rLLM is a framework for training language agents with reinforcement learning that decouples agent code from execution harness, sandbox, and training backend.
The tool addresses the fragmentation in agent training by letting developers write agent logic once and run it across different environments—whether using Claude Code, Codex, or custom LangGraph implementations—while swapping training backends (distributed multi-GPU via verl, single-machine via tinker, or Fireworks platform) with a single flag. It works by wrapping agents with a decorator, routing their LLM calls through a model gateway that captures token IDs and logprobs, collecting trajectories into episodes, computing rewards, and updating model weights. The same agent code runs unchanged during both evaluation and training.
Adopt this tool if you are training agents on coding, math, or reasoning tasks and want to avoid rewriting agent logic for different execution environments. It suits projects that need to experiment with multiple training methods—GRPO, REINFORCE, RLOO, or on-policy distillation—without refactoring. The framework includes 60+ integrated benchmarks covering math, code, QA, and agentic tasks, so you can start training immediately. The README does not compare rLLM to other agent training frameworks, so no comparison is made here.
Almost all open issues are raised by outside users rather than the core team, indicating a substantial base of adopters reporting real-world use. Maintainers typically respond to new issues and pull requests within a day. Work in the issue tracker is dominated by bug reports, questions, and documentation requests.
