Most game-creation agents are evaluated on narrow tasks: a single engine or a one-prompt browser game. That misses what makes a game ship, according to a new arXiv paper introducing CEGame-Bench.

The benchmark shifts the goal from producing a build to delivering a playable system. A game counts as shipped only when a player can launch it, finish it, and return to a saved world. The authors argue that none of these are guaranteed by a successful build, so agents need to be tested on the full long-horizon process.

CEGame-Bench is positioned as a more realistic measure of game-creation agents, pushing evaluation toward the complete player lifecycle rather than isolated code generation.