Train the harness, not just the model: 6K examples of harnessed agentic RL lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%

arXiv 2608.17528·high signalXBlueskyLinkedInCopy link

Agent Lightning v1.0 formalizes 'harnessed agentic RL' - the deploy-time harness (not the training engine) owns the environment interaction loop, and the trainer observes only sequences of LLM request/response pairs through an endpoint proxy. That inversion creates concrete engineering problems the paper enumerates and solves: retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, each of which materially affects training stability. In roughly 3,500 lines of code and with only 6K training examples on modest compute, RL took Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified (+14.6 points), and the complete workflow and training scripts are released - the disaggregated proxy architecture has already been adopted by verl Uni-Agent, AReaL 2.0, slime, and Polar.