Your agent retries are probably making things worse

Been running a few agent workflows in production and the biggest pain isn't model quality, it's retry behavior. When a tool call times out, the framework just fires it again. If that call wasn't idempotent, you get duplicate writes, double charges, or two emails sent. I've started wrapping every side-effecting tool with a dedupe key and a short circuit breaker, but it feels like patching something the orchestration layer should handle. Curious what others do here: rely on the framework