Ant Group, in collaboration with Tsinghua University, has released the stable version 1.0 of its open-source reinforcement learning training framework, AReaL. This version is centered on enabling "one-click RL training integration for Agents," allowing for compatibility with various Agent frameworks without requiring code modifications, thus making intelligent agent reinforcement learning training ready for immediate use.
Since the beginning of 2026, Agent technology has continued to gain momentum, with frameworks such as LangChain, Claude Code, and OpenClaw experiencing rapid development. However, this growth has also highlighted two major challenges. First, the cost of integrating training is high: existing Agent frameworks have diverse interfaces, often necessitating the development of extensive adaptation code for each integration. Second, Agents lack the ability for continuous evolution: the capabilities of most Agents are determined by the fixed weights learned by the underlying model during the training phase. Once deployed, they cannot be further optimized for specific scenarios, meaning their performance ceiling is fixed upon delivery.
AReaL is the first large-model reinforcement learning training system to feature fully asynchronous training and inference decoupling, enabling Agents to receive feedback and continuously optimize their decision-making through interaction with real-world tasks. The newly released v1.0 version makes it possible for any Agent to integrate into RL training with zero modifications. By introducing a Proxy Worker intermediary layer between the Agent and the training system, developers only need to change a single request address to connect to the training process.
Comments