Agent Lightning v1.0 introduces a lightweight framework designed for "harnessed agentic RL," a paradigm where the deployment-time harness—which manages tools, context, and control flow—directly participates in model post-training. Unlike traditional agentic RL where the training engine manages environment interactions, this disaggregated architecture relies on an LLM endpoint proxy, meaning the trainer only observes sequences of LLM request-response pairs. This structural shift necessitates specialized handling of retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling to maintain training stability. Implemented in roughly 3,500 lines of code, the framework supports arbitrary agent harnesses and provides a reproducible pipeline for evaluating instruction-following, search, and coding agents. The researchers demonstrate the efficacy of this approach through reinforcement learning applied to a Qwen3.5-9B model. Using a limited dataset of only 6,000 training examples and modest computational resources, the method achieved a significant performance boost on the SWE-bench Verified benchmark, increasing the score from 41.8% to 56.4%—a 14.6-point absolute improvement. By decoupling the environment loop from the training engine, Agent Lightning provides a practical testbed for advancing RL methodologies within complex, tool-using agentic systems.