---
title: "OpenForgeRL: Train Harness-native Agents in Any Environment"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.21557v1"
permalink: "https://feed.astralor.com/items/itm_d3c071465ce8444859c6c248"
published: "2026-07-23T17:38:30.000Z"
score: 81
channel: "tech"
tags: ["AI Agent","Claude Code","Codex","OpenClaw","OpenForgeRL","Reinforcement Learning"]
---

# OpenForgeRL: Train Harness-native Agents in Any Environment

## AI commentary

> This paper is a must-read for AI agent engineers and researchers stuck with complex inference harnesses that hinder end-to-end training. OpenForgeRL finally bridges the gap between harnesses and RL, letting SFT/RL stacks natively use production-grade inference frameworks. Note that error recovery remains unsolved, but now there's a reproducible playground. For teams deploying reliable GUI or tool agents, this framework is essential.

## At a glance

OpenForgeRL introduces a lightweight proxy and Kubernetes orchestrator that converts model calls from complex inference harnesses (e.g., Claude Code, Codex) into trainable RL data, enabling end-to-end training. On tool-based and multimodal GUI tasks, it matches or surpasses larger models with only hundreds to thousands of samples, and provides analysis on how harness choice and RL shape agent behavior, revealing improvements in self-verification and multi-step planning while error recovery remains weak.

## Key points

- OpenForgeRL decouples training from inference via a proxy and K8s orchestrator, enabling end-to-end training on any harness and environment. 
- Achieves 31.7 pass^3, 33.7, 37.7 on ClawEval, QwenClawBench, OSWorld-Verified, outperforming open baselines of similar size. 
- RL improves agent reliability (self-verification, tool coverage, multi-step plans) but error recovery remains weak. 
- Requires only hundreds to thousands of tasks for significant performance gains.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_d3c071465ce8444859c6c248