OpenForgeRL: Train Harness-native Agents in Any Environment
2026-07-23 • Artificial Intelligence
Artificial IntelligenceComputation and Language
AI summaryⓘ
The authors introduce OpenForgeRL, a new open-source system that helps train AI agents that use complex setups (or harnesses) for tasks like reasoning and using tools. Their system makes it easier to train these agents end-to-end by separating how the agent thinks from how it learns, using a proxy and Kubernetes to run tasks remotely. They tested OpenForgeRL on various challenging AI helpers and found it performs better than similar open models, even matching bigger ones in some cases. They also studied how different setups affect learning and confirmed that reinforcement learning helps agents be more reliable, though some skills like fixing errors still need work.
AI agentreinforcement learningmulti-turn reasoninginference harnessOpenForgeRLKubernetesproxymulti-process inferencetool use agentserror recovery
Authors
Xiao Yu, Baolin Peng, Ruize Xu, Hao Zou, Qianhui Wu, Hao Cheng, Wenlin Yao, Nikhil Singh, Zhou Yu, Jianfeng Gao
Abstract
Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks cannot natively express stateful, multi-process harness inference. To address this, we present OpenForgeRL, an open-source framework for training harness-based agents end-to-end in diverse environments. OpenForgeRL achieves this with a lightweight proxy that serves the harness's model calls while recording them as training data for a standard RL codebase (e.g., veRL), and a Kubernetes orchestrator that runs each rollout in its own remote container, together enabling training on any harness in any environment at scale. By decoupling training and inference, OpenForgeRL allows researchers to easily train, study, and improve agents directly in the real harnesses and environments they are deployed with. We validate our framework across diverse, complex harnesses and environments, spanning tool/claw-based agents and multimodal GUI browser- and computer-use agents. Using only 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. Beyond benchmarks, we analyze how harness choice (e.g., ZeroClaw, OpenClaw, Codex) and RL shape agent behavior. We find that some harnesses are substantially harder to learn than others, and that RL improves agentic reliability, such as self-verification, tool coverage, and completing multi-step plans, though critical abilities such as error recovery remain weak.