AgileRL

AgileRL is an open-source Python framework for teams training AI agents on their own tasks, including LLM agents and agents for robotics, finance or logistics. Its evolutionary hyperparameter tuning adjusts training settings during a run, reducing the need for separate tuning trials. The framework uses the Apache 2.0 license.

For open-weight LLMs, it covers supervised fine-tuning, preference tuning with DPO, and reinforcement learning with algorithms such as GRPO and PPO. Agents can learn through multi-turn interactions with OpenEnv environments, or from datasets paired with reward functions. Supported model families include Qwen, Gemma, Granite and Nemotron.

Training can run on your own GPU or across multiple GPUs. LoRA and QLoRA reduce memory needs by training adapters, while PyTorch distributed training and offloading support larger workloads. The LLM stack works with Hugging Face Transformers and PEFT, with vLLM available on Linux. Classic reinforcement learning uses the same library, with support for single-agent and multi-agent tasks, offline RL and contextual bandits. It also works with Gymnasium and PettingZoo environments.

Install agilerl for classic RL, or agilerl[llm] for LLM post-training. Local runs do not require Arena, the optional managed training and deployment service. Arena commands submit jobs to external GPU clusters.