Adaptive agentic intelligence framework using Strands Agents
The Adaptive Agentic Intelligence Framework (AAIF), built with Strands Agents, handles routine tasks locally with small language models and routes complex requests to cloud models through Amazon Bedrock. This post walks through building AAIF with local SLMs through Ollama and Amazon Bedrock, using an Internet of Things monitoring scenario.
The Adaptive Agentic Intelligence Framework (AAIF) takes a novel approach to organizing AI processing by dividing it into three layers that cater to different types of tasks. At the heart of AAIF is the Strands agent, which acts as a smart orchestration layer. This orchestrator is a lightweight small language model (SLM) that classifies incoming requests and routes them to either local processing tools or cloud processing endpoints based on the complexity of the task.
Local processing is handled by SLMs running on on-device hardware through the Ollama runtime. This enables the framework to handle routine tasks such as retrieving device telemetry data and analyzing camera feeds with low latency and no reliance on cloud connectivity. To illustrate local processing, a simulated device telemetry dataset is defined, where each device type (e.g., temperature sensors, security cameras) has predefined attributes like location, temperature, and status.
The second tool in the framework performs vision analysis using a local SLM. This tool processes camera feed images directly on the local hardware, ensuring that sensitive camera data remains on-device and not transmitted to the cloud.
For tasks that require broader context and multi-step reasoning, the framework uses cloud processing through Amazon Bedrock. These requests are routed to frontier models hosted in the cloud, which can perform complex pattern analysis and strategic recommendations that the SLMs running locally are unable to handle effectively.
The framework's architecture is designed to be flexible and adaptable to various deployment scenarios. It can run on edge devices like Raspberry Pi computers with limited memory, in full cloud deployments where all inference is performed through Amazon Bedrock, or in hybrid configurations that combine the speed of local processing with the reasoning power of cloud models.
This adaptability is made possible by separating the routing logic (handled by the SLM orchestrator) from the processing tools, which can be independently modified or swapped out based on the available hardware and network conditions.
Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
