An AI system now designs, builds, and playtests its own games, then rewrites them based on its own judgment of whether they are fun.
Researchers built Recursive Game Creator, a four-part loop made up of a Designer, a Builder, a Player, and a Reviewer. The Designer turns instructions and prior feedback into a plan, the Builder codes a candidate game, and a coding-native Player runs scripted policies to generate gameplay data fast, instead of the slow, click-by-click testing other systems rely on. The Reviewer scores that data against game-specific criteria, picks the better version, and sends notes back to the Designer for the next round. On the team's own benchmarks, the system scored 77.89 overall on GameCraft-Bench and hit a 53.2% strict success rate on GameASG-Bench, a 34.1% jump over a same-model baseline, with a 93.4% runtime-check pass rate. A small user study found people played longer and rated the games higher than earlier versions produced the same way.
Most game-generation agents stop once the code compiles. This one treats a working build as a starting point, not a finish line, by creating an automated stand-in for a human playtester that can churn through gameplay sessions far faster than people can. That splits "does it run" from "is it fun," and tries to answer the second question without waiting on human testers.
The code isn't public yet, and the benchmarks are the team's own. Worth revisiting once someone outside the lab gets to decide if these games are actually fun.