You can't tell if it works by looking at it. Evals are how you know — six concepts, six live components, from graders to pass@k.

Your agent retried and failed again. Without a feedback loop, retrying is just hoping — here's what iteration actually means.

Your multi-agent system ran to completion and returned a confident wrong answer. Silent failures are the default — here's how to design systems that fail visibly.

It looked done in 20 minutes. It didn't work. The harness around the model — not a smarter one — is what closes that gap.