The reference implementation routes routine requests to local small language models through Ollama and sends more complex requests to cloud-hosted models through Amazon Bedrock.
Its routing layer uses llama3.2:1b to classify incoming requests. In the IoT monitoring example, the agent exposes tools for device telemetry, camera-feed analysis with llama3.2-vision, and an Amazon Bedrock call for expert analysis. AWS describes three layers: orchestration, local processing, and cloud processing.
AWS says the architecture can run on an edge device such as a Raspberry Pi with 8 GB of memory, as a fully cloud-based deployment, or as a hybrid setup. The post is a reference walkthrough, not a release announcement for a separately described product.
The source does not provide benchmark results for cost, latency, routing accuracy, or task quality. It also does not establish which Bedrock model should be used, whether the implementation is production-ready, or what access, licensing, and operating costs apply.