Modern AI agents do not operate as standalone LLMs; they run inside agent harnesses that manage tools, context, and control flow. Recognizing the harness as a critical component, Hugging Face researchers have introduced Agent Lightning v1.0, a lightweight framework designed to enable reproducible reinforcement learning (RL) for arbitrary agent harnesses.
The framework builds on the original Agent Lightning architecture, which connects agents to RL training through an LLM endpoint proxy—an approach now adopted by frameworks like verl Uni-Agent, AReaL 2.0, slime, and Polar. This paradigm, termed "harnessed agentic RL," directly involves the deploy-time harness in model post-training, narrowing the gap between training and real-world use.
Harnessed agentic RL differs fundamentally from traditional agentic RL: the harness owns the environment interaction loop, while the training engine observes only LLM request-response pairs. This introduces challenges in retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, which can destabilize training if not addressed.
Agent Lightning v1.0 implements a solution in approximately 3,500 lines of code, prioritizing simplicity. It supports arbitrary harnesses and serves as a testbed for studying these challenges. The researchers validated it on instruction-following, search, and coding agents, providing a complete reproducible pipeline for coding-agent RL.
Using only 6K training examples and modest compute, RL improved Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%, a 14.6-point absolute gain. The complete workflow and training scripts are released to facilitate reproducible research in harnessed agentic RL.