Google and Johns Hopkins Pro2 Agent tops MLE-Bench at 76.0%
The non-obvious insight is that belief tracking of bottlenecks, not more search cycles, drove the jump from 28.8% to 57.6% first-shot medal rates — an architectural lesson transferable to any agent that iterates on code. Builders of research agents and autonomous coding systems should adopt the single-variable experiment pattern and consider whether their loop can detect a wrong-at-core solution. Teams comparing agent frameworks will find the MLE-Bench numbers a useful new benchmark bar.