AWS published a Strands Agents example that uses local small models for routine IoT tasks and Amazon Bedrock for complex analysis.
AWS published a technical walkthrough on October 21, 2026, for an Adaptive Agentic Intelligence Framework built with Strands Agents. The example routes routine work to small language models running locally and sends more complex requests to models through Amazon Bedrock. AWS presents the design as an alternative to sending every agent request to a cloud-hosted large model, which can add network latency and API costs.
The demonstration has three layers. A lightweight llama3.2:1b model classifies incoming requests and decides where they should go. Local tools handle tasks such as retrieving device telemetry and analyzing camera images, using models served through Ollama. Requests requiring broader context or multi-step reasoning are directed to a cloud model through Bedrock.
AWS builds the example around IoT monitoring. Its sample agent exposes tools for retrieving sensor data, analyzing a camera feed with llama3.2-vision, and requesting more complex analysis from Bedrock. The walkthrough uses simulated device data rather than a live sensor deployment. AWS says the local vision step keeps camera feeds on the device in this setup, while complex analysis is sent to the cloud service.
The post describes several deployment options, including edge hardware, a cloud-only configuration, and a hybrid arrangement. It specifically says the framework can run on an edge device such as a Raspberry Pi with 8 GB of memory. Its routing model and processing tools are separated so developers can swap or reconfigure components based on hardware and network conditions.
This is an implementation guide, not a report of a newly launched managed AWS service or measured production results. The example requires an AWS account with Bedrock access, Python 3.10 or later, Ollama, the named models, Strands Agents, and AWS tooling. Its practical contribution is a reproducible architecture for developers exploring how to keep selected agent operations local while retaining access to cloud models for harder tasks.