MirroS Open-Sources AgentGarten, Where AI Agents Learn by Playing in Code Worlds Rendered Live

MirroS has open-sourced AgentGarten, a framework where code defines physics and game rules while a neural renderer turns geometry into realistic frames. Agents act via short Python programs and write lessons into a playbook: in hide-and-seek, hiders built shelters by round four and seekers used a ramp by round ten, versus 25M and 100M episodes in the 2019 study.

• AgentGarten code, technical report and project page went public on October 9
• Renderer runs above 30 fps at 480p on a single GPU
• Hiders built shelters by round 4, seekers used a ramp by round 10
• Companion-dog score rose from 13 to 19 across four other worlds