Agentank uses Claude to evolve tank AI through 1,000 battles

Agentank is a browser-based arena where AI agents like Claude or Codex write the code that controls a tank in battle. Unlike traditional games, you don't drive—you coach. After each fight, you tell Claude what went wrong (e.g., 'be more aggressive' or 'dodge left'), and Claude rewrites the tank's JavaScript code before the next match. The creator ran over 1,000 battles, spending roughly $200 in Claude credits, and observed the tank's behavior shift from cautious to aggressive to bizarrely creative. Each iteration is immediately visible, turning abstract agent loops into a tangible, gamified debugging experience. The platform is open for anyone to bring their own AI agent and see if it can build a surviving tank.

What makes Agentank standout is its transparency: you see exactly how Claude interprets your feedback and how that changes the tank's actions. The creator noted that the agent's code often produces unexpected strategies—like hiding behind obstacles or charging recklessly—that reveal both the power and the quirks of current LLMs. This isn't just a game; it's a sandbox for exploring AI alignment, iterative improvement, and the limits of agentic code generation. For developers, it's a hands-on way to understand how feedback loops shape AI behavior without needing a robotics lab.