LEGO-RL lifts coding agents across harnesses

LEGO-RL tackles the engineering bottleneck that makes reinforcement-learning coding agents difficult to train reliably: preserving harness behavior while keeping rollouts aligned with policy updates.

• –In-process proxying captures raw generation streams, mitigating train-inference drift caused by compaction and re-serialization
• –Stage-wise sandbox defenses directly target reward hacking and unreliable execution signals
• –SWE-bench Verified gains reached 6.4 points for OpenHands, 5.8 for Claude Code, and 9.4 for OpenCode
• –A reported rollout-training probability correlation above 0.99 suggests unusually strong optimization fidelity
• –The framework could make existing agent harnesses reusable RL environments instead of requiring bespoke training stacks