We are a open-source initiative spawning from Berkeley Sky Computing Lab to democratize reinforcement learning (RL) techniques and develop scalable systems for large language models (LLMs) and agents.
We are excited to announce rLLM v0.2, a major upgrade of our RL training framework. In v0.1, rLLM provided agent and OpenAI Gym-like environment abstractions to support training ReACT-style agents. In v0.2, we introduce AgentWorkflowEngine and AgentWorkflowTrainer—more general abstractions that enable arbitrary agentic programs to be trained. Agent builders and researchers can now define multi-agent systems, complex workflows (such as solver-judge, planner-executor, or MCTS), or agentic programs with custom reward functions, and train them with reinforcement learning without rewriting their production code.
We introduce Pepper, a real-time, event-driven architecture enabling the next generation of proactive agentic systems. With Pepper, we built a personal assistant that proactively fetches and summarizes emails, provides context before you even start a conversation, and continues to follow up while working on tasks. We open-source Pepper to advance the creation of proactive agentic systems.
We release rLLM, an open-source framework for post-training language agents via reinforcement learning. With rLLM, you can easily build their custom agents and environments, train them with reinforcement learning, and deploy them for real-world workloads.
