Multi-agent reinforcement learning (MARL) is increasingly used in real-world coordination tasks, yet little is known about how MARL systems are engineered in practice. We present an empirical repository-mining study of 187 open-source GitHub repositories comprising 12,205 files (9,957 Python, 2,090 YAML, and 158 JSON). Using automated static analysis and manual repository screening, we examine engineering decisions in algorithm adoption and hyperparameter configuration, environment integration, and multi-agent implementation patterns. Our results show that open-source MARL development is concentrated around a small set of algorithm families, especially PPO, DQN and QMIX, and that algorithm use co-occurs with recurring library and environment combinations. Hyperparameter practice is largely limited to a small set of recurring defaults, while specialized variants tend to involve larger hyperparameter sets than their base algorithms. Repositories most often realize multi-agent behavior through simple structural patterns, such as instantiating multiple agents or defining multiple agent classes, whereas specialized multi-agent abstractions occur less frequently. These findings provide an empirical baseline for current MARL engineering decisions and highlight opportunities to improve standardization in hyperparameter specification, environment integration, and multi-agent abstractions.

Top