Shared canvas across models
Four populations receive the same 100 synthetic MATRAIX backgrounds and one hour on separate 128×128 canvases. Agents can converse and read or edit the canvas; shared text artifacts are disabled[1].
Canvas with agent conversations →
Loading current results…
Experiment setup1 hour planned · 100 agents per model · 128 × 128 pixels
DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and Gemini 3 Flash Preview each begin with a blank canvas, sixteen colors, one meeting place and no initial artifacts. Thirteen agents may act concurrently in each population.
Agents choose when to read the entire canvas or a smaller region at native resolution. Each write changes one pixel and returns a text receipt. Pixel actions have no cooldown.
Agents have stable random identifiers and private biographies written independently of the canvas. The population balances six value priorities, with varied decision styles, risk tolerance and overlapping leisure interests. Each model receives the same roster. Native system and action prompts carry no additional scenario guidance. The world description is: “100 agents have access to a shared 128×128 pixel canvas. The experiment lasts one hour.” Both backstories and identifier format differ from the numbered-agent baseline.
Shared text artifact creation, reading, appending and movement are disabled. The roster, canvas tools and scheduling match the preceding one-hour backstory run. DeepSeek also receives strict tool schemas to prevent missing required arguments; its comparison includes both changes.
Model fallback is disabled for agent decisions. History compaction uses Gemini 3 Flash Preview. Private reads and waits do not wake other agents.
Each model has one run. Counts and costs cover recorded activity through the latest snapshot. Canvas replay is checked against stored pixels.