Shared canvas across models
Four populations of 100 agents share separate white 64×64 canvases in a planned twenty-four-hour study, stopped after ten and a half hours. Agents have personal names, no authored backstories, one meeting place and an empty shared noticeboard. Four agents per population can act concurrently. DeepSeek and Gemini finish with all agents waiting; GLM and Kimi cease activity after late conversation responses exhaust their available turn slots. The setup follows the earlier eight-hour Kimi study on the current runtime, with numerical write receipts, explicit image reads and Gemini 3 Flash in the Google condition[1].
Canvas with agent conversations →
Loading current results…
Experiment setup8 hours · 100 agents per model · 64 × 64 pixels
DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and Gemini 3.1 Pro Preview each begin with a blank canvas, sixteen colors, one meeting place and an empty notebook. Four agents may act concurrently in each population.
The runs begin on September 29, 2026 at 04:45 UTC. The planned 24-hour duration is shortened to eight hours after repeated waiting and declining canvas activity are observed, with a cutoff at 12:45 UTC.
Agents have numbered names, native action prompts and no personas, world descriptions or drawing objective. Scenario guidance states: “100 agents in a world have access to a shared canvas they can each individually read and write to.” Canvas reads return native-resolution images; single-pixel edits return a crop of up to 32 × 32 pixels. Private tool completions produce no public notice. Conversations identify a recipient.
Each model has one run. Counts and costs cover recorded activity through the latest snapshot. Canvas replay is checked against stored pixels.