AWS demonstrates evaluating a supply-chain multi-agent system with AgentCore

AWS’s new technical walkthrough focuses on assessing multi-agent systems for accuracy, task success, and explainability in real-world scenarios. Its example uses fictional multinational retailer AnyCompany Retail, whose proposed assistant handles inventory allocation, distribution adjustments, inventory-health analysis, and routing or fulfillment simulations. Amazon Bedrock AgentCore Evaluations became generally available on March 31, 2026; AWS describes it as a fully managed capability for assessing agent performance in development and production.

The demonstrated system is built with Strands Agents SDK, Amazon Bedrock AgentCore MCP Server, and Amazon Bedrock AgentCore Evaluations. It has an orchestrator and four specialized agents covering optimization, distribution, routing, and analytics. The orchestrator receives a planner’s request and delegates work to the specialized agents, which are exposed as tools. Each agent runs on Amazon Bedrock AgentCore runtime with AgentCore memory and observability enabled, while foundation models on Amazon Bedrock power the agent loop.

The agents connect to different tools: the optimization agent calls MCP tools backed by mock Amazon API Gateway REST interfaces; the distribution agent uses recommendation APIs for inventory rebalancing; the routing agent uses logistics APIs for carrier and route recommendations; and the analytics agent answers supply-chain diagnostics questions. AWS distinguishes AgentCore Evaluations from Amazon Bedrock Guardrails: evaluations assess agent quality after execution, while Guardrails enforce safety constraints during execution. The listed configurable Guardrails include content filtering, denied-topic detection, and grounding validation.