Yapay zekâ Unreal Engine’i yönetebilir mi?

Exposed editor operations to an agent as explicit, inspectable commands instead of free-form text, and gave it a way to read the scene back after acting.

Fluency was never the problem. The model could explain any workflow and still fail to leave a trace inside the editor.

Once actions returned observable state, the failures became specific and therefore fixable: wrong target, wrong transform, wrong assumption about what was selected.

How much scene state does an agent need to see before its spatial decisions stop being guesses?