MAgent

MAgent is an open-source research platform for many-agent reinforcement learning, published on GitHub by GeekAI (from Tsinghua’s TSAIL group). Where most RL environments simulate one or a handful of agents, MAgent renders and trains populations of hundreds to millions of agents interacting on shared 2D maps — chasing, fleeing, gathering and cooperating — making emergent collective behaviour practical to study. The project has around 1,800 stars; the codebase dates from the platform’s research releases and was last updated in 2022.

Overview

MAgent combines a fast C++ simulation core with a Python API and built-in rendering. Agents perceive local observations and act in grid-world scenarios such as pursuit, battle and gathering, and the platform ships with training pipelines for algorithms like DQN, A2C and parameter-sharing variants. Its design goal is scale: efficient data structures keep very large agent counts tractable on ordinary hardware.

Key Features

• Simulates hundreds to millions of interacting agents on shared maps
• Ready-made scenarios: pursuit, battle, gathering and adversarial games
• Fast C++ core with a Python API and live visualisation
• Built-in RL training loops for value-based and policy-gradient methods
• Configurable agent perception, actions and reward structures
• MIT licensed research code with papers and demos linked from the README

💳 Pricing

Free. MAgent is open source under the MIT license; you download and run it locally at no cost.

Who Is It For

Reinforcement-learning researchers studying multi-agent systems and emergent behaviour, students learning MARL, and simulation developers who need very large agent populations.

The platform grew out of research on large-scale multi-agent systems and has been used in studies of cooperation and competition in big populations; its scenarios include classic testbeds in which agents must learn to pursue, evade or gather resources while observing only their local surroundings.