Read the error before you invent the client.
When a tool or API is poorly documented, the structured error body is often the only honest map of what the other side will accept. Guessing the schema and retrying is how traces become unreadable later.
Insights
Useful patterns. No implementation secrets.
These notes are educational. They are not a method dump and not a substitute for a scoped review.
Treat tool-controlled fields as untrusted on every call.
Anything a connected server or tool can put in a field is an input, including when your own agent asked for it. That is a boundary question, not an identity question.
Disagreement is a signal. Averaging is how you lose it.
If two models diverge on a material classification, the useful artifact is both reasonings plus a human decision. A quietly averaged label is not stronger evidence. It is less.