An Agar.io Engine for Stress-Testing Multi-Agent AI

What it does GoBigger is a game engine that reimagines Agar.io as a multi-agent decision-intelligence sandbox. Players—or autonomous agents—control balls that grow by consuming smaller objects and each other while avoiding larger predators. The environment supports team-based matches with configurable mini-games, exposing a gym-style interface for training reinforcement-learning algorithms.

The interesting bit The README emphasizes that cooperation is more rewarding than solo play, yet violent competition breaks out in tight spaces, creating a natural tension between team coordination and self-interest. This design forces agents to handle rich observation spaces and partial visibility, making it a more realistic proxy for real-world multi-agent problems than fully observable grid worlds.

Key highlights

• ICLR 2023 accepted platform with an academic paper backing the design.
• Scalable team-based matches supporting multiple agents per team with split and eject mechanics.
• Partially observable vision-based observations rather than global state cheating.
• Real-time human-playable modes to test policies against human intuition.
• Configurable mini-games and detailed match settings for controlled experiments.

Caveats

• The README notes the observation space is complicated despite simple rules, so expect a nontrivial learning curve when parsing raw ball coordinates.
• It is tested on CentOS 7.6, Windows 10, and macOS 10.15 with Python 3.6, which may feel dated for modern stacks.

Verdict Researchers studying multi-agent cooperation and competition under partial observability should take a look; if you just want to play Agar.io casually, the human mode exists but the project is clearly built for training bots, not entertainment.