This is a really interesting idea. Testing with agents and evaluating with them has been one of the harder parts of building LLMs, so this seems really useful.