AWS's AgentCore workflow mixes Bedrock and SageMaker AI for agentic systems

Amazon Web Services released a technical deep-dive on combining Amazon SageMaker AI with Amazon Bedrock AgentCore to create production-ready agentic workflows. The core problem: developers often want to mix managed foundation models with their own cost-optimized or domain-specific models, but struggle to avoid rewriting their agent framework. AWS's solution connects both hosting paths through a single Bedrock AgentCore container, using the agents-as-tools pattern from Strands Agents. In the reference architecture, an orchestrator agent running Claude Haiku 4.5 classifies user intent and routes tasks via cross-Region inference. A budget agent uses Claude Sonnet 4.6 for 50/30/20 budget breakdowns with structured Pydantic output. A financial analysis agent deploys Qwen 3.5 9B on a SageMaker AI real-time endpoint using the OpenAI-compatible API, handling stock analysis and portfolio construction with tool-calling. The entire workflow runs inside Bedrock AgentCore, enabling collaborative execution across heterogeneous models.

The post walks through concrete implementation details: deploying Qwen 3.5 9B using the vLLM Deep Learning Container (vllm:0.22.1-gpu-py312-cu130) on an ml.g6e.2xlarge instance with 1x L40S 48GB VRAM, and setting up auto-refreshing bearer tokens via httpx.Auth because SageMaker's OpenAI-compatible API requires token renewal for long-running sessions. AWS highlights three benefits: cost optimization (self-hosted Qwen for specialized financial work), data residency (keeping sensitive data on your own endpoint), and model flexibility. They also solve an observability gap by showing how to obtain token-level metrics from SageMaker endpoints—something Strands doesn't provide by default. Complete source code is available in a linked GitHub repository, with prerequisites including IAM roles with sagemaker:InvokeEndpoint and sagemaker:CallWithBearerToken permissions, plus Bedrock model access.