An AWS machine learning blog post presents the Adaptive Agentic Intelligence Framework (AAIF), an architecture built with Strands Agents that routes requests between local small language models (SLMs) and cloud-hosted models through Amazon Bedrock.

According to AWS, AAIF uses a lightweight llama3.2:1b model as an orchestrator to classify incoming requests. Local tools handle tasks such as retrieving simulated IoT device telemetry and analyzing camera images with the llama3.2-vision model through Ollama. Requests requiring broader context and multi-step reasoning are sent to a cloud processing tool that calls Amazon Bedrock.

The walkthrough uses an IoT monitoring scenario and includes tools named get_device_telemetry, analyze_camera_feed, and get_expert_insights. The telemetry example uses simulated device data rather than connected production sensors. AWS says the local vision processing keeps sensitive camera feeds on the device, while the local processing layer is designed to operate without a cloud dependency.

The post describes deployments ranging from edge devices, including a Raspberry Pi with 8 GB of memory, to full-cloud and hybrid configurations. It says the routing logic and processing tools can be changed independently to fit available hardware and network connectivity.

The demonstrated setup requires an AWS account with Amazon Bedrock access, Python 3.10 or later, Ollama, specified local models, Strands Agents, the AWS CLI, and Boto3. The post presents the framework as a way to address the cost, latency, and connectivity considerations associated with sending every request to a large cloud model; those benefits are described by AWS rather than independently verified in the post.