Researchers behind a new paper titled Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness argue that generating a playable game is not the same as generating an enjoyable one. While recent game-design agents have made substantial progress in producing playable titles, the authors note that program correctness alone does not guarantee a good player experience.
To close that gap, the framework organizes development around four collaborative roles inspired by real game studios: a Designer, a Builder, a Player, and a Reviewer. The Designer turns user instructions and the Reviewer's feedback into detailed plans, which the Builder converts into candidate games.
A central innovation is the coding-native Player, which creates and executes reusable policies through programmatic interfaces. This allows it to collect diverse gameplay trajectories efficiently, avoiding the evaluation bias the authors attribute to slower GUI-based collection.
The Reviewer then uses carefully designed trajectory-based metrics to infer player preferences, combining visual evidence with explicit textual preferences to judge games against game-specific criteria. The better version is accepted and improvement reviews feed the next round, closing what the authors describe as a recursive loop.
The method reports state-of-the-art overall performance of 77.89 on GameCraft-Bench. On GameASG-Bench it reaches a strict task success rate of 53.2% — a 34.1% improvement over the same-model baseline — and the highest mean runtime-check pass rate among compared methods at 93.4%. A user study indicates longer playtime and higher ratings, and the authors say code is coming soon.