ART, the Agent Reinforcement Trainer from OpenPipe, is an open-source framework for training multi-step AI agents with reinforcement learning. Instead of only fine-tuning on static examples, ART lets agents learn from experience: they attempt real tasks, get scored, and improve through GRPO training runs. The project has about 10,800 stars on GitHub and is actively developed in 2026.
ART plugs into the agent you already have — a LangGraph workflow, a custom tool-using loop or an MCP-based assistant — and wraps its rollouts into training trajectories. Rewards can come from your own logic or from RULER, ART’s automatic grading system that scores trajectories without hand-labeled data. Training runs on open-weight models such as Qwen, Llama and GPT-OSS, using efficient LoRA updates served through vLLM, so a single GPU node can meaningfully improve an agent’s success rate on tasks like email triage, SQL generation and web research. The result is a practical path from a prompt-engineered prototype to an agent that measurably gets better at its job. ART’s design stays agnostic about where agents run: rollouts can execute inside your existing application, in a sandbox or in a cloud environment, while training happens wherever your GPUs are. Checkpoints and trajectory inspection tools help you see exactly which behaviours a training run reinforced. OpenPipe, the company behind the project, built its reputation on managed fine-tuning for demanding customers, and ART distills that operational experience into open source.
ART is free and open source under the Apache 2.0 license. You run training on your own GPUs and only pay for your infrastructure.
ART is for AI engineers whose agents work but plateau with prompting alone, and who want on-the-job training that raises real task success rates on open-weight models.