The open-source reinforcement learning infrastructure project AReaL recently released its 2.0 version, specifically designed for agents applied in real-world business scenarios, providing systematic support for continuous learning. Through a unified reasoning gateway, AReaL 2.0 can record the interaction processes and feedback of agents during real-world task execution. This data is then used to train and optimize the underlying models, ensuring continuous improvement of agents in a safe and controlled environment. This version effectively addresses the challenge of agents' inability to self-improve after deployment, allowing agents to integrate into the online reinforcement learning process without requiring redevelopment, transforming multi-round dialogues, tool invocations, and other interactions in real tasks into learning resources. For enterprise application scenarios, AReaL 2.0 introduces a data proxy mechanism for agent trajectories, ensuring that security requirements such as permission control and data desensitization are met. This mechanism connects agent services, real-world task trajectories, data governance, and online reinforcement learning training, providing a solid engineering foundation for agents' continuous learning after deployment and promoting an evolutionary model where agents continuously enhance their capabilities through feedback in real-world environments. The AReaL project was jointly initiated in 2024 by teams from Ant Group, Tsinghua University, and The Hong Kong University of Science and Technology, and in May 2026, it evolved from Ant InclusionAI into an independent open-source community, joining the PyTorch Foundation Ecosystem project. Currently, the technical report and code for AReaL 2.0 have been fully open-sourced.