OpenPipe ART

Getting reinforcement learning into an existing agent usually means a rewrite; ART, the Agent Reinforcement Trainer, avoids that by splitting into a client and server that speak the OpenAI chat completions API, so agent code keeps calling what looks like a normal endpoint while the backend records trajectories, computes GRPO updates onto LoRA adapters, and reloads improved weights into a vLLM inference server between rounds. Rewards can be hand-written or delegated to RULER, OpenPipe's judge system that ranks a batch of trajectories with an LLM and eliminates most reward engineering. Training targets models supported by Unsloth and vLLM, including the Qwen and Llama families, running on a local CUDA GPU or on serverless backends such as W&B Training, and example notebooks train agents for games, email research, and tool-use tasks end to end. Distributed as the openpipe-art package on PyPI under Apache 2.0 with docs at art.openpipe.ai, it is among the most practical routes from a working agent prototype to one that learns from its own experience.