LEGO-RL: harness-native reinforcement learning for coding agents
Lego-X · hf · 2026-08-20
LEGO-RL is a reinforcement learning framework for coding agents that connects native coding-agent harnesses directly to scalable policy-gradient training, avoiding a rewritten training environment.
Key components:
- In-process LLM proxying for transparent integration between training and the harness's model calls;
- Sandbox orchestration to safely execute agent-generated code and actions in parallel;
- Integrated monitoring of agent behavior and metrics during training.
The authors report improved sparse MoE model performance across multiple harnesses.
More from coding & agent
- TrueFoundry Releases TrueForge, an Open-Source Agent Runtime Harness — omarsar0 · 2026-08-20
- TrueForge open-source agent harness cuts costs by 30%-75% in testing — omarsar0 · 2026-08-20
- Anthropic engineer: Build systems that prompt themselves, don't just prompt — ifioknkem · 2026-08-20
- Hermes Agent integrates Spotify plugin, picks music while you work — alexcovo_eth · 2026-08-20
- Devin is now powerful enough, Cognition says only user ambition limits usage — marvinvonhagen · 2026-08-20
- Eve Launches: Build Durable AI Agents with Next.js-Style Architecture — evilrabbit_ · 2026-08-20