AWS publishes reference architecture for building multi-agent AI systems using NVIDIA NIM, Strands Agents, and Amazon Bedrock AgentCore

The solution demonstrates a three-agent campaign review system running in parallel, with a persona reviewer agent evaluating content and producing resonance scores, a validator agent checking against legal and brand guidelines, and a finalizer agent aggregating outputs into consolidated recommendations.

The architecture combines NVIDIA NIM (a managed inference service providing GPU-accelerated LLM responses via OpenAI-compatible APIs), Strands Agents (a multi-agent orchestration framework for coordinating tool-based reasoning workflows), and Amazon Bedrock AgentCore (which provides managed runtime, shared memory across agent invocations, checkpointing and recovery capabilities, and built-in observability for inspecting execution paths and monitoring latency and token usage).

Deployment uses an AWS Serverless Application Model (AWS SAM) template that packages Strands agents and dependencies into a Docker container, provisioning an Amazon API Gateway interface to invoke the agents and enabling AgentCore Observability and AgentCore Memory without manual infrastructure management.