We’ve recently validated a new approach to accelerating Agent RL. 🚀
🔗 Technical blog:
legox.net/blog/lego-rl-sao/
A major cost bottleneck in Coding Agent RL is rollout generation.
Each rollout requires an agent to explore code, edit files, run tests, and iteratively fix errors. Sampling multiple trajectories for the same task can involve extensive tool interactions.
But what if we could train effectively with just ONE rollout per task?
In LEGO-RL, we explored SAO (Single-Rollout Asynchronous Optimization).
SAO uses a learned Critic as a value baseline, enabling meaningful training signals from a single trajectory and reducing reliance on group-based sampling.
📊 Results first
Starting from Qwen3.5-35B-A3B, trained on OpenSWE:
• SWE-bench Verified: 64.0% → 68.6%
• SWE-bench Pro: 39.8% → 42.1%
• SWE-bench Multilingual: 51.7% → 56.0%
⚡ Now the efficiency gains
To reach their respective best validation performance:
• SAO: 12.8k trajectories
• GSPO: 48.6k trajectories
That’s 73.7% fewer sampled trajectories.
Under our current setup, SAO also achieves 5.4× faster training steps.
The takeaway: Scaling Agent RL doesn’t always require more rollouts. Better learning signals can make each rollout count.
1
2
120
How Can Coding-Agent RL Generalize Across Harnesses?
We train one shared model with mixed-harness reinforcement learning (MH-RL) across three agent harnesses. We compare it with single-harness reinforcement learning and model merging on all three harnesses. It shows that MH-RL supports all three harnesses with one shared checkpoint. Its average accuracy is close to deploying a separately trained model for each harness, while reducing selected cross-harness performance gaps.
Check our recent post at legox.net/blog/lego-rl-cross…
🚀 We’re open-sourcing LegoFlow — a framework for building and improving long-horizon coding-agent data with Claude Code / Codex.
Long-horizon task construction is still painfully expensive: too many steps, too much manual work, and painfully slow iteration.
LegoFlow turns the whole loop into a programmable workflow:
⚙️ One-click skills for task construction, rollouts generation, training & evaluation
🔁 End-to-end iteration + self-improvement
📊 Visual Dashboard for the full pipeline
📦 5,000 curated tasks + 2,780 successful trajectories
The goal: make high-quality coding-agent data much easier to build, iterate on, and scale.
🔗 Code: github.com/LegoX/LegoFlow
📝 Blog: legox.net/blog/legoflow/
🤗 HF: huggingface.co/collections/L…
And we’re just getting started — more types of long-horizon agent data are coming soon. 🔥
Nice job by MiMo RL, but the cost ($1M/day) can be too expensive. Here we release the full RL training dashboards of agentic coding, a cheap alternative from Lego-RL for more audience in the community. rl-dashboard-openhands.legox…
Nearly half a year of silence. We spent it studying one problem: how far RL can scale.
MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.
Streaming the run: mimo.xiaomi.com/rl/
Lego-RL live training dashboards:
🧩 OpenHands / GSPO
rl-dashboard-openhands.legox…
🧩 OpenCode
rl-dashboard-opencode.legox.…
🧩 Claude Code
rl-dashboard-claudecode.lego…
🔀 Mixed Harness
rl-dashboard-mixed.legox.net
⚙️ SAO
rl-dashboard-sao.legox.net
🌍 Multilingual
rl-dashboard-multilingual.le…
1
67
Thrilled to announce Lego-RL🚀, a harness-native reinforcement learning framework for coding agents like Claude Code, OpenCode and OpenHands-SDK. It successfully boosts Qwen-3.5-35B-A3B from 64% to 70% on SWE-Bench verified.
Everything is open-sourced: legox.net/blog/lego-rl/
Btw some of our teams’ previous attempt on agentic coding, wellcome to drop by
- SWE-Lego: legox.net/blog/swe-lego/
- SWE-Review: legox.net/blog/swe-review/
- Terminal-Lego: legox.net/blog/terminal-lego…
3
53





