According to @godofprompt, Langfuse logs every LLM call, tool, and cost with versioned prompts, evals, and datasets to debug failing agents.