Microsoft Open-Sources OpenForgeRL for Training Complex Agents in Native Environments
microsoft · hf · 2026-07-25
Microsoft released OpenForgeRL, an open-source framework designed to tackle the challenge of end-to-end training for modern AI agents that rely on complex inference harnesses (like Claude Code and Codex).
Core Mechanism
- Decoupled Training & Inference: It uses a lightweight proxy to record model calls from real harnesses as training data, executing rollouts in isolated remote Kubernetes containers.
- Compatibility: Integrates smoothly with standard RL codebases like veRL, enabling scaled training across any environment.
Performance
- Tool-based Agents: Trained on only hundreds to thousands of tasks, OpenForgeClaw achieved 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench.
- GUI Agents: OpenForgeGUI outperformed similarly sized baselines and matched or surpassed much larger models on OSWorld-Verified (37.7) and Online-Mind2Web (63.0).
Key Findings
The research highlights that harness choice significantly impacts learning difficulty. While RL notably improves agentic reliability (e.g., self-verification, tool coverage), critical abilities like error recovery remain weak.
Related event: Microsoft Open-Sources OpenForgeRL for End-to-End Agent Training(4 posts)→
More from coding & agent
- Andrew Ng’s OpenWorker open beta runs local or cloud models and plugs into Slack — Saboo_Shubham_ · 2026-07-26
- Grok 4.5 tops several coding-agent benchmarks while claiming a lower cost per fix — XFreeze · 2026-07-26
- Recursive self-improvement may start with AI improving the code that builds its successors — imjustnewatai · 2026-07-26
- Agent inference costs are exploding as multi-agent workflows burn through tokens — 机器之心 · 2026-07-26
- Parallel scout agents can improve context understanding while cutting latency — pvncher · 2026-07-26
- Someone built a full 3D Doom-style game with SQL queries — thursdai_pod · 2026-07-26