The eval suite we build before we let an agent touch production

Agentic systems fail differently from traditional software: not with a stack trace, but with a plausible-sounding wrong action. That means the usual approach of shipping and watching the logs is not enough.

Before any agent touches a real user or a real system, it runs against a hand-curated evaluation set of at least 150 scenarios, including adversarial ones designed to trigger its worst failure modes. We track task success rate, but more importantly, we track the rate of confident but wrong actions, since those are the ones that erode user trust fastest.

Then it runs in shadow mode, making decisions that get logged and reviewed but never executed, for at least a week of real traffic. Only after that do we grant it write access, and even then, behind a kill switch a human can flip in under a second.