After reading another comment I'm not sure my suggestion is any good, it may not test looking at code and improving it and instead test "writing optimized mandlebrot in java" which it has probably seen some great examples of.
This matches my experience with AI agents. Wiring up the correct feedback and paying attention to ensure they use it is important. Tests and linters are great, but there's usually much more that human devs look at for feedback, including perceived speed and efficiency.