In this paper, I describe the development of a unique approach to developing strategies for games in which success can only be measured by the final outcome of the game, preventing the use of heuristics. I created and evaluated evolutionary algorithms, applying them to develop strategies for tic-tac-toe. Strategies are comprised of neural networks with randomly initiated weights. A population of candidate strategies are created, each strategy competes individually against each other strategy, and evolutionary operators are applied to create subsequent generations of strategies. The set of strategies within a generation of the evolutionary algorithm forms a metagame that evolves as the algorithm progresses. Hypothesis testing shows that strategies produced by this approach significantly outperform a baseline of entirely random action, although they are still far from optimal gameplay.