Reproduce “AgentGarten: Code Worlds for Evolving Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiawei Chi, Shangchen Miao, Zhiyuan Shi, Kailu Wu, Hanyang Wang, Weiliang Chen, Qiyu Dai, Chengyang Li and 7 more
AgentGarten separates executable world state and rules from a shared neural renderer that generates agents’ observations from engine-provided geometry. Agents act on rendered first-person views, review outcomes, and carry playbooks into later rounds. In hide-and-seek, agents built shelters by round 4 and used ramps by round 10; these milestones reflect pretrained agents and a different learning setup, so they are not directly comparable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.