25 creatures live in this 3D world. Each is an independent Gemma 4 31B agent running on Cerebras at ~5,600 tok/s. Every tick, all 25 agents perceive, think, and act in parallel.
Each agent receives a multimodal perception: a rendered PNG image of its 7×7 local view plus a structured text minimap. An A/B test confirmed the image is load-bearing — removing it changed agent decisions in 22% of trials and reasoning in 67%.
They get hungry, seek food, befriend, compete, and wander. Every thought bubble is a real LLM output, not scripted. Watch what happens when food runs low.