AWS pairs LLM-as-a-Judge quality scoring with an autonomous DevOps agent to monitor production multi-agent systems

AWS ML Blog · rss · 2026-09-12

An AWS engineering blog argues that production multi-agent failures often slip past traditional infrastructure monitoring: a poorly scoped supervisor prompt raises no errors but routes 20% of requests to the wrong specialist, and a missing IAM permission can make an agent return empty responses while logs show only successful tool calls. CloudWatch proves systems executed, not that agents helped users.

Using a four-specialist airline reservation system (multi-city bookings, loyalty benefits, travel-policy compliance in a single conversational turn), the authors demonstrate a dual-layer architecture: AgentCore Evaluations continuously scores live interactions with LLM-as-a-Judge across helpfulness, correctness, and goal completion, with reasoning attached to every score; on quality drops it runs pattern analysis over recent low-scoring sessions and recommends prompt, tool-selection, or orchestration fixes. AWS DevOps Agent acts as an on-call engineer, autonomously pulling CloudWatch logs, building resource topology graphs, correlating IAM policies and invocation traces to link an empty response to a missing permission.

The stack includes AgentCore runtime with built-in OpenTelemetry, the Strands Agents SDK (Swarm/Graph/Agents-as-Tools patterns), and the FAST fullstack template, forming a continuous monitor-analyze-improve-deploy feedback loop.

Original post →

More from coding & agent

coding & agent channel →