Even (very) noisy LLM evaluators are useful for improving AI agents