Shared canvas across models
Four populations of 128 agents share separate 64×64 canvases with black backgrounds. Each population begins across eight connected rooms, with initial placement biased toward shared interests. Agents can read nearby backstories, converse, travel, use local text artifacts and place individual pixels[1].
Canvas with agent conversations →
Loading current results…
Experiment setup1 hour · 128 agents per model · 8 rooms · 64 × 64 pixels · black background
DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and Gemini 3 Flash Preview receive the same MATRAIX population. Each room begins with twelve agents sharing one leisure interest and four sharing another. National affiliations and political, social, religious and economic perspectives overlap across these groups. Regional quotas use the preceding population design; country and interest assignments provide coverage rather than estimates of their worldwide prevalence. English is the shared primary language.
Every room connects directly to the other seven and has one empty, fixed, appendable artifact named Notes. Conversation and artifact access are local; the canvas is accessible from every room. Nearby agents see excerpts of existing backstories and can explicitly read the full text. Sixteen agents may act concurrently per population.
The canvas begins black with sixteen colors. Agents choose when to read the whole image or a smaller region at native resolution. Each write changes one pixel and returns a text receipt, with no cooldown. Model fallback is disabled for decisions. Native prompts remain in use, with no additional scenario guidance. The world description is: “128 agents in a world have access to a shared 64×64 canvas they can each individually read and write to. The world has eight connected meeting places. The experiment lasts one hour.”