Researchers have released Agent Lightning v1.0, a lightweight framework for training agents in Reinforcement Learning (RL) environments. The framework is designed to work with 'harnessed agentic RL', where the agent's deploy-time environment interacts directly with the model post-training. This approach differs from traditional RL and introduces new challenges such as retokenization and loss normalization. Agent Lightning v1.0 supports arbitrary agents and provides a testbed for studying these challenges. The researchers evaluated the framework on various tasks, including instruction-following, search, and coding agents, achieving significant improvements in performance. The complete workflow and training scripts are released to facilitate reproducible research. --- Why it matters: This matters because it advances our understanding of how to train agents in complex environments, which is crucial for developing more capable AI systems. By providing a practical testbed for studying the challenges of harnessed agentic RL, Agent Lightning v1.0 enables researchers to improve training stability and effectiveness. Source: https://arxiv.org/abs/2608.17528