Which models does the example architecture use?

The example uses Claude Haiku 4.5 on Bedrock as the orchestrator agent, Claude Sonnet 4.6 on Bedrock for budget analysis, and Qwen 3.5 9B deployed on Amazon SageMaker AI for financial analysis and portfolio construction.

Why is the `stream_options: {"include_usage": true}` parameter critical?

By default, vLLM does not include token usage in streaming responses, so Strands receives text chunks but never a final usage object. Adding this parameter tells vLLM to send an extra final chunk with token counts; without it, token tracking reports zero and cost monitoring becomes impossible.

How do you get token-level visibility for SageMaker endpoints in Bedrock AgentCore?

You must manually emit a custom `gen_ai.chat` OpenTelemetry span that wraps the SageMaker agent invocation and extracts token usage from Strands' internal `AgentResult.metrics.accumulated_usage`, since Bedrock AgentCore's automatic instrumentation only recognizes Amazon Bedrock model calls as generative AI operations.

Large Language ModelsAI Coding AssistantsAmazon AI BlogAug 14, 2026Aug 15, 2026, 01:00 JST

AWS shows how to blend SageMaker and Bedrock models in one agent workflow

Send on LINE

3 Key Points

What happened
AWS published a technical guide demonstrating how to build multi-agent systems that combine models hosted on Amazon SageMaker AI (like Qwen 3.5 9B) with managed models on Amazon Bedrock (Claude Haiku 4.5 and Claude Sonnet 4.6) in a single architecture using Amazon Bedrock AgentCore runtime.
Why it matters
Teams building AI agents often need cost optimization, data residency control, or domain-specific models that don't exist in managed services. This approach lets you use the cheapest or best model for each task without rewriting your agent framework, while keeping everything in a production-ready architecture.
What to watch
The post emphasizes a critical implementation detail: custom OpenTelemetry spans are required to capture token usage from SageMaker endpoints, since Bedrock AgentCore's default instrumentation doesn't track SageMaker inference metrics. Without the `stream_options: {"include_usage": true}` parameter on vLLM, token counts report as zero, making cost monitoring impossible.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The post addresses a real operational friction point: most AI teams have models scattered across different platforms. Bedrock hosts managed models, but specialized or cost-optimized models may live on SageMaker, or teams need to keep data in a specific region. Rather than force everything into one service or maintain separate agent frameworks, AWS shows how to unify them under a single orchestrator using the agents-as-tools pattern from Strands Agents. The architecture is pragmatic: an orchestrator agent (Claude Haiku 4.5) classifies user intent and routes to specialized agents, each of which can call its ideal model backend—Bedrock for managed APIs, SageMaker for custom or open-weights models accessed via OpenAI-compatible endpoints.

The deeper lesson in the post is about observability gaps. Bedrock AgentCore automatically instruments agents with OpenTelemetry, which sounds convenient until you realize it only recognizes Bedrock model calls. SageMaker endpoints remain invisible in traces, making cost monitoring and latency debugging impossible. The post's solution—manually emitting `gen_ai.chat` spans with token counts extracted from Strands' accumulated usage metrics—is practical but also a sign that cross-platform agent orchestration requires careful instrumentation planning. The emphasis on `stream_options: {"include_usage": true}` highlights another subtle pitfall: vLLM's default streaming behavior strips token counts, so without this flag, custom observability spans report zero usage even when implemented correctly.

FAQ

Which models does the example architecture use?: The example uses Claude Haiku 4.5 on Bedrock as the orchestrator agent, Claude Sonnet 4.6 on Bedrock for budget analysis, and Qwen 3.5 9B deployed on Amazon SageMaker AI for financial analysis and portfolio construction.
Why is the `stream_options: {"include_usage": true}` parameter critical?: By default, vLLM does not include token usage in streaming responses, so Strands receives text chunks but never a final usage object. Adding this parameter tells vLLM to send an extra final chunk with token counts; without it, token tracking reports zero and cost monitoring becomes impossible.
How do you get token-level visibility for SageMaker endpoints in Bedrock AgentCore?: You must manually emit a custom `gen_ai.chat` OpenTelemetry span that wraps the SageMaker agent invocation and extracts token usage from Strands' internal `AgentResult.metrics.accumulated_usage`, since Bedrock AgentCore's automatic instrumentation only recognizes Amazon Bedrock model calls as generative AI operations.

Amazon AI BlogRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.