Observability for AI Agents

Traces for agent runs look nothing like traces for request/response services: they branch, retry and argue with themselves. This talk covers the span model we settled on, cost attribution per step, and how to debug a run that took forty tool calls.