OpenForgeRL trains harness-native AI agents end-to-end across environments

Researchers present OpenForgeRL, an open-source framework for training harness-based AI agents end-to-end in diverse environments, in an arXiv preprint. The abstract says modern agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw for multi-turn reasoning, tool use, and external systems, but those harnesses make end-to-end training hard on open SFT and RL stacks that cannot natively express stateful, multi-process harness inference. OpenForgeRL uses a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase such as veRL, plus a Kubernetes orchestrator that runs each rollout in its own remote container, enabling training on any harness in any environment at scale by decoupling training and inference. The authors validate the framework on tool and claw-based agents and multimodal GUI browser and computer-use agents. With hundreds to a few thousand tasks, OpenForgeClaw reaches 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench. OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager. Both outperform open baselines of similar size on nearly all benchmarks, and in the GUI setting match or surpass models several times larger, the abstract states. Analysis of harness choice, including ZeroClaw, OpenClaw, and Codex, finds some harnesses substantially harder to learn than others. RL improves agentic reliability such as self-verification, tool coverage, and multi-step plan completion, though error recovery remains weak.