Evaluating AI agents needs environments with two properties that usually trade off: the rules have to be small enough to formalise cleanly, and the strategy has to be rich enough that simple heuristics get crushed by stronger play. Most environments are one or the other. Tic-tac-toe is too thin; modern video games are too thick. Planet Wars sits in the right middle.

The game itself: a few planets that produce ships over time, a clock, no fog of war (in the full-information variant). You can send ships to capture neutral planets, attack enemy ones, or sit back and accumulate. Five-minute rules, deep tactics. A single match finishes in seconds, so you can run thousands of agent-vs-agent matches per evaluation round — which is what evolutionary loops need.

The repo packages the game two ways. Headless batch mode for evolutionary training (pack a generation against a roster of Hall-of-Fame opponents, score on win rate). GUI viewer for replay analysis (game_log_to_html_string exports a self-contained game replay you can scrub through). Both fully observable and partial-information variants are supported.

Built as the agent-evaluation testbed for the fsgp neuroevolution agenda — recent work has been on tier-0 tick-loop cleanup, packed-game scheduling so each evolved agent plays a multi-opponent round per Hall-of-Fame sample, and a GameProgressCallback that streams per-tick state to a W&B dashboard.