The Self-Play AI Passed Its First Gauntlet
One of the three hardest problems in ArcoMage was never the art or the netcode — it was balance. A game with a few hundred hand-written spells has an astronomical number of interactions, and no human team can playtest them all. So we built an opponent that plays the game against itself, millions of times, with no graphics and no waiting.
This week it crossed a real threshold: the current neural-net snapshot now reliably beats its own older versions in headless self-play. That matters because the same engine that makes a tough single-player opponent is also the engine that auto-balances every spell. When a card is too strong, the AI discovers and exploits it long before a human would — and because every spell is defined as data, not buried in code, rebalancing is a one-line change the AI itself proposes.
Translation: ArcoMage gets more balanced the more it's played, not less. No power-creep, no dead cards, no “banned list.” Just a book that keeps getting fairer.