A new open-source framework called Agent Lightning treats the tools and control flow around an AI agent as something you train, not just the model underneath.
Agent Lightning v1.0 connects any agent harness, the tool-calling and context-management layer that sits between a model and its tasks, to reinforcement learning through a simple LLM endpoint proxy. The approach, first introduced in an earlier version of the project, has already been adopted by other RL frameworks including verl Uni-Agent, AReaL 2.0, slime, and Polar. The new release is a compact codebase, about 3,500 lines, built to test how well this setup handles instruction-following, search, and coding agents. For coding, the team trained a Qwen model on just 6,000 examples and modest compute, raising its SWE-bench Verified score from 41.8% to 56.4%, a 14.6-point gain.
That distinction matters because harnessed agentic RL flips who is in charge during training. The harness, not the training engine, owns the loop that talks to tools and the environment, while the trainer only sees raw request-and-response traffic to the model, which creates real engineering headaches like retokenizing conversations and merging samples correctly. As more coding agents ship as a harness bolted onto a general-purpose model, whoever solves that plumbing well gains a real edge over rivals still training models in isolation.
One asterisk: the paper credits the gains to 'Qwen3.5-9B,' a label that does not match any published Qwen release, since Qwen has no 3.5 series and no 9B-parameter Qwen3 variant, so treat that specific number as unverified until the authors clarify it.