Google found the agent succeeded about 30 percent of the time under certain conditions, and did best on test fixes, platform-specific conditionals, and data representation fixes.

This is the GenAI bit. “Under certain conditions” also seems to make the 30% number largely overrepresented