rLLM
An open-source framework for training language agents with reinforcement learning, letting a developer bring any agent harness and sandbox and switch training backends with one flag.
rLLM lets a developer bring an existing agent harness (Claude Code, Codex, Terminus-2, mini-swe-agent, opencode and others, or a Harbor-compatible task directory) or wrap their own agent, such as one built with LangGraph or the OpenAI Agents SDK, using an @rllm.rollout decorator, and run it in a sandbox such as Docker, Daytona, Modal, or locally.
The same agent code drives both evaluation and training, switching between training backends such as verl for distributed multi-GPU training, tinker for a single machine, or fireworks for the Fireworks platform, with one flag. It integrates more than 60 benchmarks, including Terminal-Bench, SWE-bench, AIME, MATH-500 and GPQA, and supports training methods such as GRPO, REINFORCE, RLOO, SFT and on-policy distillation.
It requires Python 3.11 or later, installs from GitHub with uv pip install, is written in Python, and is licensed under Apache-2.0.