The author used an AI agent to master the 1988 DOS game Arkanoid, successfully defeating the final boss, DOH, in all 10 test runs. The project involved creating an autonomous player that analyzed the game screen, developed its own strategies, and wrote the necessary code without prior instructions on how to play.
Initial attempts using reinforcement learning on a downscaled screen failed to produce significant results, with the agent struggling to master mechanics like ball deflection. The author then shifted to a new approach where an LLM, specifically Claude Opus 5.5, acted as an autonomous researcher within a sandbox environment, using full-frame visuals and save states.
The AI researcher independently discovered the game's physics, such as the six-zone deflection mechanic, and optimized its gameplay to avoid losing the ball. It also learned to dodge projectiles from the final boss and implemented workarounds for game-breaking glitches, such as ball loops, by analyzing recorded gameplay frames.
A separate, independent program served as a judge to verify the AI's performance. In the final evaluation, the agent achieved 10 consecutive victories against the boss with an average score of 343,977 points. The total cost for the AI's work, based on API usage, was approximately 31 USD, with the entire project costing around 136 USD.
The author concludes that successful AI implementation requires a safe sandbox environment for experimentation, a clear mechanism for automated verification, and integrating the LLM directly into the process rather than using it merely as a coding assistant. This approach mirrors the shift toward Software 3.0, where LLMs act as interpreters for high-level tasks.
