Wrong, not broken
Summary
AWS Insights discusses shifting observability for agentic AI from pre-release assumptions to run-level correctness in production. It introduces CloudWatch Omni for unified visibility across agents, apps, and infrastructure, with 17 evaluators to measure behavior and a workflow to turn production traffic into evaluative datasets. The article highlights the need for explicit criteria, chain-of-causality analysis, and dynamic authority as AI agents become more capable within applications.