Keeping every error and slow trace covers most services, but model calls tend to fail quietly. I had an agent trace come back green end to end while the agent had done nothing, and a run like that is fast and error-free. So for LLM calls I wouldn't push the routine sample rate anywhere near zero.