OpenClaw's AReaL v1.0 Framework Enables Concurrent AI Agent Training

On March 4th, Ant Group, in collaboration with Tsinghua University, released the stable version of the open-source reinforcement learning training framework, AReaL v1.0. This release centers on enabling "one-click RL training access for Agents." It requires no code modifications, is compatible with diverse Agent frameworks, and allows intelligent agents to begin reinforcement learning training immediately.

Since early 2026, Agents have maintained strong growth momentum. Frameworks like LangChain, Claude Code, and OpenClaw have advanced rapidly, yet they highlight two significant bottlenecks. First, the barrier to training access is high: existing agent frameworks use varied interfaces, often necessitating extensive adaptation code for integration. Second, Agents lack continuous evolution: most rely on fixed model weights from their initial training phase. Once deployed, they cannot be optimized further for specific scenarios, with their capabilities capped at launch.

AReaL is the first fully asynchronous, training-inference decoupled large-scale reinforcement learning system. It enables Agents to receive feedback and continuously refine their decisions through real-world task interactions. The v1.0 release allows any Agent to connect to RL training without changes. By inserting a Proxy Worker layer between the agent and the training system, developers only need to redirect a single request address to enable training.

(Figure: AReaL's asynchronous training architecture seamlessly integrated into agents)

Take the popular OpenClaw framework as an example. Developers simply need to point the `base_url` and `api_key` in OpenClaw's configuration to the AReaL gateway. Their OpenClaw agent is then connected to reinforcement learning training. The agent continues performing tasks normally, while users periodically rate its performance. AReaL automatically gathers this training data and updates the model in the background, enabling the agent to evolve autonomously through continuous use.

The AReaL v1.0 release also introduces the native training engine, Archon. Built on PyTorch's native capabilities, it achieves full 5D parallelism—data, pipeline, tensor, context, and expert parallelism—lowering the installation and debugging barrier. It also offers multiple backend options for training and inference, facilitating flexible deployment across different environments. Remarkably, this complex distributed system was developed and validated from scratch in just one person-month. Within 32 days, nearly a million lines of code were modified to fully implement the Archon engine, enabling it to train billion-parameter Mixture-of-Experts (MoE) models.

The secret behind this efficiency breakthrough is AReaL's integrated AI-assisted development system, which automates highly complex engineering tasks.

AReaL v1.0 features an AI-assisted development workflow that provides end-to-end support, from planning and coding to verification and PR creation. When working on core modules like MoE parallelism, memory optimization, and algorithm implementation, a dedicated AI programming assistant acts as a senior expert. It provides timely, targeted guidance during code changes, ensuring the correctness of every modification. AReaL's AI-assisted programming is more than just a productivity tool; it can undertake "deliverable" R&D work in complex infrastructure projects, pioneering the next paradigm for AI infrastructure engineering.

The AReaL team has stated they will continue to iterate on the training engines, usability, and multimodal agent training. The code and documentation for AReaL v1.0 are now open-sourced in the inclusionAI community.

· GitHub repository: https://github.com/inclusionAI/AReaL

· Related paper: https://arxiv.org/abs/2505.24298