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

The paper introduces OpenForgeRL, an open-source framework for training AI agents using harnesses like Claude Code, Codex, and OpenClaw in various environments. It allows end-to-end training of these complex harnesses with open infrastructure, which is usually challenging due to the stateful, multi-process nature of these harnesses. The framework uses a lightweight proxy to serve model calls from the harness and record them as training data for a standard RL codebase (e.g., veRL). A Kubernetes orchestrator is used to run rollouts in separate remote containers, enabling large-scale training in any environment. The authors validate their framework across diverse harnesses and environments, showing that it outperforms open baselines of similar size on various benchmarks.

What the Document Says

  • OpenForgeRL is a framework for end-to-end training of AI agents with complex inference harnesses.
  • The framework uses a lightweight proxy and Kubernetes orchestrator to achieve this goal.
  • It allows training on any harness in any environment at scale.
  • OpenForgeRL was validated across diverse, complex harnesses and environments.
  • Results show that OpenForgeClaw performs well on ClawEval and QwenClawBench.
  • OpenForgeGUI performs well on OSWorld-Verified, Online-Mind2Web, and WebVoyager.
  • The framework outperforms open baselines of similar size on nearly all benchmarks.
  • In the GUI setting, it matches or surpasses models several times larger.

Key Concepts Mentioned

  • AI agents
  • inference harnesses
  • Claude Code
  • Codex
  • OpenClaw
  • SFT/RL stacks
  • veRL
  • Kubernetes
  • rollouts
  • ClawEval
  • QwenClawBench
  • OSWorld-Verified
  • Online-Mind2Web
  • WebVoyager

Claims Made by the Document

  • OpenForgeRL allows end-to-end training of harness-based agents.
  • It improves agentic reliability, such as self-verification and tool coverage.

Evidence Used

  • The text describes the OpenForgeRL framework and its capabilities.
  • Results are provided showing the performance of OpenForgeClaw and OpenForgeGUI on various benchmarks.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The agent may struggle to understand the specific meanings of 'pass^3', 'pass@3', and 'harness choice' without additional context.
  • The agent may not be able to accurately interpret the implications of the findings without human analysis.

Open Questions

  • What is the significance of the improvements in agentic reliability?
  • Why are some harnesses harder to learn than others according to the authors?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper doesn't specify how the authors define 'pass^3' and 'pass@3'.
  • It's unclear what is meant by 'harness choice (e.g., ZeroClaw, OpenClaw, Codex)' as these are not mentioned elsewhere in the text.
  • The paper doesn't discuss the limitations of the authors or any potential biases in their research.

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.