A team of researchers posted Agent Lightning v1.0 to arXiv on Aug. 18, a lightweight reinforcement learning framework for training agents without rewriting the tools and control logic that already run them. Applying RL through the framework raised the accuracy of Qwen3.5-9B, Alibaba’s open-weight model, on SWE-bench Verified from 41.8 percent to 56.4 percent, a jump of 14.6 percentage points, using only 6,000 training examples. That result matters because most agentic reinforcement learning work assumes deep access to a training engine’s internals. Agent Lightning’s core bet is that it does not need that access.
The framework’s first version introduced what its authors call harnessed agentic RL. Instead of embedding a trainer inside an agent’s code, it sits between the agent and the model as a proxy, watching a stream of prompts and responses rather than issuing the agent’s tool calls itself. The design has spread. AReaL 2.0, slime, Polar, and verl’s Uni-Agent framework have each adopted a similar proxy layer, according to the paper.
That design comes with a cost. The harness now owns the loop that talks to tools and the environment, while the training engine sees only a raw sequence of requests and responses sent to the model. Turning that log into a clean training signal is harder than it sounds, and the paper names five specific failure points: backend scheduling, sample merging, loss normalization, retokenization, and advantage calculation. Any one of those, mishandled, can destabilize the training run.
SWE-bench Verified is a widely used coding benchmark built from real-world GitHub issues that were screened for solvability. Solving one means producing a patch that passes a project’s hidden test suite, not just describing a fix. A 14.6-point jump from a training run of only 6,000 examples is a notable efficiency claim. Agentic RL results on that benchmark typically come from much larger training sets or reward models built specifically for the task. The paper does not compare its compute budget to any specific competing system, so the efficiency claim rests on the training-set size alone.
The team also tested the framework on instruction-following and search agents beyond coding, and released a full reproducible pipeline with training scripts specifically for the coding-agent experiment. That is worth noting on its own. Agentic RL papers often publish benchmark numbers without a runnable pipeline attached, which leaves replication to whoever is willing to reverse-engineer the setup.
Teams already building custom RL integrations for coding agents have a lighter-weight option to test: a proxy-based harness that reports comparable gains with a much smaller implementation footprint. Anyone fine-tuning agents against SWE-bench Verified should try to reproduce the 41.8-to-56.4 jump using Agent Lightning’s published scripts before committing engineering time to a custom RL stack.
arXiv, paper 2608.17528 (“Agent Lightning v1.0: Towards Harnessed Agentic RL”), submitted by Zhiyuan He and co-authors on Aug. 18, 2026.
