AWS Releases Integration of SageMaker AI with Bedrock AgentCore: Enabling Multi-Agent Workflows and Token-Level Observab
By Mr.Xu
Published: · 8 views
Summary:AWS has introduced a new integration scheme that combines Amazon SageMaker AI's OpenAI-compatible endpoints with Amazon Bedrock AgentCore runtime, enabling developers to build multi-agent workflows. This setup allows each specialized agent to utilize the model best suited for its task, optimizing for cost, data residency, and model flexibility. Additionally, the solution demonstrates how to achieve token-level observability for SageMaker endpoints through custom OpenTelemetry spans, addressing t
AWS Releases Integration of SageMaker AI with Bedrock AgentCore: Enabling Multi-Agent Workflows and Token-Level Observability
AWS has recently introduced a new integration scheme that combines Amazon SageMaker AI's OpenAI-compatible endpoints with Amazon Bedrock AgentCore runtime, empowering developers to build sophisticated multi-agent workflows. The key advantages of this solution include:
- Model Flexibility: Each agent can leverage the model best suited for its task, such as using Claude Haiku 4.5 for user intent classification, Claude Sonnet 4.6 for budget allocation, and Qwen 3.5 9B for financial analysis.
- Cost Optimization and Data Residency: By deploying self-hosted models on Amazon SageMaker AI, developers can better manage costs and data residency strategies.
- Token-Level Observability: The solution also demonstrates how to achieve token-level observability for SageMaker endpoints through custom OpenTelemetry spans, addressing the default limitations of Strands Agents.
Technical Highlights
- Multi-Agent Architecture: The Amazon Bedrock AgentCore runtime enables the creation of complex multi-agent systems where each agent handles specific tasks and collaborates through a unified orchestrator.
- OpenAI-Compatible Endpoint Integration: SageMaker AI's OpenAI-compatible endpoints allow seamless integration of OpenAI models while leveraging SageMaker's GPU acceleration and elastic scaling capabilities.
- Custom OpenTelemetry Spans: To address the default lack of token-level observability in Strands Agents, the solution proposes a method using custom OpenTelemetry spans to extract and record token usage from SageMaker endpoints.
Industry Impact and Developer Recommendations
- Multi-Agent System Development: For developers needing to build complex AI workflows, this solution offers a flexible and efficient implementation path.
- Cost and Performance Optimization: By carefully selecting and deploying models, developers can optimize AI inference costs while maintaining performance.
- Observability Enhancement: Developers are encouraged to focus on token-level observability to better monitor resource consumption and performance of AI models.
Conclusion
AWS's integration of SageMaker AI with Bedrock AgentCore provides a powerful multi-agent workflow construction scheme and addresses the critical issue of token-level observability. This solution not only enhances the flexibility and scalability of AI systems but also offers developers more granular control over costs and performance.
Related Resources
- OpenAI-compatible API for SageMaker AI
- Strands Agents — agents as tools
- Amazon Bedrock AgentCore Observability
- OpenTelemetry generative AI semantic conventions
— END —Source: AWS Machine Learning Blog (2026-08-14)
Tags: #AWS #SageMaker AI #Bedrock AgentCore #Multi-Agent Systems #OpenTelemetry
Community Comments