OpenForgeRL: Train Harness-native Agents in Any Environment

Open-source framework from Microsoft Research (with Columbia and Dartmouth; MSR's Baolin Peng is project lead, senior author Jianfeng Gao) for training harness-native agents end-to-end. Modern agents live inside inference harnesses like Claude Code, Codex, and OpenClaw, but open SFT/RL stacks cannot natively express their stateful, multi-process inference. OpenForgeRL decouples the two with a lightweight proxy that serves the harness's model calls while recording them as training samples for a standard RL codebase (e.g., veRL), plus a Kubernetes orchestrator that runs each rollout in its own remote container (building on Orchard Env), scaling elastically on Azure — enabling SFT and RL on any harness in any environment.

Recipe: Qwen3-30B-A3B-Thinking backbone, SFT distilled from successful MiniMax-M2.5 teacher rollouts, then GRPO-style RL from the SFT checkpoint. Using only hundreds to a few thousand tasks, OpenForge-Claw reaches 31.7 pass³ / 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench; OpenForge-GUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager — matching or surpassing models several times larger in the GUI setting. The analysis shows harness choice shapes learnability (some harnesses — ZeroClaw vs OpenClaw vs Codex — are substantially harder to learn than others) and that RL improves agentic reliability (self-verification, tool coverage, multi-step plans) while error recovery remains weak. Part of the model–harness execution layer we track alongside OpenHarness and Harness-Bench. Code, data, and models are promised but not yet public as of late July 2026 — watch for the repo.