AWS shows how to evaluate multi-agent systems for explainability

Why agent evaluation needs more than response quality

According to AWS, a central challenge as multi-agent systems move from experimentation to production is making sure they stay helpful, accurate and explainable in real-world conditions. Enterprises use such systems for work that spans data sources, tools and business constraints, from supply chain planning to financial analysis and customer operations. These systems coordinate several specialized agents to make decisions, execute workflows and generate recommendations.

AWS argues that traditional evaluation, which looks only at the quality of a model's response, is insufficient for agentic systems. Correctness also depends on tool selection, workflow execution and adherence to business constraints. Amazon Bedrock AgentCore Evaluations is described as a fully managed capability for assessing agent performance across development and production.

The post also points to Amazon Bedrock Guardrails, which provides safeguards such as content filtering, denied topic detection and grounding validation. AWS draws a line between the two: evaluations assess agent quality after execution, while Guardrails enforce safety constraints during execution.

A supply chain assistant as the test case

To make the concepts concrete, AWS builds its example around AnyCompany Retail, a fictitious multinational retailer with ecommerce channels, regional fulfillment centers, distribution centers and thousands of physical stores. The company faces inventory imbalances, with stockouts during promotions in some regions and excess stock in others. Transportation teams must balance delivery speed, carrier capacity and cost.

The solution is built with the Strands Agents SDK, Amazon Bedrock AgentCore MCP Server and AgentCore Evaluations. An orchestrator agent receives a planner's request and delegates to four specialized sub-agents exposed as tools:

• An optimization agent, which calls MCP tools backed by mock Amazon API Gateway REST interfaces that return optimization decisions.
• A distribution agent, which calls recommendation APIs to suggest inventory rebalancing across fulfillment centers, stores and digital channels.
• A routing agent, which calls logistics APIs to recommend carrier and route options.
• An analytics agent, which answers supply chain diagnostics questions.

Each agent runs on Amazon Bedrock AgentCore runtime with AgentCore memory and AgentCore Observability enabled, and the agent loop uses foundation models on Amazon Bedrock.

Three layers of evaluation

AWS proposes a three-layer approach that progressively builds enterprise trust. The first layer uses built-in evaluators that need no setup. Helpfulness serves as a universal baseline, paired with a second evaluator aimed at each agent's main failure mode: Tool Selection Accuracy for the orchestrator, Response Relevance for the optimization and distribution agents, Instruction Following for routing, and Faithfulness for analytics.

The second layer adds custom evaluators that encode domain-specific business rules: constraint satisfaction for optimization, data grounding for distribution, route feasibility for routing, SQL correctness for analytics, and plan coherence for orchestration. These check whether a recommendation respected budget limits, used real inventory data and produced operationally correct output. For the optimization agent, the constraints include budget, warehouse capacity and inventory coverage of demand ≤ qty ≤ 2× demand. Other custom checks include a KPI attainment evaluator for fill-rate or revenue improvement, a risk impact evaluator for stockout and overstock risk, and a tool trajectory evaluator that tests whether the orchestrator routed to the correct sub-agent.

The third layer adds explainability evaluators for trust and auditability. Built-in evaluators assess general response clarity, while custom evaluators verify that agents explicitly articulate their decision rationale, reference supporting data or tool outputs, and explain tradeoffs such as cost versus service level.

On-demand and online modes

AgentCore Evaluations supports two modes. On-demand mode is meant for development benchmarking, regression testing and CI/CD gates. Online mode is for continuous production monitoring and alerts.

The same custom evaluators can be reused online through an OnlineEvaluationConfig object that references the evaluators' Amazon Resource Names and sets a sampling rate, for example 1 to 10% of production traces, plus optional session filters. According to AWS, the service then reads traces from AgentCore Observability, scores them and streams results to Amazon CloudWatch dashboards and alarms. The post itself uses on-demand mode to test the solution.

What it means for teams building agents

For teams moving agents into production, the post offers a pattern: start with built-in evaluators for a quick baseline, add custom evaluators that reflect business rules, then score explainability so that decisions can be audited. Because the same custom evaluators work in both modes, checks written for testing before release can also be applied to live traffic. AWS says this lets teams evaluate both the language quality of a response and the business validity of an agent's decision.