Rllm

rLLM is an open source framework for training AI agents with reinforcement learning. A developer keeps the agent code they already have, swaps in a tracked client, defines a reward function, and lets the framework handle training. The aim is minimal code changes regardless of how the agent was built. It works with agent frameworks including LangGraph, SmolAgents, Strands, the OpenAI Agents SDK, Google ADK, or a plain OpenAI client. Adding a rollout decorator to the agent code lets rLLM trace every model call automatically. The documentation lists agents trained with the framework, including a 4B financial analysis agent, a 32B software engineering agent, a 14B coding model and a 1.5B math model, and it links a catalog of community projects built on it. The rLLM command line interface covers the common workflow. Users configure a model provider, evaluate a model against more than 50 benchmarks, launch reinforcement learning training on a dataset, and scaffold a custom agent. Datasets are pulled from HuggingFace automatically, and a local LiteLLM proxy handles API routing. The documentation site covers getting started, cookbooks, projects, an API reference and a blog. It is aimed at researchers and engineering teams training their own agent models. The site publishes no pricing.