Love the idea of turning agent benchmarks into a real-time Bomberman match between LLMs — super fun way to surface speed vs reasoning tradeoffs.