AI agent builds a Godot game, playtests it, and revises based on the data

An AI agent was prompted to create a Godot game called Kêr-Is, inspired by the Breton legend of Ys, in which the player must survive twelve tides. The prompt required procedurally generated art and sound, no downloaded assets, and rules separated from rendering so automated playtests could run headlessly. The agent then ran bots of varying strength, including idle, greedy, fixed and steward players, across 200 seeds to check the design. The first attempt met all seven numerical targets, but the results revealed that no tested policy won on 24 of the seeds and that the all-knowing Seer bot won 87% of the time. In a second turn, the agent wrote a full-future search that found winning plans for every seed, which was replayed through the game's rules without changing balance parameters. The author notes this single recorded run is not a benchmark and that a full-future search is not the same as the original playtest bots.