Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

OpenForgeRL: Train Harness-native Agents in Any Environment

Source

arXiv

arxiv.org/abs/2607.21557v1

Read the source

Source

Plain-Language Brief

OpenForgeRL is an open-source framework designed to train AI agents end-to-end in various environments using complex harnesses like Claude Code, Codex, and OpenClaw. The framework enables large-scale training by creating a lightweight proxy for model calls and a Kubernetes orchestrator for rollouts in remote containers. The study demonstrates the effectiveness of the framework across diverse harnesses and environments, with impressive results on benchmarks like ClawEval, QwenClawBench, OSWorld-Verified, Online-Mind2Web, and WebVoyager.

What the Document Says

  • OpenForgeRL is a framework for training harness-based AI agents end-to-end in diverse environments.
  • The framework consists of a lightweight proxy for model calls and a Kubernetes orchestrator for rollouts in remote containers.
  • OpenForgeRL allows researchers to train, study, and improve agents directly in the real harnesses they are deployed with.
  • The study validates OpenForgeRL across diverse, complex harnesses and environments, including tool/claw-based agents and multimodal GUI browser- and computer-use agents.
  • Results show that OpenForgeClaw achieves 31.7 pass^3 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench, while OpenForgeGUI reaches 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager.
  • OpenForgeRL outperforms open baselines of similar size on nearly all benchmarks and in the GUI setting matches or surpasses models several times larger.

Key Concepts Mentioned

  • AI agents
  • harness-based agents
  • inference harnesses
  • Claude Code
  • Codex
  • OpenClaw
  • OpenForgeRL
  • SFT/RL stacks
  • veRL
  • Kubernetes
  • multi-turn reasoning
  • tool use
  • external systems

Claims Made by the Document

  • OpenForgeRL allows training harness-based agents end-to-end in diverse environments.
  • OpenForgeClaw and OpenForgeGUI show impressive results on various benchmarks.

Evidence Used

  • The description of OpenForgeRL's components and its ability to train agents across diverse harnesses and environments.
  • The presented results from ClawEval, QwenClawBench, OSWorld-Verified, Online-Mind2Web, and WebVoyager benchmarks.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.