Is agent X getting worse on real repos?

The weekly score of each model on the public benchmarks, one chart per model, drawn from published runs only.

Every point carries its 95% interval and the number of bugs and trials behind it, so a dip inside the interval is not read as a regression.