LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents

Summary

LEGO-RL is a framework that bridges native coding-agent harnesses with scalable policy-gradient optimization without modifying their internal control flow. It ensures faithful optimization via in-process LLM proxying for token-level alignment, reliable execution through scalable sandbox orchestration with image caching and stage-wise defenses against reward hacking, and observable training via an integrated plugin and Live UI. When training the sparse MoE model Qwen3.5-35B-A3B with GSPO across three native harnesses, LEGO-RL improved performance on SWE-bench Verified: OpenHands SDK from 64.0% to 70.4%, Claude Code from 62.4% to 68.2%, and OpenCode from 57.2% to 66.6%, while maintaining a rollout-training probability correlation above 0.99.