A 40-game PuzzleScript benchmark for AI agents is nearly complete

The creator says they handpicked 40 games after reviewing more than 300 and refined the prompts. They expect the benchmark to be the most reliable way to assess agents' abstract reasoning in game environments, but say it's already somewhat saturated and models remain inefficient in their actions. Models are tested through Codex or Claude Code harnesses without code generation. The creator says they won't release the game set publicly.