models
Building agentic workflows with SageMaker AI and Bedrock AgentCore
This article explains how to combine Amazon SageMaker AI with Amazon Bedrock AgentCore to create multi-agent workflows, highlighting deployment steps, integration mechanics, and observability improvements.
AS1 NewsSource: aws.amazon.com
AMZN
COSTBuilding complex, agentic workflows often involves integrating various models and managing their interactions efficiently. This article demonstrates how to combine OpenAI-compatible endpoints on Amazon SageMaker AI with Amazon Bedrock AgentCore runtime, enabling specialized agents to collaborate on tasks while each uses the most suitable model.
The architecture connects three model-hosting paths through a single Amazon Bedrock AgentCore container: an orchestrator agent running Claude Haiku 4.5 on Bedrock, a budget agent utilizing Claude Sonnet 4.6 on Bedrock, and a financial analysis agent deploying Qwen 3.5 9B on Amazon SageMaker AI. The orchestrator routes user requests to the appropriate specialized agent, which then calls its respective model, with results flowing back through the orchestrator.
Deploying Qwen 3.5 9B on SageMaker AI involves using the vLLM Deep Learning Container on an ml.g6e.2xlarge instance, with specific environment configurations. The multi-agent system is built using Strands Agents’ tools pattern, with a custom authentication class to handle token refresh for SageMaker endpoints, ensuring secure and continuous operation.
A key aspect of this setup is achieving token-level observability, which is not provided by default for SageMaker endpoints. The solution involves manually emitting OpenTelemetry spans around model calls, extracting token usage from Strands’ internal metrics, and adding stream options to enable token count reporting in streaming responses.
The article also discusses how to deploy the entire system to Amazon Bedrock AgentCore runtime, configure observability, and extend the architecture for further experimentation, such as swapping in fine-tuned models or implementing cost-aware routing.
In conclusion, this architecture enables cost-effective, flexible, and observable multi-agent workflows by integrating SageMaker AI models with Bedrock AgentCore, providing detailed insights into token usage and system performance.
Provides a comprehensive guide for deploying and monitoring multi-agent AI workflows on AWS, enhancing operational observability and model integration.