Awesome AI Agents › AI Agent Development

OpenPipe/ART

⭐ 10712 Python added to this list on 2025-07-30 repository created 2025-03-10

Agent Reinforcement Trainer (ART) is an open-source reinforcement learning framework designed to train multi-step agents for real-world tasks using the GRPO algorithm. It enables agents to learn from experience and improve their performance on complex tasks by leveraging large language models (LLMs) such as Qwen2.5, Qwen3, Llama, and Kimi. ART introduces a novel approach called RULER (Relative Universal LLM-Elicited Rewards), which eliminates the need for hand-crafted reward functions by using an LLM as a judge to automatically score agent trajectories. This innovation significantly speeds up development by removing the labor-intensive process of reward engineering, making it general-purpose and effective across various tasks without requiring labeled data or expert feedback. The framework provides an ergonomic and modular training harness that integrates GRPO into any Python application, allowing users to train agents from anywhere, whether on local machines or GPU-enabled environments. ART supports flexible observability and debugging through integrations with platforms like W&B, Langfuse, and OpenPipe. It offers intelligent defaults for training parameters and inference configurations, ensuring efficient and stable training while allowing customization to meet specific needs. ART's architecture separates the client and server components: the client interfaces with the user's codebase to perform agent workflows and gather data, while the server handles inference and training on GPU-equipped machines. The training loop involves executing rollouts, scoring trajectories with rewards, training the model using GRPO, and updating the model iteratively. ART supports various agent tasks demonstrated through example notebooks, including playing games like 2048, Tic Tac Toe, and Codenames, as well as solving temporal clue problems and email retrieval tasks. Installation is straightforward via pip, and the project provides extensive documentation, example notebooks, and community support through Discord. ART is designed to improve the reliability and capability of LLM-based agents by enabling on-the-job training and continuous learning, making it a powerful tool for developing intelligent agents that can adapt to real-world challenges.

https://github.com/OpenPipe/ART

2048-gameagent-trainingcodenamesdebuggingemail-retrievalgpugrpokimilangfusellamallmloramulti-step-agentsobservabilityopen-sourceopenpipepythonqwen2.5qwen3reinforcement-learningreward-functionsrulertemporal-cluetic-tac-toetraining-loopvllmw&bzero-shot-rewards