Integrating Amazon SageMaker AI with Bedrock AgentCore for Multi-Agent Workflows
AWS has detailed a method for building agentic workflows that combine managed foundation models on Amazon Bedrock with custom models hosted on Amazon SageMaker AI, including guidance on achieving end-to-end observability.
Original source published: August 14, 2026
Amazon Web Services has published a guide on creating multi-agent systems that integrate Amazon Bedrock AgentCore with Amazon SageMaker AI. This architecture allows developers to use managed foundation models for orchestration while utilizing domain-specific or cost-optimized models, such as Qwen 3.5 9B, hosted on SageMaker real-time endpoints via an OpenAI-compatible API.
A primary focus of the documentation is implementing observability across these hybrid environments. While Amazon Bedrock AgentCore automatically instruments Bedrock model calls with OpenTelemetry, it does not natively provide generative AI spans for SageMaker-hosted models. To monitor metrics like latency and token usage, developers must manually emit custom OpenTelemetry spans and configure vLLM to include usage data in streaming responses. This approach enables unified tracking of token consumption and performance across different model providers within a single production-ready workflow.