Traditional applications fail loudly: an exception, a timeout, a red dashboard. AI agents fail quietly. They call the wrong tool, loop three extra times, or answer a question they should never have touched, and still return a well-formatted response with a 200 OK. Every metric we have spent fifteen years perfecting says the request was healthy.

In this session, Vikash Agrawal, who leads engineering in Amazon CloudWatch and previously helped build Amazon Q Developer, debugs a production agent in front of the audience. Using a real customer-support agent, real traces, and one genuinely embarrassing out-of-scope answer, he shows why observability for agents is a loop rather than a dashboard: observe the reasoning path, evaluate quality rather than just latency, and prove a fix on real data before shipping it.

You will see what an agent trace actually contains, how a tool failure cascades into model behaviour, why built-in evaluators score a wrong answer as "very helpful," and how a ten-line custom evaluator catches what they miss. Everything is built on OpenTelemetry and works with LangGraph, CrewAI, Strands, the OpenAI Agents SDK and others, wherever they run.

You will leave with a checklist of what to have in place before your agents move beyond the demo stage, and a clear picture of what breaks when you don't.

AI agents fail quietly: wrong tool, wrong answer, 200 OK. Vikash Agrawal (Amazon CloudWatch) debugs a agent on stage, showing how to trace reasoning paths, evaluate quality rather than just latency, and prove fixes on real data. Built on OpenTelemetry, framework-agnostic. Leave with a checklist for taking agents past the demo stage. Demos the recently launched CloudWatch Omni.