Huawei & CUHK open-source Lego-RL: plug real coding agent harnesses into RL, SWE-bench 64.0→70.4
青稞AI · wechat · 2026-09-06
Huawei and CUHK researchers open-sourced Lego-RL, a Harness-Native RL training framework for coding agents that plugs real harnesses (OpenHands SDK, Claude Code, OpenCode) into RL without changing a single line of their code. On Qwen3.5-35B-A3B, SWE-bench Verified scores rose from 64.0/62.4/57.2 to 70.4/68.2/66.6 across three harnesses.
The framework tackles three failure modes of RL on real agent harnesses:
- Faithful: keeping policy updates aligned with true agent behavior despite train/inference environment gaps;
- Reliable: preventing environment failures and reward hacking from polluting training signal;
- Observable: AgentPlugin enables training validation, live diagnostics, and human review in a semi-automated data-to-ops loop.
Paper: arxiv.org/abs/2608.17393; code: github.com/LegoX/Lego-RL. Core developer Du Yiming (Huawei Leibniz Institute, CUHK PhD) will present the work in a Qingke AI talk on Sept 8.
More from coding & agent
- Early Astra in GitHub Copilot hits regressions: revert or let it self-fix? — DanWahlin · 2026-09-06
- Dev builds complete Roblox idle game entirely with agentic tool Roblox Build — rms80 · 2026-09-06
- Karpathy joins Anthropic: 700 experiments in 2 days, knowledge graphs boost his loop — blaizedsouza · 2026-09-06
- Agent speeds up a JSON scanner 1.7-2.1x on late-field lookups with a short-string fast path — arthurcolle · 2026-09-06
- Garry Tan: Aside might be the best AI agent harness, outpacing frontier labs — garrytan · 2026-09-06
- Astra coding review: 20-minute tasks drop to 4 minutes but burns 60% of Pro quota in a day — andimarafioti · 2026-09-06