An open-source collection advancing frontier agentic intelligence
LegoX is a series of open-source projects spanning agentic data pipelines, training frameworks, evaluation benchmarks and foundation models. Our mission is to connect these pieces into reproducible and extensible systems that advance frontier intelligence.
Can one rollout provide enough signal to train a coding agent? SAO makes single-rollout training practical with a learned critic, warmup, and advantage whitening, achieving competitive performance with more efficient training.
Can one RL-trained checkpoint work across OpenHands SDK, Claude Code, and OpenCode? We compare mixed-harness reinforcement learning with single-harness training and model merging across all three harnesses.
Agent-native workflows connect task curation, trajectory rollout, training, and evaluation. We also release LegoFlow-SWE, with 5,000 open-source tasks and 2,780 successful trajectories filtered from 12M PRs.
An open-source framework for harness-native reinforcement learning of coding agents. Code, data, and models are available now.
Our open models and datasets have reached a new cumulative download milestone on Hugging Face—made possible by the community using and building on them.
Collaborating institutions