KradleBench - Multi-agent AI benchmark in Minecraft

Show HN: KradleBench – Multi-agent AI evals in Minecraft

KradleBench - Multi-agent AI benchmark in Minecraft

KradleBench is a dynamic benchmark that evaluates AI models in Minecraft, testing collaboration, economy, conflict, logic, and creativity. Unlike static benchmarks, it places agents in messy, noisy environments with other goal-driven agents, requiring responsiveness and adaptation. Top models like GPT-6.2 Astra and GPT-6.2 Sol coordinate actions, generate recognizable objects, and even devise surprising strategies—such as turning diamonds into weapons to eliminate rivals. Every run is replayable as an interactive session. KradleBench reveals that while AI excels at solo challenges, skilled humans still outperform models in strategy and conflict, offering a realistic measure of multi-agent intelligence.

Most AI benchmarks are static. But the real world is not static. It’s messy, it’s noisy, filled with other agents that are pursuing their own goals. Being smart isn’t enough. AIs must be responsive and adapt to their environment.