Modern systems have good observability and better tooling than ever, yet testing often remains manual and slow. In your experience, what aspects of testing can actually run autonomously today, and what still seems fundamentally hard or unsolved?