rubric-eval benchmarks LLM agent execution traces

Focusing on agent traces instead of final outputs is essential for moving LLM agents from prototypes to reliable production systems. Process-first evaluation identifies hidden risks and inefficiencies that output-based benchmarks miss, while local execution via Ollama ensures data privacy and eliminates API costs. The ability to penalize forbidden tool usage represents a significant step toward safer autonomous agent deployments.