AI Is Superhuman Wherever the Scoreboard Is Cheap

Excellent article. It's why agentic AI has diminishing returns as the scale of the project grows. Towards the mid to end development, the features and fixes become large enough that they are hard to check. Features interact with increasingly many other features and have more surface area. Bugs are more complex and higher risk. For example with a game, at the beginning making mechanics exist, i.e. player jumping, is relatively easy, but towards the end the jump has to have multiple states that are fine tuned if it is to have any value, and this is hard to check and relies heavily on how the gameplay feels.