OpenForgeRL: Microsoft Open-Sources End-to-End RL Training for Claude Code and Codex-Style Harnesses

arXiv / HuggingFace Daily Papers (Microsoft)·high signalXBlueskyLinkedInCopy link

Microsoft's OpenForgeRL (arXiv 2607.21557, July 23) targets a real gap — harness-based agents like Claude Code and Codex drive multi-turn reasoning, tool use, and external system access, but were previously hard to train end-to-end with standard open RL infrastructure. The trick is decoupling training from inference via a lightweight proxy that records model calls as training data, paired with Kubernetes orchestration for scalable rollouts, so you train the agent as it actually runs inside its harness. Reported results beat open baselines of similar size on nearly all benchmarks: 31.7 pass@1 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench for tool use; 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager for GUI agents.