Building agentic workflows with SageMaker AI and Bedrock AgentCore
Amazon Web Services Ayush Sharma ● Covered by 2 sources
AWS showed how to run a multi-agent setup that mixes Bedrock models with a Qwen model on SageMaker AI. The twist: SageMaker calls need extra tracing work, or token usage stays invisible.
Based on reporting by Amazon Web Services, Ayush Sharma — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pushing a practical answer to a common agent problem: how to mix managed foundation models with your own cheaper or more specialized models without rebuilding the whole framework around them. In its example, Amazon Bedrock AgentCore runtime sits in the middle, while one agent runs on Bedrock and another reaches out to a Qwen model hosted on SageMaker AI through an OpenAI-compatible endpoint.
The setup is split into three roles. An orchestrator agent uses Claude Haiku 4.5 on Bedrock to classify the request and route it. A budget agent uses Claude Sonnet 4.6 for 50/30/20 budget breakdowns with structured Pydantic output. A financial analysis agent uses Qwen 3.5 9B on a SageMaker AI real-time endpoint for stock analysis and portfolio construction.
The interesting part is not just that this works, but that it works inside one production-style container on AgentCore. AWS says the model choice gives cost optimization, data residency, and flexibility. The financial agent is deployed with the vLLM container, and the example points to an ml.g6e.2xlarge instance with a 48 GB L40S GPU. The endpoint is then wired into a Strands multi-agent system alongside the Bedrock models.
But the post spends a lot of time on observability, and for good reason. Bedrock model calls get token-level spans automatically through AgentCore’s OpenTelemetry instrumentation. SageMaker OpenAI-compatible calls do not. Strands emits spans for agent activity, but not the gen_ai.chat span with token attributes that you’d want if you care about cost or latency. So AWS wraps the SageMaker call in a custom OpenTelemetry span and pulls token counts from Strands’ accumulated_usage metrics.
There’s one more catch: vLLM won’t send usage data in streaming responses unless stream_options includes include_usage set to true. Without that, the token count stays at zero, which makes the whole tracing exercise pointless. AWS also calls out a few operational details: use fresh agent instances per request, turn on AGENT_OBSERVABILITY_ENABLED, and raise X-Ray sampling from the default 1 percent to 100 percent during development if you actually want to see most traces.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous part of agent work that matters most: the glue, the tracing, the boring auth token refreshes. The industry loves talking about multi-agent magic, but real systems are won by people who can explain where every token went. That’s less sexy than a demo, and far more useful.
Read more about this at: Amazon Web Services