An agent produces correct answers in development but slows down and costs more after launch. Which monitoring combination best supports investigation?

1. Only the proportion of responses with more than ten words.
2. Quality samples, request latency, token usage, errors, and relevant request traces.
3. Only a weekly count of registered model names.
4. Only the number of developers on the project.