Friday, 9 October 2026 · 7 min read
AgentGarten enables evolving agents via code-driven worlds and neural rendering
AgentGarten couples explicit simulators with a shared neural renderer trained by Adversarial Forcing to create real-time interactive environments. Pretrained agents learn efficiently from rendered observations, improving through playbooks across rounds, achieving tool-use strategies in just a few games.
Based on the paper AgentGarten: Code Worlds for Evolving Agents by Jiawei Chi, Shangchen Miao, Zhiyuan Shi, Kailu Wu and 10 others
Challenges in building faithful and realistic interactive environments
Agents learn by interacting with environments, but creating environments that are both faithful (with consistent state, rules, and dynamics) and realistic (with observations matching real-world visuals) is difficult. Simulators provide explicit state and programmable rules but require costly 3D asset creation for visual diversity, limiting scalability.
Video world models generate rich visual observations from data but embed state and dynamics implicitly, making it impossible to inspect or edit environment behavior directly. This trade-off limits the ability to build diverse, controllable, and visually realistic interactive worlds for agent learning.
Scaling agent capabilities requires environments that maintain persistent, editable state and support diverse visual observations. Achieving this across many environments without bespoke asset creation remains a bottleneck for embodied learning and agent evolution.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short