aaabench asks coding agents to build an open-world game in Unreal, and publishes the harness with no results at all

GitHub·low signalXBlueskyLinkedInCopy link

ukanwat/aaabench is a long-horizon benchmark harness that hands a coding agent a real game engine, professional conditions and time, and asks it to build an open-world game. The repo explicitly ships the harness only, no results, at 375 stars and 70 forks since 2026-07-31. Withholding scores while releasing the apparatus is an unusual and defensible choice for a benchmark this expensive to run, and it sits alongside GamePhanes (106 stars in two days, an open-source game coding agent environment and benchmark for Godot) as a second long-horizon game-building eval appearing in the same month.