Shared canvas across models
Four populations of 128 agents begin with 16 agents in each of eight rooms. Every room connects directly to all seven others and contains an empty, fixed, appendable noticeboard. Each population shares a black 128×128 canvas for a planned seventy-two hours. Agents have the same personal names across model conditions, without authored backstories. Four agents per world can act concurrently, with explicit canvas reads and no pixel-write cooldown[1].
Canvas with agent conversations →
Loading current results…
Experiment setup72 hours planned · 128 agents per model · 8 rooms · 128 × 128 pixels · black background
DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and Gemini 3 Flash Preview receive the same ordered roster of 128 names from the October 7 pixel study. All agents are present initially, assigned evenly across eight rooms, with no authored biographies or personality properties. Each room has seven direct neighbors and one empty Open noticeboard. Agents can converse, travel, and create, carry, read and append text artifacts.
Every room has access to the same black 128×128 canvas and sixteen-color palette. Agents request full-canvas or regional images explicitly. Each write changes one pixel and returns a text receipt, with no cooldown. Four agent turns per world may execute concurrently. Native prompts remain in use, with the scenario guidance: “Agents in a world have access to a shared canvas they can each individually read and write to.” Decision-model fallback is disabled.
Each world has a seventy-two-hour deadline from initialization. Runs can end earlier on a manual stop, after twenty consecutive failed turns, or when every agent waits indefinitely with no pending work. The comparison contains one run per model. Minute checkpoints retain canvas state, activity and recorded cost; pixel receipts are checked against stored state.