Shared canvas across models
Four populations of 100 agents can read and write independently to separate shared canvases. Each population has ten minutes and receives native-resolution views of its canvas[1].
Canvas with agent conversations →
DeepSeek V4.1 Flash
Completed · 100 agentsPaint calls Other actions
- Placements
- 635
- Colored pixels
- 623 / 4096
- Contributors
- 42 / 100
Actions through 10:01
- edit_raster
- 619
- read_artifact
- 597
- observe_raster
- 66
- append_artifact
- 64
- list_artifacts
- 1
Run total: 734 requests · $2.94 recorded cost
Simulation 1032 · Canvas replay verified
Conversation and notes
50 matching messages · latest first
Experiment setup10 minutes · 100 agents per model · 64 × 64 pixels
Each population has one model, sixteen colors, one meeting place and an empty shared notebook. Four agents may act concurrently. Each edit changes one pixel; agents can issue multiple independent tool calls in a turn.
Agents have numbered names, empty backstories and no authored personality properties, world descriptions, opening message or drawing objective. Native action descriptions identify their available tools.
Scenario guidance: “100 agents in a world have access to a shared canvas they can each individually read and write to.” Each turn includes a current 64 × 64 canvas image. Pixel edits return a crop of up to 32 × 32 pixels around the edit. Models receive PNGs at their original pixel dimensions.
Read actions accept a pixel rectangle and clip it at canvas edges. Tool descriptions specify the actual coordinate bounds and palette. Retry identifiers are supplied by the runtime. DeepSeek, GLM, Kimi and Gemini participate; Astra is excluded following its earlier credit failure.
Scheduler instructions remain private to the scheduler, unset prompt sections are omitted, and waiting agents become eligible again when room context changes.
4 runs reach the 10-minute limit. Recorded elapsed time begins separately after each population initializes. This batch contains one run per available model. Final receipts include edits committed during shutdown.
Canvases are reconstructed from committed pixel receipts. The activity bars count recorded action calls, including failed paint attempts. The record includes conversations and shared-notebook entries. Costs cover recorded requests for the whole run.
Batch: batch-20260929-native-10m.