with Bedrock AgentCore runtime
• Use Strands Agents' "agents as tools" pattern to route tasks between specialized agents (orchestrator, budget, financial analysis)
• Implement auto-refreshing bearer tokens with httpx.Auth for production-ready SageMaker authentication
• Add custom OpenTelemetry spans to capture token usage from SageMaker endpoints that Strands doesn't instrument by default
• Enable stream_options with include_usage=True in vLLM to expose token counts in streaming responses
• Extract token metrics from Strands' result.metrics.accumulated_usage for cost and latency monitoring
• Deploy entire workflow to Bedrock AgentCore runtime with observability enabled via CloudWatch and X-Ray

This architecture enables cost optimization, data residency, and model flexibility by letting specialized agents use the best-suited model for each task while maintaining production-ready observability across all model providers.