Google's Pro² Agent hits 76.0% medal rate on MLE-Bench
The non-obvious lesson is that writing down open problems and candidate fixes, rather than iterating blindly, is what unlocked the benchmark lead — a design pattern practitioners can copy into their own agent loops. Teams building autonomous ML or coding agents should care, as can researchers benchmarking on MLE-Bench against the previous 70.7% baseline. The trade-off is compute: the top score required doubling the time budget to 48 hours.