AWS demonstrates building AI agents using Strands Agents SDK with models deployed on SageMaker AI endpoints and MLflow observability

AWS published a guide showing how to build AI agents by combining Strands Agents SDK (an open source SDK for building AI agents) with foundation models deployed on SageMaker AI endpoints, integrating them with SageMaker Serverless MLflow for agent tracing and A/B testing across model variants.

Organizations deploying models on SageMaker AI gain infrastructure control over compute instances, networking, and scaling; support for different models including custom fine-tuned or open-source alternatives like Llama or Mistral; and cost predictability through reserved instances and spot pricing—capabilities that managed foundation model services do not provide.

The post demonstrates deploying Qwen3-4B model from SageMaker JumpStart as a SageMaker AI endpoint, then creating a SageMaker AI Model provider within Strands Agents to run agents against the deployed endpoint with OpenAI-compatible chat completions APIs; a Jupyter notebook with complete code is available in the GitHub repo.