Building agentic workflows with SageMaker AI and Bedrock AgentCore
AWS introduces a pattern for mixing managed foundation models (Bedrock) with custom/domain-specific models (SageMaker AI) in multi-agent workflows without rewriting the agent framework Amazon Bedrock AgentCore runtime enables deploying the entire multi-agent system as a single container while routing to different model hosting paths The architecture uses Strands Agents' "agents as tools" pattern with an orchestrator agent (Claude Haiku 4.5) routing to specialized agents using different models To
Analysis
TL;DR
- AWS introduces a pattern for mixing managed foundation models (Bedrock) with custom/domain-specific models (SageMaker AI) in multi-agent workflows without rewriting the agent framework
- Amazon Bedrock AgentCore runtime enables deploying the entire multi-agent system as a single container while routing to different model hosting paths
- The architecture uses Strands Agents' "agents as tools" pattern with an orchestrator agent (Claude Haiku 4.5) routing to specialized agents using different models
- Token-level observability from SageMaker endpoints requires custom implementation since Strands doesn't provide it by default
- The solution demonstrates cost optimization, data residency control, and model flexibility in a production-ready architecture
Why It Matters
This integration addresses a critical pain point for AI practitioners: the inability to combine managed FMs with custom models in agentic workflows without significant framework modifications. It provides a practical blueprint for organizations seeking cost optimization through domain-specific model deployment while maintaining the capabilities of managed models, all within AWS's ecosystem.
Technical Details
- Architecture: Three-agent system with an orchestrator (Claude Haiku 4.5 on Bedrock), budget agent (Claude Sonnet 4.6 on Bedrock), and financial analysis agent (Qwen 3.5 9B on SageMaker AI), all running inside a single Bedrock AgentCore container
- SageMaker Deployment: Qwen 3.5 9B deployed using vLLM DLC (image vllm:0.22.1-gpu-py312-cu130) on ml.g6e.2xlarge instances with 1x L40S GPU (48GB VRAM), supporting 32K context length
- Authentication: Custom bearer token handling via
SageMakerAuthsubclass with auto-refreshing tokens usingsagemaker.core.token_generator.generate_tokenfor long-running agent sessions - Framework: Strands Agents with "agents as tools" pattern, creating fresh agent instances per invocation; OpenAI-compatible API integration for SageMaker endpoints
- Observability: Token-level observability from SageMaker requires custom OpenTelemetry setup (
strands-agents[otel]) since Strands doesn't provide it by default
Industry Insight
- Organizations can now implement hybrid AI architectures that balance cost (using smaller/custom models for specific tasks) with capability (leveraging managed models for complex reasoning), without vendor lock-in to a single model hosting approach
- The "agents as tools" pattern with fresh instance creation per invocation provides a scalable template for building specialized multi-agent systems that can be deployed across different cloud regions
- AWS's integration of Bedrock AgentCore with SageMaker AI endpoints signals a strategic push toward unified multi-model agent orchestration, encouraging enterprises to adopt modular agent architectures rather than monolithic single-model solutions
Disclaimer: The above content is generated by AI and is for reference only.