Over the past six months, @LuozhuZhang and I have been deeply focused on researching agents within the Project Zomboid (by @theindiestone ) environment. Choosing this environment was no coincidence. After immersing ourselves in 500 hours of gameplay, we realized that PZ is more than just a game—it is an ideal environment for studying agents. After 300 hours of exploration, we discovered that PZ's challenges and value for AGI research far exceeded our expectations. Games have always been a testing ground for AGI research. From early Atari games to Go (as demonstrated by AlphaGo), StarCraft II (OpenAI Five, AlphaStar), and more recently Minecraft (MineDojo, Voyager [xitter.cf/DrJimFan/status/159546…]), these games provide clear success metrics that allow research progress to be empirically validated and compared with human performance. However, PZ stands out due to its complex survival systems and highly realistic environmental interactions. Unlike open-ended sandbox worlds like Minecraft, PZ combines exploratory freedom with more complex and dynamic elements. For example, players must adapt strategies in real time, weighing whether to use noise to attract or avoid hordes of zombies. Additionally, PZ introduces emotional states, health metrics, injuries, seasonal changes, and realistic physical interactions, such as manually placing furniture, build shelters, collecting supplies, and the effects of item weight on movement efficiency. These multidimensional dynamics and realistic interactions bring PZ closer to the complexity of the physical world, setting higher demands on agents' perception and decision-making abilities. In PZ, we have conducted exploratory experiments that include integrating Mixture of Experts (MoE) with HydraLoRA to enhance reasoning capabilities, generating synthetic data through self-play, recording inference paths with the "Trace" system, and implementing a "reasoning-action" loop via LLM+RL. Additionally, we have transformed PZ into a universal engine for agent research, enabling efficient testing and development for various reinforcement learning and reasoning tasks. As @LuozhuZhang mentioned, our ultimate goal is to train an agent capable of outperforming 99% of human players—becoming a "PZ Master." xitter.cf/LuozhuZhang/status/187… We also aim to create a multi-agent evolutionary system where agents are randomly deployed across different regions of PZ, making autonomous decisions, collaborating, or competing in complex environments. Imagine live-streaming 20 agents rebuilding civilization in a post-apocalyptic world—how entertaining would that be?🤠 While the final form of AGI remains unknown, we believe our research within PZ can help us develop smarter agents and achieve breakthroughs in reasoning, multi-agent collaboration, dynamic learning mechanisms, and agent personality shaping. These explorations bring us one step closer to realizing the goal of general artificial intelligence.