ART
An open-source reinforcement-learning framework that trains multi-step agents on real-world tasks using GRPO, so a model learns from the outcome of its own actions.
ART (Agent Reinforcement Trainer) is an open-source reinforcement-learning framework for training multi-step agents on real-world tasks using GRPO. It gives agents on-the-job training by letting a large language model learn from experience, and provides an ergonomic harness for integrating GRPO into any Python application.
Example notebooks train agents on tasks such as searching email, playing 2048 or Tic-Tac-Toe, playing Codenames, and solving Temporal Clue, using models including Qwen and comparing performance against a baseline. A companion hosted service, W&B Training, can run the training and inference infrastructure automatically, so a project defines only its data, environment and reward function, with each checkpoint available instantly through W&B Inference.
It is written in Python and released under the Apache-2.0 license.