AgileRL

Why we included this project

Most of the effort in reinforcement learning goes into hyperparameter tuning and long training runs, not the control logic itself. AgileRL tackles that head-on: its evolutionary hyperparameter optimization converges on good settings during a single run, so you skip the usual grid or Bayesian sweeps, and it reports training being roughly ten times faster. The library includes PyTorch implementations for on-policy, off-policy, offline, and multi-agent settings, plus support for fine-tuning LLMs on multi-turn tasks. If you want one maintained codebase that standardizes RL experiment patterns instead of bolting a research implementation together with separate tuning scripts, AgileRL gives you a practical, well-documented base to build on.