AI Garden: A Pixel Art Ecosystem Co-Constructed by AI Agents

Core Point: AI Garden is an innovative open-source project where multiple AI agents autonomously collaborate in a shared pixel grid world to build a dynamically evolving virtual ecosystem with over 130 plants and 19 buildings. Combining generative art, multi-agent systems, and large language models (LLMs), the project exhibits rich emergent behaviors. Users can observe agent interactions and contribute to the project. Original author: juliosuas, published on GitHub on June 10, 2026 (Link: https://github.com/juliosuas/ai-garden).

In recent years, the concept of AI agents has risen. Unlike single-task AI, they can perceive the environment and make autonomous decisions. What kind of emergent behaviors will their interactions produce when multiple AI agents are placed in the same virtual space? AI Garden is exactly a creative experiment exploring this question—it is not only a technical demonstration but also a 'living' digital ecosystem autonomously built and evolved by AI agents, integrating generative art, multi-agent systems, and LLMs to create a unique virtual landscape.

Multi-agent Collaboration System

Each AI agent has unique personality, preferences, and goals (e.g., planting flowers, building structures, maintaining plants). They can perceive each other's actions and adjust their strategies (e.g., if an area has many trees, other agents may build treehouses or develop open spaces). Simple rules lead to complex collective behaviors.

LLM-Driven Decision Making

Agents generate prompts based on the current environmental state (plants, buildings, positions of other agents, etc.) and obtain decision suggestions via LLMs, making their behaviors more flexible and creative (e.g., an agent might suggest "Plant a circle of willows by the lake to create a poetic atmosphere" or "Build a stone bridge connecting the two banks").

Pixel Art and Real-Time Rendering

Adopting a retro pixel style, each plant/building has carefully designed pixel sprites; real-time rendering updates are supported, allowing users to observe garden dynamics and agent activities through a browser.

The multi-agent system exhibits complex characteristics not possessed by individual agents:

1. Spontaneous Community Structure: Agents gather in resource-rich areas to form "towns" with high density of plants and buildings, showing organic planning rather than preset grids;
2. Niche Differentiation: Agents gradually specialize (forest guardians, architects, explorers), improving system efficiency;
3. Art Style Evolution: Different regions present diverse aesthetic styles (symmetric geometry or natural chaos), derived from agent preferences and decision history.

AI Garden is an open project, and community contributions are welcome:

• Add new plants/buildings: Submit pixel materials and descriptions;
• Create custom agents: Define unique personalities and behavior patterns;
• Optimize decision logic: Improve prompt engineering to enhance agent intelligence;
• Develop new features: Add weather, seasonal changes, visitor mode, etc. The project uses a permissive license, encouraging forking and experimentation, and serves as a practical platform for learning multi-agent systems and LLM applications.

The tech stack balances efficiency and performance:

• Frontend: Canvas pixel rendering engine for efficient display of large-scale tiles;
• Backend: Lightweight agent management system handling decision scheduling and state synchronization;
• LLM Integration: Supports OpenAI API and local models (e.g., Ollama);
• Data Storage: Uses event sourcing pattern to store world state, allowing playback of history at any moment;
• Cost Optimization: Caches decisions and batches LLM requests to reduce operational costs, enabling individual developers to run instances.

AI Garden touches on deep AI issues: How do multiple autonomous agents coordinate in a shared environment? How to achieve ordered structures without central control? These questions are relevant to real-world scenarios like multi-robot systems, autonomous driving coordination, and digital twin cities, providing a low-risk experimental platform. Future plans: Introduce more complex ecological simulations (plant growth cycles, resource competition), direct agent communication mechanisms, and user-agent interaction functions. The goal is to evolve into a "digital nature" co-cultivated by humans and AI.