Menu
Research & Blogs
10 min read
Single-Rollout Asynchronous Optimization for Coding Agents: Practice in Lego-RL
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 in Lego-RL.
8 min read
How Can Coding-Agent RL Generalize Across Harnesses?
Can one RL-trained checkpoint work across OpenHands SDK, Claude Code, and OpenCode? Comparing mixed-harness RL, single-harness training, and model merging with Lego-RL.
10 min read
Lessons Learned from LegoFlow-SWE
What task quality, reward hacking, and reasoning depth taught us about building reliable coding-agent training data.
15 min read
LegoFlow: Easy and Interactive Code Data Engineering
An agent-native framework for curating coding tasks, collecting trajectories, training models, and evaluating the full loop.
8–12 min read
Lego-RL: Harness-Native Reinforcement Learning for Coding Agents
Without modifying OpenHands SDK, Claude Code, or OpenCode, Lego-RL plugs native harnesses into GSPO training. The same Qwen3.5-35B-A3B gains +6.4 / +5.8 / +9.4 on SWE-bench Verified.
7–9 min read
SWE-Review: Closing the Loop on Issue Resolution with Agentic Code Review
An agentic reviewer that explores the repository, traces root causes, and returns structured feedback to close the generate–review–revise loop.
8–10 min read
What Makes Interaction Trajectories Effective for Training Terminal Agents?
Why the strongest agent is not always the best teacher, and how environment-grounded supervision enables far more data-efficient terminal-agent training.
8–10 min read
SWE-Lego: Pushing the Limits of Supervised Fine-Tuning for Software Issue Resolving
A supervised fine-tuning recipe built from curated instances, validated trajectories, error masking, a difficulty curriculum, and a test-time verifier.