we did somes stuff on evals for products, but i don't think we've focused on only harnesses. However i think the same principals apply w/o much change. Harness vs LLM API changes the implementation detail, but i suspect the eval workflow should be exactly the same! Check out: and