Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness

Recursive Game Creator. An Agentic Product-Level Experience-Oriented Game Harness. By Jiajun Chen and three others, from the MMLab at the University of Hong Kong, the University of Hong Kong, and one other institution. Published on the sixth of October.

Game design agents have made substantial progress at generating games that run and can be played. But a program that is correct is not the same thing as a game that is enjoyable. An agent can satisfy every check on whether the code works and still hand a player something rough and dull. The gap is between a prototype and an entertaining game, and correctness alone does not close it.

Recursive Game Creator is a harness aimed at that gap, oriented toward the experience of playing rather than the correctness of the program. It organises development as a loop that repeats, built from four components: a Designer, a Builder, a Player, and a Reviewer. Each round produces a candidate game, plays it, judges it, and feeds the judgment into the next round.

The Designer turns user instructions, and the Reviewer's feedback, into detailed plans. The Builder turns those plans into candidate games. The Player is coding-native: it creates and executes reusable policies through programmatic interfaces, so it can collect diverse gameplay trajectories efficiently. That matters because collecting play slowly through a graphical interface introduces evaluation bias, and the programmatic route mitigates it.

The Reviewer then uses carefully designed metrics computed on those trajectories to infer what players would prefer. It combines that with visual evidence and explicit textual preferences, and evaluates each game against criteria specific to that game. Finally it accepts the better version and writes improvement reviews for the next round, which is what closes the loop.

On GameCraft-Bench, the method reaches a state-of-the-art overall score of seventy-seven point eight nine. On GameASG-Bench, its strict task success rate is fifty-three point two percent, an improvement of thirty-four point one percent over the same-model baseline. Among the compared methods it also has the highest mean runtime-check pass rate, at ninety-three point four percent. A user study adds the human side: longer playtime, and higher ratings. Code is described as coming soon.

What shifts here is the definition of done. Instead of stopping when a game runs, the harness keeps going until a game plays better than the version before it, judged on trajectories a player actually produced.